Agentic AI for Investors - A BIG (Bengal Investment Group) Initiative for #Investor #Education

0:02 Okay. So, thank you all for joining. this is the first time we are doing this online as part of the you know we've started a forum in in Kolkata we're calling it Bengal investment group and the whole idea is that we wanted to share knowledge we wanted to have a place where you know we could interact and that's the whole idea for this group and today we are doing this online because it's it's dura puja time in kolkata prep and the usual place that we have that's you know the surrounding areas it's very close to new market so it's in fact bang opposite new market so it's extremely crowded at this time of the year so we thought we will do this online and this format is more suited for the presentation that we are have planning to have today Rudro has been a friend and a co-investor for a long time. we've known each other for very long and he has a very strong technology bank background. He's he's spent about 17 plus years in the IT industry working mainly in AI, machine learning consulting.

1:29 He's worked in you know companies like IBM, Cognizant, Accenture and in his current role in the current company he is the senior manager for data science. He's working on machine learning engineering ops capabilities etc. And he's he holds an MBA from IMT Gazyabad. These are B tech in electronics. just recently if I'm not wrong completed a chief digital and AI officer program from ISB Hyderabad and also an applied agent AI course from MIT. So he's not only a guy who's continuously been working in this field but he's also keeping himself updated. that's very important for all of us. And second you know the other part which is equally important for us is he's a investor himself and he's been an investor for many years. So we wanted this kind of a combination of somebody who is an investor and who is going to help us you know understand how the you know how to use AI in an agentic form. So today we wanted it to be a a sort of a demonstration class so that we can all end today's session with some takeaways. So that's the idea.

2:55 with that I will you know sort of hand it over to Rudro and Rudro take over from here. >> Oh thanks a lot. let me first share my screen. Let's see if the screen sharing is working. Let me know if the screen is up for everyone. >> Yeah. >> Are you able to see? >> Yes. Yes. Yes. Go ahead.

3:26 >> Perfect. And hope my voice is clear for everyone. >> Yes, it is. Absolutely. >> Perfect. Oh, thanks Da for all the kind words. I think Dada has been a mentor of sorts for not only for me for the many of us. who started their investing journey from Kolkata and he has been a pioneer and like motivating inspiring all of us. So today I think our aim is to let everyone we'll just take today's session more from a retail equity investor lens and try to understand how AI can help us in this journey.

4:02 So standard disclaimers we will use some stock names. These are just for illustration purpose. There is no recommendation whatsoever. when you are working with AI always validate the figures, dates, estimates, do your very deep due diligence and always get a advice from a SEB registered advisor before you put them in a single rupee in your investment. I think Dada mentioned I think these are some of the handles. feel free to interact with me on Twitter. I was been active in value picker and blogs earlier.

4:34 Now who is this session for right? So if you already are a pro u sort of a AI user maybe you'll not get much maybe you'll get certain things but if you are someone who is thinking hey AI is too difficult how can I get started and what we'll plan to do today is lay out AI for investors without any jarens as much as jargon free we'll try to keep it and through the session we'll try to develop our own process and we'll have to experiment so you cannot get it right from day one but you have to be flexible to start experimenting ing.

5:08 So we will distribute the next two hours into three sessions. First session we will do some primers, some basic grounding on AI, what to use, what to not to do. In the second session, we'll do some hands-on exercises. I will share a sort of a notebook where you will have all the prompts and everything curated. You just follow along with me as I do it on the screen. we'll use cloud pro for that. And in the last session, I will show you the agent tech system I've been developing exactly based on the principles we will learn in session two.

5:40 Okay. So let's get started. So the AI landscape. So I think change is always difficult. Remember when we moved from the shouting trading ring in BSC in '95, it was a big change. Then we moved from let's say the basic phone phoning your broker, hey can you place this order to internet. and the last five years I think everyone has moved to mobile trading right so these are all big changes similarly AI is there it's ubiquous now so we have to all adopt AI for our daily usage so let's start with what is an LLM for this group for a retail investor I would just want to know this is my equity analyst my junior research analyst who has read a ton of document given that it has read a ton of books Wikipedia internet sites and all it has a general understanding of the world around us. Right? So when we throw in certain terms, say hey analyze this annual report for us give me a summary of this conference call. It is able to connect the dots and understand the intent behind what you are asking.

6:41 It speaks in words not numbers. It will see in the later slides. It starts generating and thinking in words and output is generated in tokens. Most important has no memory of its own. So even if you have a lengthy chat for three four hours on one day and you're starting a fresh session the second day it will always start from scratch unless you ask it to use the memory of the previous days. So how does it actually think? So when we type in a query to today's models like cloud or gemini or openai it gives us the impression it is thinking but under the hood it is not right. So we have to be very very cognizant of the fact that a LLM cannot think. So go back to your mobile for a second, right? When you do a mobile and you type the thing like the stop closed, you automatically see three options. The this is how autocomplete works and over usage multiple days you'll see it is trying to pick up the words you use most. So it is sort of training itself. Think of the LLM as a more sophisticated broader autocomplete. Given it has seen that plethora of documents, web search, everything, it has started to identify the patterns we human beings use. So the responses it generate looks more and more plausible as of someone with very deep thinking is generating those. But under the hood, it is just connecting the dots and presenting the next token, next output. It is not able to think. So whenever you are putting a task to the AI always remember that unless a real world analyst with maybe the junior most analyst in your team which can have a person and thing the LLM cannot. So we very very cognizant in putting the guard rails the instructions so that it can actually generate the output which is reasonable for you.

8:31 Now let's go ahead and with this notion what the LLM or the agentic system can or cannot do for you. It can write and draft it's very competent at that it can summarize long notes. It can give you extract of let's say 500 page reports. it can summarize annual report based on the instructions you give and it you can give it certain patterns. It can analyze and brainstorm with you but it cannot make a final decisioning. It cannot think through. So never ask a LLM hey should I buy this stock always your capital view your decision. So do a thorough due diligence verify the numbers multiple times and then only make a very informed choice consulting your advisor.

9:14 So again never ask to predict something hey how will this stock do in the next 3 months. It can generate facts it can give you the strength and weaknesses it can give you a balanced view but never take it for an action. The action should always be come from you as an individual investor. Okay, this is an important topic hallucination. Let's understand it from our day-to-day life. I think you all have encountered a friend in different gatherings who has clueless about the context but always bombard something right it's completely factual completely fixious and later you found out hey he was talking gibberish lms also have a tendency if you do not put the guardrail it will not come up and say I don't know it will give you a fixious fact let's say if you ask it if you press it to give you the quarterly numbers for a company for the next quarter which is not even published it will come up with a plausible number which is completely factual. So unless you put in the right guard rails, you will start using a lot of fictional or completely ambiguous number which can derail your hypothesis.

10:18 There have been cases in the past Google Bard was fined. There was a very famous New York lawsuit where the lawyer was not aware of this. He sort of filed his response. There were a lot of fictional cases and was fined by $5,000. So in general well by doing equity analysis when we've been doing a very factual research we'll have to put very streamlined guardrails and ask it hey if you can't find this information tell us I can't find it. So unless you put this guardrails it will try to hallucinate will be confidently wrong and that will sort of derail your whole thesis and research.

10:53 Okay. Now this is another topic we need to understand. So I want you to visualize this whiteboard on the left right. So when we bombard a LLM with a large number of documents, large number of context, it cannot process it. So think of this whiteboard. So first you start with a prompt, some information from a call transcript or anything, the board is fail. The moment you start putting more information, it start erasing the earlier one. Right? So we have this notion that I'll dump everything and ask very intelligent questions. It will not work that way because your earlier provided context is wiped out by the time you bring in the new ones. So we have to be very mindful that what it actually is holding and there are ways to do that we will see later in the session. So be very mindful that give it very limited context specific to the task we are doing.

11:47 Let's go ahead. Now this is also an important paradigm. It is not be relevant today as much it was very critical earlier. all models have a knowledge cut off date means they have been trained on books information website everything till that cut off date. You can query a model and say hey what is the knowledge cut off date you will come back with a response. Earlier what was happening is let's say if the model was not connected to the internet and you're asking it more latency what happened to the stock or this company in this week it was not able to answer. So let's say you're using claude or any other AI tool be mindful to give it the access of internet. So it can complement this knowledge gap. So it has the understanding of the world based on the training and it needs to complement from its knowledge cut to present using internet and other tools. So unless you give it that it will be very difficult for the LLM to work and operate.

12:47 Okay let's go ahead. Now this is also very interesting. So we tend to think that sorry I think there is a hands raise. We'll not be able to take questions as of now. Please put it on the chat. We'll try to come back to your questions on the chat. So this is nondeterministic. So what happens is we are very used to hey I will give you a recipe. Let's say I want you to do certain things and it will always give me the same answer. So people often get baffled. Hey I wrote this query yesterday. it give me something and today it is giving me something else. As we'll see it is guessing right as we saw it's a sophisticated autocomplete. It doesn't have a memory of its own. It doesn't know what it has given you yesterday. So the same query will yield reasonable but two different answers. So every time you repeat the query let's say you ask me what are the top three risks. So both the answers both the responses are relevant but will be deter different and nondeterministic. So we have to weaken this mind that if I am working with the LLM always if my guards are strong the answer is reasonable but never accurate or let's say never exactly same as I was doing yesterday.

14:00 Okay. Now why a human needs to be in the charge? So we saw some of these pitfalls of hallucination where it is trying to guess on its own the context being wiped out. So it might have an incomplete context. it might miss some numbers. So it is always important to be a human in the loop. So whatever it is giving you, you have to be accountable. You have to validate. So no matter how sophisticated is your agentic AI system, the final onus lies with you. Like a co-pilot in a aircraft, you have to take control of the decision making. You have to probe deeper and account for all the information before you make a trade.

14:41 Okay? And if we see this envision this from a risk mitigation point of view how that control helps. So think of it that they it has proposed sort of a strategy. Hey this is the range this is kind of a good buy in this range. Now you need to understand that what are the risks first that risk control is very important. Was any figure hallucinated? Was any outdated facts used? Did it miss any recent facts? So first that governance gate has to be checked by you the data integrity part. Next you have to give it very clear rules what to consider and what not to because if those rules are not clear and think of it every individual investor can be a very different persona. Someone can be taking a very short-term trading bet.

15:27 Someone can be taking a short to medium-term bet. Someone can have a very long-term horizon. So your governance and flow should be crafted to what you are trying to achieve from that investment proposal right it will be very very fine-tuned and crafted by you and you would need to put that governance authorization layer and then once you see and validated and tranqued the fact then only you approve that so these things needs to be fine- tuned and crafted it cannot be a one size fits all every investor here I see 150 of us in the call have a very different persona have a expectation what he or she's expecting from a certain investment certain trade bet you have to fine-tune and craft that put in your own judgment put your own governance and then execute so the human loop is very very critical and we saw even today there are some negative news from anthropy right so tool is never to be dreamed it is doing what it was meant to you have to take the owners and control to get this done now let's come to tokconomics So you will hear this word tokens. We'll keep it very simple. a token is roughly around four character. This is the unit of work with which LLM or my equity analyst reads and writes things. Right?

16:45 So you see here in the example I put a simple query is HDFC bank a good stock to buy. Now this is broken into nine tokens and whatever response comes out let's say there is a simple oneline response. This is a response is 120 tokens. So just for a simple study we need to understand that ALMs operate our entire budget when whatever tool you go they will give you a quarterly limit or let's say monthly limit. Hey this month this is your total AI limit and that limit is determined by tokens. So you have a volume of tokens. So the moment you read something the moment you write something tokens are expensed. Every query, every document you put on the chat has been picked up and it is broken down into tokens and these are sort of calculated when you go to cloud and see your usage that hey I have used 10% 12% of my weekly limit. This is based on the tokens.

17:44 Now if I'm hoping that a lot of you have already subscribed to cloud pro we'll be heavily focusing today on the cloud models. We will also look at other families and their corresponding choices. But let's say if you're in the plot world, you will see that there are different models, right? How they differ is that these models, some of the models are meant for quick response. So think of that I have three different vehicles in my garage. I can use a bike, I can use a car, I can use a truck. Now for daily errors, right? So quick chats, rewriting something. Hey, can I just quickly refactor this email? Can you quickly summarize this paragraph into one line? Very simple task which we don't need to think a lot. Use the basic model. Use a bike for small errors. So haiku is the name of that basic model.

18:32 So if you are in cloud you will use haiko for your day-to-day very basic task reformatting something summarizing etc. Sonnet is your card for your daily usage in the city. So if you are doing some everyday research quickly looking up a stock quickly analyzing a trans transcript in industrial presentation anything sonnet is your go-to model and if you are doing some heavy task right so moving between cities or getting a truckload of people from one place to other you need to call the big guy which is opus now you might say hey if I have everything I will always use opus the challenge is on the right Right. So for the same weekly consumption we saw there is a cost of tokens. So haiku if the cost of haiku is x sonnet is twice as expensive and opus is four times as expensive. So that means if you are able to walk for let's say 3 hours with haiku that time will be immediately shortened for sonnet and opus. So be mindful your weekly limit have to optimize and use it for the daily tasks. So from planning to thinking to deep research balance it out from these three different models.

19:48 Okay. Now I think we all are aware of these three broader family of models. These think of their three different brokers we use all doing the more or less a similar things similar offerings at similar price points. Roughly 2000 is the pro license for use. You can use openAI you can use Gemini you can use cloud for any of the purposes. Today's session is focused on cloud but does not mean you have to use only cloud. You can use any three of them. So what are the corresponding choices? We already looked at the cloud repository. We have hikosetic and opus. Corresponding choices on the open stack are Luna, Terra and Astra. So for example, if you are a heavy user of open, you'll use your daily basic stuff for with Luna, the high stuff with Terra and only go for the most complex and expensive task with the Astra model. Similar ratios of token consumption, token pricing between the three different models. And on the Gemini stack, it was just use flashlight for very quick tasks. Use flash is a allpurpose model and 3.1 Pro is your deep thinking model. Google has recently launched the Gemini 4 argon but it is now still in preview for security and other research areas. It is not open for public yet.

21:06 Okay. Now with that said, now we know about models. Now we know what are the three choices. We know what to use when. Now let's come to the prompting. So if we start with a very basic prompt is this talk good? It's a very vague thing, right? So this guy, our equity research analyst has no thinking of its own like we saw. So if your prompt is very vague, you will come back with a very vague response and then you will blame the tool. Hey the tool is not working. So we have to be very specific, give it a context, give it a format and then iterate. Most of the times when you're starting your AI journey, it might feel frustrated that hey I gave it a query, it was not able to understand. So sit with it. It is just the intern in the office day one. Sit with him. Give him the context. Tell him how you work. Show him some of your examples. Hey, I want you to do some things. Here are three examples. Can you do this for now? On the fourth instance, the LM is smart enough to understand your process, understand what you are looking for and able to give you the right response. So, you'll see as we go through the exercises, our quality of output will drastically increase. the it will be more pinpointed, more focused when we move from this very basic weight to a curated one.

22:25 So now the next part as we'll discuss later in the call in the workout session is prong to skills. So let's take a pause. if someone might say hey I'm doing very good with prongs. What is the need to do everything else? Now think of it. You have a very detailed prompt that work for one company. Maybe you want to do a analysis of annual reports. Maybe you want to do equity research but it is not generalizable and every time you come back you have to paste that prompt. You have to craft that. It is not a repeatable process. A prompt is telling the models what you are doing. A skill is a mapping of your process. A skill tells you how to go about certain thing. We will see with more examples. So what is happening is a skill let's say for equity research you might have 12 steps and if you start putting the 12th step in a broad prompt it will fill the entire context window the module will not be able to tackle with everything else. So when we move from prompt to skill what happens is they just load the name and description of the skill. We'll see some examples later. So the entire body let's say your 12 steps are hidden. So the LLM knows that here is the skill this is the name of the skill and in the description it briefly talks about the 12 things it does. So when later in the workflow step six is required it goes to the skill file picks up the detailed prompt for only step six. So it helps protect the the context window help you work better with large number of documents and information and makes the process repeatable. So you do a very detailed prompt once you save it as a skill and going forward you don't need to tell anything just say hey can you do this the lum would know that here is a recipe you have already given for me I can go look up the certain step and walk with that now very quickly a bot versus agent I think bot we are all familiar we have all set let's say price alerts in screener right if the stock is down 20% from the high send me an alert the problem with bot is they are just good for one small specific task. It cannot sort of adapt on a dynamic nature of things. If let's say a particular stock price feed is not present, right? It is let's say listening to a feed from NSE or BSC. If that path is blocking, it cannot do anything. The alert will fail.

24:51 Compared to that, an agent is more dynamic, more adaptable in nature. It you give it a go. We'll see the more drill down on how an agent operates later in the workshop. It sort of breaks down the task and it can use the different tools. So for example, if you ask an agent to an analyze an annual report and maybe the report is not there, it will try to see can I go to the website, can I download the report and do it. A bot will not do that. So an agent uses tools and it has a plan of action to meet the goal. It will trying to do this pathways. It will always come back to you for decisions. But this is more dynamic, more adaptive and can use tools, check errors on its own and can do the stuff for you.

25:36 Okay. Now this is a great book. So it malik is one of the greatest thinkers today. He's a associate professor at Wharton. He has a huge follower on Twitter on Substack. So he has written a great book co- intelligence. This is a book without any jarens. It is just lay out how to sort of start working with AI in a day-to-day life from different environments from equity research from creative sites. I recommend all to read this book. So Ethan talks about four broad things. Invite it everywhere.

26:05 Start including the AI into your day-to-day processes. Always be the human in the loop like we discussed and give it a persona. Right? So throughout the day like you might be doing a trading you might be doing investing always give it a persona so that it can act based on what you're looking for and give you the response and the fourth point is today's AI is the flow so you might be frustrated hey this is not working today every month these families Gemini anthropic they are all coming with the latest models so today's model what you're working with might be the flow it's not the selling so 2 3 months down the models might evolve and the thing you are not able to do today can be achieved. So be flexible, work with them and adapt them.

26:53 Okay. So that brings us to the end of the first session. So key things we want to do is LLM's predict language. We know they don't know facts. Then they can always sound confident when they're still wrong. Always verify and validate. we know the different families of models what to use and a human might be always in the loop. Okay. So this brings us to the first session any quick thing on the chat we might take one or two questions anything you want to call out before you go to session two.

27:32 >> No I think this is fine. >> okay >> let's go ahead. >> Okay perfect. So then we will go to the session two. So session two I I hope everyone has their plot ready. So we'll and please be on the laptop it will be effective there. Let's start on the session two. Okay. So what we will do we will look at certain nuggets. Now we introduced what is prompt, what is skill, what is agent. Now we will go a bit deeper into each of these as we go through the exercises. So we'll start with the first our first couple of exercises will be on prompting.

28:09 So there are different formats you can do with prompting. We will focus from a equity research point of view. A very simple format just four things. A role what role you want the LM to take what is the context what you actually trying to achieve from this prompt the task you want to do and a format. Okay. So let's do that. Okay. Okay. The one important call out is today you will see we will generate a lot of output in HTML and not on PPD because we have this notion in equity research you are very sort of comfortable with seeing long decks and all but decks have a lot of limitation in terms of design in terms of layout. LLMs work much better with HTML. So today we will switch to HTML. All the outputs we will generate you'll see you can use more interactive tabs and all. Let me show you one quick example given we have the time.

29:03 Just one second. Let me know if this page is visible for you the HTML page. Are you still seeing the screen? >> I'm seeing the screen. >> Okay, let me reshare and see if I can show you the full desktop. Let me the HTML page is there.

29:35 >> Yes. >> Okay. Perfect. So, this is an example. let's say we read a lot of ton of books, right? And often I want to go back and see, hey, can I revisit my favorite books? Now, a PDF or a PPT is a very fixed format. You always need PowerPoint open. So, look at this. is this is an interactive sort of a page where it has laid out all the books with very curated custom covers for it. So I can go into atomic habits. I can read hey what are the key things the author wanted me to do what are the some things I can take to my work look at some of the memorable quotes and come back. So just a simple prompt I was just playing around and it created for 25 books. so any recent books you read we can start adding to the library. So HTML is a very nice sort of a format where you can do this kind of interactive things which is not possible in a PowerPoint. So switch your output to HTML. So that gives more power to the LLMs.

30:33 So switching back to the PPT. Hopefully you're able to see the PPT now. Perfect. Okay. Now, so we this is done. Now let's go to the today's session. So one action I need from all of you is use the QR. You will go to a drive link which is there on top and you'll be able to download the companion books for Windows and Mac. So once you are able to do that just give me a thumbs up so that we can start with the first exercise today.

31:05 Let me know if this QR is working. Yes. No. I I can try putting the link on the chat. Would that help? >> Yes. >> Are people able to use the QR? I see some thumbs ups. >> Yeah, it would be good if you can put the link in. >> Yeah, put it on the chat as well. Let's see.

31:35 >> Okay. I think I'm not able to. So, Bisha, can you do that? I'm not able to share on the chat. You have the link. If you can share, please. So otherwise like I think the QR should work. If the QR is not working the other is sort of putting that link on the chat. >> I am hoping that others will be able to see. So what we are looking at is you will be able to download a companion again HTML with all these tabs which comprises all the exercises we are doing today. H so be it on there >> I just added in the chat if if you are able to see it on the chat.

32:25 >> No I think people are not able to see. Okay. So was the QR working? I can still put the QR then on the slide. Just give me one second. We're here. >> Yeah, somebody has put in. Let's see if others can see. Pri has put in the link on the chat. >> Okay, >> I see it. I don't know if others are able to see it. >> No. Or is the QR working? I think the QR should also help you with the link.

32:49 >> Yes, QR is working. >> QR is working. >> QR is working. QR should be fine. Then just go to the link download that companion book either you are on Mac or Windows and open it. Let's just spend a minute give everyone a minute and give me a thumbs up when you have downloaded. this companion should look something like this. We'll maybe give one more minute and then we'll start exercises.

33:23 Okay. So what we'll do is if let's say someone do not have cloud subscription or you're not able to rank today I have added the lot of the outputs also right for each of these things. So you can just follow along I will do it on the screen you can just follow along. So one more minute and hopefully everyone should have this downloaded and open on their side along with Claude and we told like >> if you go back to that slide where you have the QR.

33:56 >> Yeah sure. >> Yeah just hold it there for some time. >> Sure I'll hold that. So first I think while we are downloading this we will do some basic u sanity check that if you have a cloud pro subscription are you able to connect to the internet like we said internet connectivity is a must and if you this cloud has access and rights to create folders. So it will do a simple first sanity check the first prompts and then we will go to exercises some basic prompting then we'll do skills and then we'll do agent.

34:31 Hey guys, let's go. >> Okay, perfect. So let's go to the first one. Let's understand. So what we doing here in the first one is just the folder setup. I want to do a basic sanity check. So here you come here. But you don't need to type a thing. Just come here. This is the double boxes. You just click it and it'll get copied. H everyone should be able to do it. Give me a thumbs up if you're able to do it.

34:56 Come to plot and just paste it as it is. So it is just doing a sanity check. >> Sorry I think can someone go on mute please? Yeah. So it is just doing that is it able to go to the internet and get some data for me and if it is able to get a fold out. So it is trying to do a okay not okay so that I can proceed with the next steps.

35:29 So it is asking which folder I should do. I'm giving it a path. I'm just put put any folder like a test folder on your C drive. Anything should work here. All the prompts will be saved. All the skills will be saved. Whatever you do take will be saved here in this location. So see Claude is asking for the permissions. Now one important thing I want all of you to note here on Claude is we talked about different models.

35:55 Here if you see my mouse click here you can see the models. So see I have selected set which is my everyday model and you can see all the model options. Here we talked about haiko right. So you can switch for simple task to haiko. We can switch to purpose for heavy lifting and use sonnet for your day-to-day tasks. And the other thing is the effort right the higher is the effort the quicker your weekly limit trains. So keep it medium. And so this is recommended by cloud also. Unless you have a particular challenge, you should continue with medium for the right kind of optimal budget usage.

36:32 Okay. So cloud will always ask for your permission. Here it is asking can I use this folder? I allowing it. So it is now just doing a basic sanity check that he's able to use my folder, able to use the internet. >> Can you please show that scanner again? >> Sure. Sure. Yeah.

37:04 Able to use the QR. >> Yeah. So I got this Windows setup. >> I copied that. What should I do after that? I should save it as you just save the HTML and open it in any browser. We will just do the all the exercises step by step. We on the first exercise just the setup check. >> Okay. >> Okay sir, you can move on. >> Okay. Perfect. >> Okay. >> Rudra just one question. when we are doing it, we are not able to select the folders, browse the folder. How do we do that on the cloud?

37:40 >> So just from the windows just create any folder just copy the path and ask clot. So let's say you come to claude and just type in hey use this folder like you can just say use this folder. Let's say I have a this path right and then you can just paste that path. So for example I'm using a test folder on my desktop. Just copy that path and say use this folder. >> Okay. Thanks. >> Okay. So then >> can you send the link again? it's working.

38:11 >> So maybe I think I have the link dada can send the link. I am not able to put up the chat window. but can you use the QR? I can maybe use the QR one more. Just try scanning from the QR. >> Has everything got that now? Can we move ahead? We have a lot to cover. >> Rudra, one question. >> Sure. >> I added the file path link the clouds responses. I only run in sandboxed Linux environment. I'm using a Windows laptop by the way, >> right?

38:49 >> Done in a sandbox Linux environment with no access to a drive. >> Okay. So, how do you recommend that? >> Yeah, I think we have to take this individual challenges later. Maybe what you can do is just follow along. Like I said, even if you're not able to run, all the outputs are there. just see the session go through and you can try your clip. We might not be able to do individual troubleshooting now on the call.

39:11 >> We can cater to that later. >> Okay. Now let's let us go back. So I think the first part is done. So that means it was able to go. Now I'm setting the context. So what it will do it it will I'm saying hey we're doing a workshop today. I want you to create folders. I want you to save everything. All the prompts we do all the skills we do all the agents we do. So everything you again come back here nothing to do just copy it. So select the two boxes copy it come to cloud and then just paste it right just enter. So that folder you have selected whatever path it is C test or anything. So it is again asking for the confirmation hey which folder just confirm it. So I confirmed the folder. Now it is able to create quickly all the subfolders and everything. It knows how to save the prompts what is the file format everything.

40:06 So let that happen quickly and you can always see how it is behaving what are the steps it is doing right so sonet was a is a medium power model it's a medium mode so always see it's very fascinating to see what it is doing at the all the types he it is created a workshop folder it created prom skills inputs and outputs for you it has given a summary that hey I was not able to make empty folders I was put a dummy text file in each folder. Here is everything ready for you. So I'm ready for your first prompt.

40:38 So if people you are following along, you should see something similar on your side. Give me a thumbs up so that we can move to the exercise one. Okay, maybe I'll just give one more minute quickly. Okay, perfect. I see some thumbs up. Now let's go to exercise one. So here go to the first exercise. So let's say my before was a very vague prompt like we saw tell me about this company. So we come here we go to cloud and say hey tell me about this company.

41:07 So I have no context I just wrote it on the chat. So like we talked about nondeterministic nature of things everyone will see slightly different output or this one right everyone will see something reasonable but none of the things will match. So whatever is an output on the screen will not match with yours. So again always important to look at what the model is doing. so just don't wait for the output. look at the way it is decoding the problem. Look at the way it is thinking so that you can guide it to course correct later. So it is looking at the is searching the web because I enabled web search. It it went there. It is reading some blog articles. It went to screener automatically. I never asked it to do anything. It is again reading some blogs. Running some commands and coming up with a quick analysis. But it has no context whatsoever what I'm trying to extract here.

42:04 So let's wait for it to finish. It will be I think hopefully soon it four pages and hopefully you're able to follow along. You're able to paste this prompt and I think Mac users there is a separate MacBook and we have tested that with some of our Mac users from the forum. You should be able to use it on the Mac as well. Give me a thumbs up. Any Mac user, you're able to use it. Perfect.

42:37 Okay. See, I think it came up with a good visible answer, but it was not clear what was asked of it, right? So, it came out with the basic summary, what happened in the result, etc. Today, we'll not go deeply into the content of it because we'll just more focus on the process and the outcomes. But you can review and you can play around. Now let's define this. Let's add some structure to it. So here is a prompt where I have a role, contest, task and format. The four things we discussed we would do. Now I'll just paste that second prompt here. Now look at it. You are a equity research assistant. I'm giving it a persona. You are writing this for a retail individual investor.

43:15 summarize HBL's F26 revenue growth, the debt levels and two key risks and in plain English as five bullet points. So it has the persona, the context, so what role it is playing and how to produce an output. So if I now enter you will find a more defined more advanced output. Again it is sort of we'll do web search. will look at the numbers and all. So what we just did, let it run for a second. Let us just come back to the slide for a second. So what we just did is we converted a very basic prompt into the four things, right? So make it a yard stick that I will not write any vague prompt because vague prompt will always give me a vague response. I'll give it a role. I'll give it the context. I'll give it the task. I'll give it a format.

44:09 So if you use this simple prompt formatting, your quality of your response and usage of LM will drastically improve. Now let's go back to see where cloud is today. Okay, see now it has given me I asked it for five bullet point. Now no other what do you see that earlier was very lengy. It has a lot of things. It was not sure what I was looking for but now I told it very specifically. It is sort of says sh profit grew faster debt is low to whisks it has given exactly what I wanted very precise very concise and then it serves my purpose okay let's practice a bit more now we will go again a very vague prompt exactly the thing we should not do in the before prompt is this stock a good buy right now one thing you will notice it will immediately come with a disclaimer that I'm not your equity advisor And if you see always there is a disclaimer here that claude is AI can make mistakes. So be very very careful.

45:10 I think the first response you will see in the same or similar lines is that I'm not a good advisor. Please consult someone. Let's see what it turns out. Reading some pages. Going to trend line. Going to markets mojo. So everyone should be able to follow along with exercise two now on your side as well.

45:50 Okay. So see I cannot tell you it's a goodbye. Sometime it would come with a disclaimer that I'm not your equity advisor. I think it will give for some of you it will not come here the prompt is saved it has come with a vague response now again let's finetune it so let me go to web workbook again now we will do a find a prompt let's come back what we are doing now let's close it you are cautious equity analyst right it took a role now consider the latest available quarterly results for HPL engineering it should go back to a Q1 Less three reason the markets may be excited. Three reason it it might recommend caution as a short table. Give me the output with a balanced summary not a buy or sell call. This is perfect for an LLM because it is not supposed to give you a buy or sell call but you can give it a balanced view measuring the pros and cons. Now here my output is more deterministic because the first prompt everyone will see wider variation of result. This is one way the more crafted your prompt is, you will be able to channelize it towards a more curated output. Tokens are not wasted because remember every output you write is a token cost. Right? So if the LLM is giving a huge lengthy output which is of not relevance to you, you're just wasting your tokens.

47:16 Now look at the prompt. Look at the output. So last output was much bigger, much longer. Here what to have is three reasons the market to be excited. three reasons for curation or caution and per balance summary. So this is exactly what I wanted to know. Not something vague as it is a good buy yourself, right? Because that is completely based on my own criteria. Okay, I think we are done with exercise two. Let's go back to the slides. Give me a thumbs up if people are able to do complete up to this much. Okay, perfect. I see some hands up. Okay, great. We'll go back to the slides now. Okay, so we covered exercise one. We covered exercise two. We saw that a curative prompt will help us. Now, now come to the actual task. So, annual report analysis is a big problem for equity investors, right? We all want to read annual report, but they're very lengthy, very time consuming. I don't have the time in the world, right? I want something quick. Now, understand the magnitude of the problem. So, we will work with the HPL engineering annual report. It's a 260page annual report.

48:26 Now, understand the nuances of it. Each page we saw that how it was broken down token each page is around 900 tokens multiplied by 260 means you are actually asking the NL to consume 230,000 tokens right and the output is just simple the analysis output will be more expensive output tokens are more expensive than input but it's much lesser so in this case when you are reading heavy documents your input is 75 times more tokenheavy than your output Now what we will do in this exercise 3 is we will give it a sort of a recipe a curated prompt where I would want it to go step by step. I'll not boil the ocean and dump the entire 260 page to it. I'll give it a very simple thing.

49:13 I'll say hey can you through this prompt read five different aspects or create a scorecard for me segments a capital allocation red flag and guidance. This is just for this example and I'm I'm giving some ground rules that consolidate it first keep the output under 800 words so that we do not end up with the burning all the tokens in the session always mention exact four figures and if something is not present just do not hallucinate put that not disclosed right so and I've also we'll like we saw we will always use the HTML example right so output should be in HTML should have by tabs and so on now let's Let's go back and see in our notebook that what we can do.

49:57 So let's start here. So first thing you have to do is you have to save the Android report. So I think hope people who are there you just right click this. This is the link for the admin report and click the save link as. Hopefully you're able to do that. So this is the document. So I go to my test folder. You all should see similar folder in your desktop. It has created this AI workshop. under this it has created all the folders in the input I'm saving this file right so I just hit save so my first part is done you all should be able to do the similar just right click save as done otherwise you'll try to save now let's spend a few minutes on the prompt so this is a very detailed curated prompt to analyze a Indian annual report so I am giving again the role you are a principal equity research analyst and forensic accountant specializing in Indian capital markets it knows the persona I need to take what I need to read. We have the extraction protocol. How I want you to go through different sections.

50:59 How I want to create the output. how I want to segment this everything. and some constraint like always use consolidated financial first if it is available then only fall back to standalone numbers. zero fluff do not explain basic stuff. I'm already aware of absolute fact grounding. Use exact figures. Do not hallucinate. No speculation. Output theorem. Right. So first here we'll do first copy this lengthy prompt. Do here copy. Come here in claude and just do we will see claude will attach this not in the prompt but as an attachment the prompt. So what I did is I did not copy the first one. I copied the second one. H I came to the bottom of it. I scrolled down. I just selected copy. I came to claude and selected at the paste. And it is now it knows that it is a sort of a very lengthy text file you have pasted. This is not the instruction. Now then I will go back and I will copy the actual instruction which is this to use that prompt. So I'll come to cloud. Now I will paste it here. Now it will be pasted as a instruction prompt. Right?

52:05 It knows that you are looking at the particular PDF to analyze and use the prompt to do that. All with me. So you are able to download the annual report. First paste the actual prompt at the second part as an attachment which is showing here. Then doing the first part which is the prompt. Give me a thumbs up if you are following along. Perfect. So let me just hit. So now you're using that attach detailed prompt which we just curated and it is using that and like again look at it.

52:41 Rudra I don't know how much you've covered and how much more is remaining but is it possible for you to slow down a bit because there are people who might not be as conversant >> okay >> you know so give it a little bit more time maybe a little bit more >> I will do that >> so might spill over to a bit later than 12 p.m. I think I'll try to slow down and >> that's fine. That's fine. I mean anyway we are recording it so people who need to sort of drop off can do >> at least people who are following along give them a little bit. I'm not saying lot of time but a little bit more.

53:19 >> Yeah >> I think just one more pause I think Sanjay you have a question you raised your hand. >> Yeah Sanjay unmute yourself. Sanjay can you unmute yourself and ask the question if you want there's another person Ashwini Aaral again if you have a question can you unmute and ask >> I could not get how to paste it as an attachment instead of the text and whenever I'm pasting it is just going as the text >> okay see because these are two problems what you can do if you're not able to do like just paste the first one okay Then one while cloud will analyze it will say that hey step B is missing you have not given me the prompt in the second then paste the second one. Okay sometime it happens it is normal behavior H. So people who are not able to paste it as an attachment like I showed you start with the first one.

54:20 I'll go back to the HTML. >> can you show me how you pasteed once again? Actually it was >> because see I have done a couple of test runs right. So my clot has this context what I'm trying to do right of the entire script. So it might have already sort of accommodated that using the past memory. But since you're doing for the first time you might not be able to do that absolutely fine. You skip the second prompt. You start with just the first prompt.

54:48 You just copy it here and see here it is mentioned that I will give you a next prompt in state B. If you can see my screen right? So in the first prompt it is there's a clear indication that I will give you a full prompt for step B. The moment you just face this one cloud will come back and say hey I analyzed your prompt but the full prompt is missing. Can you send me that? Then you come here. There's a long prompt.

55:14 There's a shown window. You scroll down to the bottom right. Then you select this copy box. Everyone able to see it? Just click it. Right. Then you give it that long prompt in the state B. It will work just fine. >> Great. Got it. Got it. Thanks so much. >> Perfect. Perfect. Okay. I think I had one more question. Sanjay Kumar, please go ahead. >> Dada is covered. You just go little slow. >> Okay. Sure. Sure. Absolutely. I think I was trying to make it to the time window because we had just two hours. We'll try to go slow. Thank you.

55:50 >> Yeah. True. True. But little slow. >> Sure. Sure. Absolutely. Yeah. I think I see another hand raised. You can take a question. >> Hello. >> Yeah. >> Yeah. Sir, am I audible? >> Yeah. Yeah, you're audible. Please go ahead. >> Sir, can you please cover the process to create agents and automate them in today's workshop? >> Yes. Yes, it is there. It will be covered. >> Okay. Thank you. Okay. So, we had put that our exercise three. Let's come back to cloud and see where we are. Right. So it saved the prompt and it analyzed the annual report. Right. so see it is intelligent enough to see that it scanned my inputs folder. It saw that I had already downloaded. It used the prompt I just sent it as the attachment here. Right? This pasted one and it me. So it is just giving a summary here.

56:45 If you want to see the result, you have to go to the outputs folder where it has created this. And again it will be slightly different for each one of you. So let me go quickly to see if I can bring that file for you. So I'll just bring it. Hopefully everyone is able to see my screen. The output is created. give me a thumbs up if you're able to see is the HTML. Okay. Perfect. See what we asked it to do is we talked about five things right. We didn't ask them to boil the ocean on the entire annual report because we just want to craft this so that it finishes quicker. I asked five things in that prompt. I asked for a financial performance. I asked for the segment deep dive and catalyst. I asked for the management strategy for governance and the guidance. So it was able to follow my prompt and everything was crafted within 800 words. That limit is important because if you want to conserve the token if I have not given that limit maybe if I created a very lengthy document which had taken a significant amount of time in this session plus I burnt a lot of my weekly limit. I didn't want to do that I restricted it. Now if I come back to plot let's read through it is saying that I put the output and it is giving some caveats also right what stands out. So the idea is a first prompt will never work right. So if you are doing it for the first time this is a more of a curated prompt but let's say you started with one thing you then you iterate you keep playing with this annual report and do it let's at the second thing third thing fourth thing so over and like after a week and so you will see that now my annual report prompt is very crafted and curated I'm covering 20 books try it out with two three different companies it working then finalize the prompt so you have to iterate and bring that journey Okay, I see one question. I think I might not have time to take many question but let's send this one. Nabi, please go ahead.

58:51 >> Just a quick question. Instead of saving the annual report into a folder, will a normal prompt not do the same job? >> Yes, it will do that. See, you are here which is more optimized for the workshop and in the prompt itself it is written. For example, if you're not able to save it, it will be able to download automatically for you. If it has internet access, you can give it a path. It can automatically search either from the HBL website or from NCBSC it will be able to download. Just for simplicity for today's session, I have added the link. I already saved it to make it easier for cloud to follow along.

59:27 >> Great. Okay. Thanks. >> Okay. So, I'll give a another minute if everyone is done or people need more time with exercise three. Are you seeing the output? Are you seeing any errors? Okay. So I think once it is generated and like I said it will be different for everyone. I showcase the response I got which was saved and don't worry everything you are doing all the prompts all the outputs is saved in your AI workshop folder.

60:03 >> We have we have called for the car we are leaving in 5 minutes. Sorry, I think I see some bit of a background noise. Okay, >> we just done. >> Okay, perfect. So, let me know. Let's go to the next one. H. Okay, so we were at exercise three. So what we saw from exercise 3 like we discussed we were able to download a annual report give it a crafted prompt and it was able to analyze the annual report based on the prompt annual report for one year right now we talked about skill right now let's see a skill in action so what is the problem with prompt like we said that it will it works for HPL but every time I have to give this prompt and we saw with some users it is a convert some process right that hey I have to paste a link the prompt something can go wrong right I can miss a word here and there it's not a very hassle-free sort of issue right wouldn't it be better if let's say it already saves this process and knows what to do when I just say hey analyze the annual report for HP just this much it would make my life much easier so now we will convert a good prompt into a still so like I said you cannot do it on the first go play around with the prompt fine-tune it, craft it, balance it. Hey, how much time it is taking? What output I want? You want to focus on risk, maybe have elongated risk section, you want to f focus on growth, keep that. So, write it, test it, test it on a few examples, then save it as a usable skill. And once we do it as scale, we'll see what is the difference between the prompt file you saw and the scale it will do.

61:56 Our exercise four is exactly that. Now that we have a scale we will take this package it into a scale now. So it will convert the prompt to a skill. cloud will see annual report analysis as the name of the skill. It is basically a standard operating procedure. It knows that whenever the user asks me to analyze any annual report, this is my method. This is the steps I need to do. work with any company, keep all the file models and rules from size three.

62:28 take the company name as input and save each finish analysis in the output folder as an HTML. So that is the objective. Now let's go to my notebook. So this is the thing done. Let's go back to the notebook. So everybody is following along. Any quick question before we go to start exercise four? I can take one or two questions quickly. If is the pace okay? If anyone is not able to follow along something maybe just raise it quickly.

63:00 I'm not able to see questions on the chat but can take a few. Okay. So, now what you see is sorry there's some hands raised. Pavin want to ask something quickly. Please go ahead. >> Yeah. Hi. very quickly I very often use both mobile as well as the laptop. So if I set this up in my laptop and I'm using a mobile let's say for to on the go I'm somewhere and I use it on the mobile save the skill. So is it automatically save on my laptop next time I open it?

63:36 >> So good news Pravin is I think we can cover it later but now you can run all the jobs from mobile. So claude is cloud has announced some major change this week. So earlier it was not possible you have your laptop needs to be turned on because it runs all the operation in a virtual machine in your laptop earlier. Now this week starting cloud has made a big change they are moving to the cloud. So now once the scale or amp is saved you will be able to operate everything on the core from the mobile.

64:05 >> Thanks. >> Okay. >> Yeah I think I hush you out of your go ahead please. Yeah. So just like if you use chip and all they have many connectors. >> So from a equity research point of view you have maybe for a sector you have skill. So how is there anything as you said you have to okay sure >> h will cover that later in the next >> one second guys I think let's hold on to the questions because there are a lot of things that truth is going to cover.

64:36 So instead of trying to preempt all of that, let's go through this. If required, we can have a follow-on session you know sometime later as well. But let him finish first. Go on. >> Perfect. Thank you. Thank you. Okay. So now I'm hoping that all of you are following along so far. You're able to do the three exercises. Now we're exercise four. Right. So our objective now is to package this as a skill. So just again everything is in plain English. And the good part is so you might think hey Uda you are giving me all these very lengthy things man how do I do it myself just speak to it you don't need to type a single word so come to plot whatever you want to do just turn on the microphone and just speak your heart out hey I want you to do this this this whatever you want in whatever language it follows Hindi even today plot is able to follow Hindi and English other Indian languages will be available so do not feel restrained that I need to understand technical jargon and what you were showing no just speak to it that I want you to build a skill covering these 10 things and speak your heart out it will do everything for you okay let's come back so here what I'm doing is I have there is a great thing what claude has done for us is it has created a meta skill what is a meta skill for example so a layman user would not understand all this complexity what is skill what is that so you don't need to so claude has created a utility within themselves which can help you create an optimized skill. So you have just to invoke it. So even in your day-to-day work right you come to cloud and when you see just if you put a slash you will see all the different plugins they're already there. So when you say skill for example you can see what is skill creator. This is a meta skill within cloud. So everyone has this by default. So create new skills modify and improve existing skills and measure skill performance. So optimization, creation, fine-tuning everything will be taken care of by cloud you don't need to do a thing. Just tell hey when you are using when you are building a skill for me please use your skill creator that's it.

66:40 Okay let's come back to the example. So I am doing the same. I'm asking Claude that hey I want you to use my prompt which I just did to a skill and use the skill creator man because you know it best do it for me and make sure there are some sort of guide rails I want to put and again the 800 watt thing etc. I want to preserve it. So I put those in the guard again what I'll do I'll just copy this and I'll come to plot and I will sort of run it down.

67:09 So it is saving that prompt then using the skill creator. So this might take a bit of time and because it will analyze using the meta skill and always just follow along what it is doing because it's fascinating to see how their thought process evolves the step by step it is doing just follow along. So if everyone is able to do it then this much give me a thumbs up. If you're on exercise four with me.

67:47 Okay, I think we'll go to agents after this. Let's continue here. And if you are wondering what is the output look like, I've already added all the output snippets, right? So if you are not able to run it, you can see. So this will be the kind of the sample output this speed creator will produce. So people who are not able to run anything all the outputs are saved right for each of the tabs even the sample scorecard just open the HTML on your site spend time on understanding what change between the two prompts what we're doing what is the output okay so this is something the skill now what I want to show you till it is running I mentioned couple of things to you right the context window see what a skill is at runtime it just load this There is something in the world of AI which is called progressive disclosure.

68:40 That means entire skill is hidden. This entire thing is not loaded. It just loads this much in the runtime. It loves okay I have a annual report analysis skill forensic equity research analysis of an Indian listed company delivered as a dense file model HTML file use wherever the user asked to analyze review triage or run annual report. So it is basically codifying the shortcuts when to use this skill. It is not loading the rest of it.

69:09 So my context is not bloated. I have just this part loaded in the runtime that the skill is enabled. The moment in the entire workflow you ask for an annual report analysis. It will then use the skill open this skill. MD file go to a step one do the thing come back again. Go to the st file read step two and come back. So this is the fundamental difference why your context is preserved while you're working with the multitude of skills.

69:40 Now let's come back see. Okay, perfect. It has done for some of your machines it might take a bit longer. So it has already done this skill. It has saved the skill. >> Sorry. Sorry, Rudra. You just said that just the skill and using a normal prompt, right? >> The difference is in the normal prompt, it has everything in the context window, right? >> And in a skill, it's just it just takes it part by part. That's the main difference.

70:07 >> Not even part by part. So bisha what happens is let's say I am doing something else let's say I'm researching something else after 40 minutes in my work today let's say annual report council till that 40 minute only this part will be loaded in the context the rest of it will not touched so if I have 20 skills it is just loading the 20 paragraphs and it's maybe a very small 100 tokens each all of the rest it knows that I have a path and a skill where to look up. So basically this works as a lookup. The moment this sort of trigger statements come in do run annual report research annual report. The moment this kind of terms come in clock knows that hey I have a recipe at that point of time after the 40th minute it will go again. It will not load everything. In step three when we were doing the prompt every the entire prompt what we saw was loaded on the context. My context was loaded.

71:05 >> Right. Right. just move step by step and a skill we'll see the skill. Let's just look at the skill for a second. Let me just open the skill file for you. So everyone should see in the skills folder an annual report analysis skills have been created and I'm using a >> you're not not seeing anything. We're just seeing your HTML. >> I'm I'm coming I'm sharing one second. >> So can you send this HTML file?

71:34 >> It's already there. That is what we are downloading. >> See, so I think I wish to your point. See this is the skill which it has written, right? >> If you you able to see the skill now. >> Yes. Yes. Yes. >> So see like I told the initial part the name and the description is only the rest of it all if you see it is very crafted step by step what to do in step.

71:59 It could be very lengthy. I can I have skills which is going into hundreds of pages because it has a sectoral reading what to do in each sector and all it does not load any of it in the context window. It just loads this much. So progressive disclosure means as in when required it will go it has all the mapping now it has all the tools in place. It will go and say just if it is needs to download annual reporting just go to step three and this is how to download annual report right. So that is a fundament to your work becomes much more optimized much more efficient when you go to skills.

72:34 >> Perfect. Correct. Thank you. >> So we'll come back to I think people who are following along are you here? I have done exercise for let me move. >> I joined a little late. I didn't get the can you send it? >> Okay we will share that. I think because let me go back. Okay. This is the Can you just quickly scan the QR? I'll go back to the presentation then real quick. Just scan the QR you will able to get the path to download.

73:05 Okay. And everything is recorded. You'll get all the material. So don't worry about that. You can follow along. >> Okay. >> Okay. Let's go ahead. Now is I think very important. Now we are moving from scales to agents. I think someone was asking the question. So what is a difference like we saw the bot versus agent we have the understanding now right a bot was very sort of a simply task deterministic it cannot adapt it cannot do agent can sort of move beyond summarizing the annual report it can research and company end because annual report creating is not the end of it right you have to do multic parts so what we'll do is now let's dive deeper into the sort of agent it has a four part. Whenever you tell him, hey, I want to research HQL engineering, it first plans. So the user what is the intent? Again, if the more curated you are in your prompt, the better for it to plan. If you give it a very vague idea, it will not be able to plan. It will normally execute it plan. So it can use a lot of tools. It can have access to different APIs. It can have access to different connectors. For example, if you have all the data BigQuery on on Gemini, you can it can use BigQuery to fetch your data. You have plugins for different market databases. It can connect to that bring data from there.

74:27 It can scrape the web. So it has a different set of tools. So the agent will plan intelligently that what kind of tools APIs I have access to. based on that I will at that I will continuous check what the user asked for what was the guard rates it provided it sort of working on that sort and it will always do course correction. It will repeat till the final output is reasonable. So agent who works more autonomously, how much autonomy, how much agency will maneuver.

75:01 >> Sorry, I think if someone can go on mute. So how much agency will give, how much autonomy you will allow depends on you. You are in control. You are the human in the loop. Like we discussed whatever you ask for it will. Now what is the benefit of the agent? Right? Now we talked about the particular task being first done by a prompt. Right? We were all good right? We I gave a good prompt then I converted it to a skill because then it becomes a full recipe what to do at the full structure. The standard operating procedure is laid out. It can follow piece by piece. Now where the skill or agent helps is now it can sort of connect the dots. It can go from one skill to the other. It can see hey in this part of the flow I will use this tool in this part of the tool I will call this API and everything is connected. It can now share the context between multiple calls. So the knowledge cumulatively builds up. It is able to do something much more complex which was not possible maybe just giving a skill or just giving a problem.

76:08 So what we are trying to do here is we will practice a simple paging chaining. Right. So what think of it a very simple structure for your research desk. You have a intern or a equator analyst. you give the simple task hey can you research hail engineering. Under the hood there are three other analysts working. Someone is gathering the data. Someone is analyzing the data. Someone is producing the final report. So we will try to build now an agentic system that will do all three.

76:38 So I need to do a researcher that will gather facts. I need to give you a very detailed curative prompt right to so that that prompt is the at the heart of the planning for it to understand what I'm actually trying to do. I will have to guide it to do the annual report analysis and then write the final report. So here let's say if you do not have this system and you start with write me a quick research prompt right let's start with that and we'll come back to the prompt.

77:08 we'll go to the notebook and see but before going to the notebook I just want to give you the structure and what we will the output will differ so in the after what we will do we'll use three different occasions one is a prompt generator that will extract the detailed prompt from that particular company and the way I want to write it it can work for every company right I don't want to tie it back to HP engineering it was just an example so I would want a first agent to do a prompt for Second agent we will just reuse what we just did. We will use that skill annual report analysis that is already ready for me. I know when I get an annual report these are the five things I want to see. And the third agent is a deep research agent. Now I need a orchestrator which is my front desk analyst which will chain all of it.

77:58 So in the output we'll go to the cloud shortly but what we're trying to do is we are trying to do a prompt generator. So these agents are hidden probably. I don't need to talk to them. It is my not my job to talk to it. I am introducing my research analyst called Arya automated research and field analyst. So going forward I don't want to get into any complexity. I just want to talk to Ara help research this company. That is my goal. I want to spend my time reading through the research. I don't want to go to the complexity. How Arya manages with these agents is left up to her. So now let's see if we can do it. H so before we jump to claude and actually do it is everyone clear what is the goal here I can take a couple of questions in the goal here and somebody has question in terms of moving from prompt to skill to agents now let's let's take a couple of questions quickly anyone is it clear okay I see a lot of thumbs up no specific questions so we'll move along okay so like I said let's first move to cloud and let's start with our companion. Okay. So, first we will start with a very basic vague prompt, right? Let's see what again we can skip this one now. If you'll see that if you put this kind of a prompt, what comes back is just a very basic thing. It will not workable. We will just directly jump onto this. So, let's do it together. We will step by step. What we'll do is first create three agents. Agent one, which is a prompt generator. Agent two is annual report analysis. Agent three which is the researcher and area will be our orchestrator on top. We'll do it one by one. So let's start with prong generator. What we are doing is I'm asking that hey you are a prong generator when run research and cover which industry sector the company belongs to industry's tailwinds headwinds and relevant government policies. Company's key business segments highlights and capex and growth bands. So I am giving it a structure which is company agnostic. It need not be anything you see here is nothing to do with the Hville engineering. It was just an example. So and this is just a very very basic problem right you can start with this you can finetune you can instead of five you can add 20 things whatever you want to see in your research problem. The more specific you are the more curated this is for you. So for now I'll go with this basic one. I will start by generation. So I again copy this agent one come to cloud and hit the paste button.

80:38 So it will start creating the prompt for me. And again look at what it is doing. It is trying to set up the prompt generator skill file because it knows that you are trying to create a recipe. You are not trying to get an action done. So it is creating a recipe and not an outcome. Go ahead.

81:10 You have a question. >> Yeah. So, sir, do we need to have the paid resources for scraping and you know API calling or let's say for coding let's say in in open air there is a codeex. So in this claude do we need cloud core and all those things python and all >> I think you don't need it this session is meant for let's say people who are non-technical users right I'm not going into that but short answer is you can do it you don't need any paid license I will show you the refined output in session three which is completely free of any paid licenses so you can you'll understand what is the possibility here yes AIT go ahead in our response is if you go to that skills folder you will see it has been saved >> okay thank you I think let's take couple of more question then I will move on yeah Vishal Ashwini and Sata quickly >> hi Rodra thanks for the question normally when I you know open claude it asks me two option. One is the desktop version and the second is the online version.

82:28 >> Mhm. >> So any preference over there in terms also in terms of security and everything. >> No no I think like I said there is a major change this week. Now you can do everything in the online version as well which were not possible till last Sunday. Till last Sunday cloud desktop was the only option but now that is changing. >> Okay sir. Thank you. No >> Rudra one quick question going back from agentic AI to skills >> right >> in one folder for annual reports I saw a folder by the name evals it may be a little bit off track question but I heard this term evals in the context of AI can you just spend a >> I think it is more technical so we can take it offline >> okay take it separately >> yeah yeah yeah okay okay >> ludra the thing is that right now what I understand is once we make that skill it works like an application sits on my computer even if I'm offline everything runs with the with my system only the processor of my system only works for doing the computation is that right >> that changing that is changing so earlier what is happening is the cloud was spinning up a virtual machine a isolated environment within your own windows or Mac system and doing that from this week they are shifting that they will now run that virtual machine on cloud so that any workflow any skills you have can be now accessible on the go as well.

83:52 >> And what about the output files and the data that we create that will also be saved on the virtual >> these virtual machines are ephemeral they do not exist beyond the session. So you have to give a cloud path let's say a one drive or G drive where it will be sort of written to and then when you resume cloud co-op you are back in your laptop you can ask cloud hey whatever you've been in the last session copy from my cloud to my local a best practice is all the folders keep it directly on the cloud like my entire cloud work is in one drive so I can access it from anywhere >> got it got it thanks so much >> yeah I think let's see the last question maybe we have to move So just one thing can we have an example of how to use this skills?

84:40 >> Yeah. Yeah. We are using that right. So we'll see that video. >> Okay. Sure. >> So here I created if you see now I created a prom generator skill which will back my this agent which is my chrome generation agent. Let's quickly come back and do the other two. I will do the second one. So here I think exactly to the question I already have this skill so I'm saying that wherever if the skill already exist if one already exists reuse it do not redo the work so because I created a skill in step four it would automatically reuse it and this is very important in your prompt is called progressive update that means whenever you are using a skill any changes happen claude is instructed to auto update the skill so over multiple iterations s are three to four weeks of work. You will see all the skills evolving.

85:32 So I'll quickly run this one. This is my second agent and then we'll come back to the agent three. Let me quickly copy that. Go back. Okay, let it run. Let's come back to the deck in the meantime. So we we have done the prompt generator. Now we we did annual report analysis and then we doing the equity research, right? So all of it in is going to ar and now once you have this entire thing set up if you are able to run all the XY 1 to 5 from tomorrow if you open clot you will be able to use claude like this area please help research company any company under the sun right you can say research company this area check the report so we abstracted all the complexity of prompt scale and agent into a multi-agentic orchestration now where claude is able to manage all the complexity and you are just able to talk in single line. This is the power of moving from a from to skill to agent pick.

86:36 We will see that where we are. It is asking for some question. Which company should I run the angle code analysis on? So, HBL and it can pause and always come back to you for specific questions when it is not sure. Okay. Then I think we will do agent three. We are running. Let's see what we are running. This is agent two. Okay.

87:09 Anyone facing any difficulty as we are running this? We are in agent two now on able to follow along. >> Rudro is asking some questions like I checked the existing HPL report against the new error checks and found three issues. >> Yeah. Yeah. >> Chairman's commission versus legal cap etc. >> Yeah. Yeah. So we are not focusing on the content. So exactly the same similar thing you're seeing here. So for now we'll skip that just for the posity of time. I'm just focusing on the skill.

87:37 Now only this step you can play around and fix everything. >> Skip and move. >> Yep. Yep. Skip and move. So claude is very interactive like as you will see as you start walking and it is so powerful now. It will point your mistakes also. If you have missed something it will probe you deeper. Hey you have missed this one. So it is fascinating how sophistry this has become now. So let's it runs let's go back to some of the examples right. So at the end of it it will create a multi-tab sort of a HTML. this is the report from area.

88:17 Right? Now remember we talked about that talk to now. So this is the area which is created. Hey, this is a snapshot key metrics. So today after all the the things are done, you will all see a report like this created by my three agents under the hood which you working in partnership with Arya and going forward this is your sort of a cheat sheet that you just open cloud and say aa please help just copy paste this thing keep handy with you so that this becomes your cheat sheet for the future.

88:49 Let's come back. I think it's still running. This will take a bit of time based on different machines and all because this is complex thing we are doing. We are building agents and analysis. I tried to keep it as optimal as possible for the live session. But when you do real stuff, it will maybe run 40 minute 1 hour also in some agents. Okay, perfect. okay, so let me go to the next one.

89:20 What is the third one? Okay, I see the equity researcher one. Let's come back here. This is the third agent I'm creating and then I'll put the orchestrator up. Yeah, I can take a couple of questions till cloud is running. Yes, >> go ahead. >> Hi, Rudra. conceptually an agent is basically a skill is it? That's what I can understand from the prompts. there there are multiple other things right so like I said agents as you start building agents it will start building an orchestration it is build start building an operating model there's a lot of supporting file agents have their own empty files okay so it has a set of instruction because different agents have different persona right it could be an orchestrator it could be a subtask there could be sub aents we'll see that in session three >> okay okay >> yeah vina go ahead >> yeah hi Right. So earlier in the first session you mentioned that we need to build we need to use the higher complexity jobs using opus but right now I see that we're building agents using sonnet only. So what would be the difference in the output which we would get from say using sonnet for a complex task and using opus >> right?

90:40 >> Mhm. Go ahead. secondly I understand that there is one feature in claude wherein we can create a project we can give it a sort of a brief that this is the sort of work which we need to do in the project and then incrementally when we chat within that particular project it makes the task easier. So what would be the difference between doing work by creating a project and by using an agent? >> Right. Right. No great great questions both of them. So for this session I'm restricting it to set right now if I go to the slide right. So let's say if you are doing the particular nature of task you can specify in the agent recipe itself that when I'm doing remember there were three agents right so you can specify which agent which model to use.

91:30 So I can say hey in the first one first one is basically a prompt generator right agnostic of the company you can say hey go with sonet and you can see for agent three always use opus that you can specify opus will burn the token much faster in this session so that I have kept everything on sonic for now but as I'll show you the examples data the deep research done in opus the difference is opus is able to distend and like think much deeper analyze things much detail it is much slower the updates are much richer it will take more time. So based on the task you can specify and sonnet will automatically cascade to the model and come back. So your orchestrator can like is always running on set but it will trigger the first two agents on set and the third agent on the opus that is a question one. Question two you asked about the project setup and everything that is right. So what helps if you see these are like the pinned ones are the projects. So projects help you to share context and memory. I don't want to disturb any existing work any individuals might be doing. That's why we are running sort of a dry session today as a test setup. towards the end I will send the link for a two-hour claude deep dive training that will explain everything from project setup prompt skills and even there is something called plug-in. So all of that you can go through that but that is a very technical training session not might be relevant for everyone that's a sort of a read from this session.

92:57 >> Got it. Thank you. >> Okay. Perfect. I think this is done. Let's quickly move on with the third one. now we have the three agents ready. We will trigger Arya now. So here if you see that now all the three are ready and I'm enabling my orchestrator which will now do an all the things. So once it is done for everyone you will see as I already show you the example and I would encourage all of you after the session play around with this area is now ready.

93:35 so try it out on the any other company you want it should work fine. I have run the test. I think there was some examples I ran. I think I ran for some other company and it it worked fine. Okay, let me come back here. So this is the output we are expecting in your output folder. You will see this kind of an output is now created with a multi-tab HTML and we'll see that result.

94:11 Once output is created, I will see and I think we are running a bit late now for session three, but we'll quickly go there. Okay, let it run. Where are we? Okay, so area skill is saved and research is finished. Okay. So this will take a bit of time but I'll just run it so they can all can just say. So area please help research engineering. Now area is ready. So I just trigger off this.

94:49 See now it is recognizing what you do. So earlier remember I had to give a ton of prompt and everything right? So now I it just boiled down to a simple one line. So I can research any company under the sun doing that and here it has already identified that running area now on a engineering the chain is a research prompt then annual report analysis then the final report first I'm saying so you will see it is breaking down step by step calling each agent so it'll done on fully autonomous way wherever it will pause it will have a gap it will come back to you and ask for your guidance okay let it run and when it ends you all know you can see kind of the example in your site. So that sort of concludes our live session. we will move now to the session three. I can take a couple of questions if they're there before we move on.

95:46 Ashwini go ahead. >> Go ahead. Yeah, RDRA thing is that I actually accidentally was working with Opus not Sonnet and the first two agents have developed given prompts in opus only. Now if I switch to set now for agent 3 and orchestrator does that work or now it'll have to start again in sonnet. >> No no no it will you can change model every block every doesn't matter. Okay.

96:20 Okay. >> Okay. Yeah. >> And it's taking a little while like it's taking a long while. So play around. I think you have to this this session is just for us to get started for you to move from this journey from prompt to skill to agent and then to multi-agentic orchestration. Now that you have the full setup ready, you have all the prompts. You can just delete everything and start a fresh in your leisure time. Go step by step. My recommendation for everyone is to start looking at what it is doing right. It's not about getting an output done. So very minutely follow the things right. So see I did not ask it anything after my area is ready. I did not ask it to do anything. I just said a few word area please help research that's much it under code automatically broke it down into the three steps. It is saving the problem.

97:08 It knows what to do. It is waiting for one agent to complete come back to the second and going ahead and and doing all the steps one by one and it will produce the final output in HTML the multitab one I'll not be able to take so many question because of the time we're already a bit late on the schedule but maybe couple of questions the people who have asked before maybe give time to others if they have fresh questions >> hello hello >> yeah go ahead >> sir so you will be touching upon the orchestration part right after this.

97:41 >> Yeah. Yeah. This is the orchestration part we discussed, right? So, see what we did is we first the three agents you ran, right? >> They different agent they are basically a goal oriented agent. First prompt genator agent's task is to create a curated prompt for that company. Whatever company comes to it, this guy's come work is to build a annual report analysis for whichever is the company. This guy's task is to build a equity research from that given company right they are good oriented agents Arya is itself an agent so Arya's task is to coordinate so Arya is your talking point he's assistant so when I said Arya go and research so let me when I going to this one right Arya please help research the company it is running the full pipeline because it knows my intent that user when they're saying please help research company this command or this tagline triggers of the entire chain when I tell Arya please check I'm not calling that agent annual report analyst anymore I'm saying area check the annual report it knows that now the user is not interested in the full research only annual report so under the hood it is just triggering my annual report analysis similarly just triggering my prompt generator or just triggering my refreshing my research so with examples with more usage you can fine-tune we'll see in the session three >> okay sir Just a last question. So it is not related to this particular thing.

99:11 So let's say if we want to attend the multiple conferences online let's say so the equity conferences. So can we create an agent for that also so that >> that is possible but beyond the scope of today's workshop we can take it offline. >> Okay. Thank you. >> Maybe one more question before we move on. >> Hello sir. Jagish here. when I try to run this I joined late u but when I try to run the second one I mean the annual report analysis agent two do I need to upload the annual reports?

99:46 >> No no no nothing because we HPL we are just working with the HBL example which annual report is already there. We have already have this in the input but for new company runs it might ask you that I'm not able to find this in the input folder. Can you download or ask? It can automatically download for you. >> Okay. So for that do I need to input anything? >> No, no, no. Like I said, you if you think that it is very difficult to find annual report like some annual reports are, you can pre-download and save it in the input folder like we did in the section.

100:17 >> If not, if it is easily locatable annual report, cloud will prompt you, hey, I don't find the annual report in the input folder. Do you want to download? You can say yes. >> Okay. Okay. Yeah. Thanks. >> Okay. Great. I think we have to move on now. >> Just one last question RDRA. >> Mhm. >> what I have seen with claude is that for certain chats it has got certain limits like number of PDF files that I can upload to a chat. So in your skill what I saw was that you have created a folder in which we have given the input and output folders from where it is taking all the information. So does this help in bypassing those limits of the cloud of no >> no no that's >> what we did is like I said me mentioned in the annual report analysis slide right the limitation is the token window right so when you're trying to read the whole annual report it will bump into the limit right it is nothing to do with the folder structure so what you can do is you can specify in the scale like we did in the annual report just go section wise do not read the entire section okay only go to the management discussion analysis part only go to the consolidated financial summary only go to the balance sheet so if you specify >> I guess what I'm trying to understand is that in a chat when I have uploaded multiple PDFs over the period of time in due course maybe not in a day over 10 weeks so after a certain time it throws up a error message that for this specific chat the number of PDFs that you could have uploaded you have reached a limit like 100 PDFs have been uploaded >> I think no got That's a different limitation of the cloud desktop app itself. But then my recommendation is do not upload any file.

101:58 >> Let it like always try to think just give it a recipe to your assistant. Do not solve the problem in the whole like here we could have just uploaded the annual report right we did not do it. We gave it a prompt so that here it will always look into the folder and the folder part is better because you can selectively pick and choose right like you can with the same area now you can research any company under the sun and you can host 200 annual reports in that input folder you don't need to load anything right so you never bump into limitation because at a specific time area will just load that one annual report and nothing else >> right right thank you sort of Go ahead.

102:42 >> My question was a little similar to what Chandan asked. >> Okay. >> So sometimes what happens is say in case of smaller lesserknown companies we generally start off our research after we read say an IC report or a report on the sector you know that is very good very exhaustive that has a lot of information a lot of inputs about what the company's plan is over the next few years. So if we have say one or two reports on that >> then how do we incorporate that or how do we use this in this recipe?

103:15 >> Yeah. Yeah. So we'll see that sort of I think was not possible to do everything in this session because it's just very tailored for to everyone can run it but we'll see that in session three for sure. >> Okay. Thank you. >> Okay. I think with that we are good. So I think this and see it has created. I think everyone can see my screen now. I said my area sort of did all the research for me. It sort of used the three agents underlying and it produced the this kind of a five page summary right snapshot key matrix segment competition and risk outlook.

103:49 So that is the power of agentic now and I can run now for any sort of company. Okay with that now I will >> your screen is not shared. Sorry, I think you not able to see that. I >> No, no, it just shared the shared right. Okay, now it is. >> Okay, perfect. So, I think that ends our claude one. So, like I said, you have the sort of companion books with you in Mac and Windows. Keep running it, keep experimenting it and learn. Now, we'll go and see the art of the possible.

104:22 Okay. So, we did here. Now what we just practiced is a good prompt has a role context task and format that card is there save it print it in your desktop always use for prompting save prompt and skills skills moving to agents and then chaining specialist agents to a doing everything. So now sorry >> hi so I mean if I want to start research on a separate company is it in do I do it in the same folder or I copy this folder and >> currently you can save and pin that session whatever let's say let me show you quickly >> because I mean right now I just ran this on another company so >> all the all the prompts and all got saved in that HBL engineering folder all everything got saved in that folder. So even let's say if cloud is not able to find it you can just say hey all my skills are in this folder can you copy it will automatically copy all skills right it is very very flexible good practice is if you want to play around just rename this like I've renamed this chat to workshop you can just pin this chat and it will be with you so you can and like I said there is a training two hours training it will show you all the best practices from project how you can use folders everything will be there thank Okay. Okay. Perfect. So now we'll continue with the slide deck. Okay. So let's now I think a lot of questions on the art of the possible. Now let's say what you show like I showcase today is a let's say very basic version of it.

105:58 Right? Because we did a quick and dirty. We sort of did it in 1 hour. So we learned from prompt to skills to agent. So multi agent. This will take time. This will take a lot of fine tuning. you it has to take your persona your way of doing research. Yeah, your perspective could be very different for the all the 100 plus people here. Now I will give you a glimpse of what I sort of developed. It is sort of just my pet project. So here like was also mentioning that I can start with some sellside report. I need to do a lot of other things from annual report to earnings call to credit rating to filings right it's not annual report is not the only source I need to analyze everything I might need an agent who will just create the recent news for me I might need an agent that need to go and check what are what is the company sort of publicizing on its website for example so just expanding the same thought of area now into a more broader more mature profile and what I arrived at There's a sort of a area now dealing with a junior analyst. So there are five analysts from data collection to research prompt deep analysis research and thesis tracker. We'll see one such example live on the screen. So what each of them are doing is data collection can download a lot of files. It can I think someone was asking about scraping. So there are some open source Python packages you can do for scraping or downloading announcements etc. You can hit screener to get download some of these data points. A very detailed risk prong generator.

107:36 Then analy analysis agent is something where it sort of breaks down each of those components. first reads annual reports then read conference calls, credit ratings, exchange files use and so on. All of this and now if you see the approval gates these are your human in the loop gates. The two boxes you see. So let's say my annual report research is incomplete. If I pass on that chain, it will be a faulty analysis. We cannot do that. So we have to pause. We have to iterate. Hey, you are not able to this annual report analysis was not good. Can you redo it?

108:13 Right? So in many such cases you have to inject the human in the loop break points where you can manually sort of validate give it specific prompts maybe required update the skill update the agent that hey this is not how I want you to do annual report analysis go deeper do this do that slowly the system will evolve when you're satisfied then move to the deep research agent and there's a thesis tracker right what are the key items I need to track and monitor for this company going forward now Let's say if these are good and you are sort of it invokes your interest and you say hey this looks interesting what about the valuation because so far we have not touched valuation. So then I call my theme specialists right there is a valuation expert. So these are like think of you the area has this kind of on call panel who are equity market specialist. So a guy deep diving on valuation there is a forensic analysis guide. someone will do a very deep peer group analysis. Someone will do a very deep technical analysis. Someone will leverage the capital allocation framework and recommend what what kind of allocation this stock should bear if you are going to buy it. And very very interesting I think this was talked in one of the forum meetings. Can I talk to some of the market veterans on this stock? Can I talk to Prashant Jan? when I talked to Kenneth Andrade on this talk. So this is basically this equity expert panel is harvesting their persona into an equity expert panel and we'll see this example in one skill. So what is it is basically then you are doing it just the expansion of what we saw in area code in a simp form.

109:58 You're expanding it to the first section 8 to H are all these different areas. Then you augment it with the valuation foreign sake peers technical capital panel and tracker and then when everything is done then you feed that the entire output to someone who is generating a initiate coverage report which is exactly the kind of salad reports we read before showing you the example I'll just take out some quick questions is the structure clear any doubt anything you want to Anyone?

110:36 >> Hi Rudra. I have one question. >> Mhm. >> How do we keep track of the tokens while working with agents and on call expert agents? >> Right. Right. Yeah. Great question. So again the your training the two-hour training I talked about which is there later will help you. So often you will face limitation because the cloud pro license has a 4hour window right or often your analysis will be stopped midway. So you can just me mention a single line prompt saying hey maintain a runbook. So your orchestrator agent which is area will maintain a runbook till what part you have done. Let's say one for analysis can span 8 hours let's say. So it will track till what part I have done. it will save all the interim outputs and resume whenever your token window your usage window opens or let's say your weekly limit expires and Thursday 8:30 it again sort of resumes for the fresh limit for the next week it will automatically prompt you to sort of resume the task so there are ways to deal about that >> but is there a way to see how many tokens have been used let's say we do one run and see >> even even in this one like even in the test runs als Also you will see in the the companion book it is giving you a measure of how much tokens roughly used based on your pattern but say for daily users like us who have a cloud pro license I would say that we have to just optimize things so if you start using something like we did for the annual report today I kept it a buffer of 800 if when you are starting an agent you speak freely to cloud like we discussed now in the first run if it is taking too long a time you ask lot can you analyze why it is taking so long how can I make this agent run faster so over this iterations you have to you can optimize and make them more stringent you can switch between different models but some ages need to run because if you break the analysis in between like peer group or valuation you might not get the full picture >> got it thank you >> kaja please go >> yeah hira what guardrails have you built in Yeah. So that it does not hallucinate what type of prompts or skills you have given.

112:53 >> Right. Right. So I have a slide for that. I I'll show you in the attendance there when you share the deck. Maybe I'll just quickly go there. one second. Yeah. I think these are the kind of guardrails I have built in all the skills. so grounding in the source allowing I don't know citing. So you'll see the example also and verify it for finalizing and use specific file links plus you can audit one model's output with the other right so I audited those output with Gemini let's say whatever the output cloud was creating using opus I sort of audited that using Gemini that is also a way you can build but then you have to go to multiple subscriptions which I don't recommend for everyone if you are not heavy user of these AI models every day h so I'll Just come back to the slide.

113:45 yeah. Okay. Sorry. Yeah. Let me just show you the art of the possible now. So that will give you some clarity where we are. Let me open that. So let me know if you are able to see the screen. >> Yes. >> Okay. Perfect. So I again will not be able to spend time on the content here too much but I'll just give you the kind of research that is possible. Now this is without any input from me right what I just talked about is completely automated and completely agentic research. So I did not write a single word in this. So this army of agents the tab a to h you saw here like I'm just going back to the s slide once. the army of agent this is basically areas run and the green boxes are all the SMB panels right so if I go there so it started with an executive summary where the stock is today it is a very detailed and now I think Raj you were asking it always ties back to the fact that this is the list of things where have I picked this from so if you see it has read the last five years of annual report it has all the confidence calls so wherever it is finding a conflicting source it's also highlight that so that I and go and validate. I did not need to take it by face value. all the information it is doing revenue and profit growth, segment by segment breakdown. Then it is going to the competitive positioning comparing with ADOR, ESUB and other companies. what kind of durable competitive mode it is having. again these are all sort of I have been fine-tuning and crafting this and some of the forum members have also helped. So thanks a lot to them to support with subing prompts. Then management and governance it is doing a very detailed analysis. what are the headline governance flag the CU departure it is flagging it's a very big risk other governance items it has picked up from the annual reports shareholding patterns then it is going to capital allocation what was the last five years have been again using a lot of data from screener from annual reports what is the capital allocation track record and always grounding in fact right it is always quoting where it picked it up from then valuation and forecast it only comes when the valuation tab is done since in this case the valuation tab was done. It was able to refer to it. Then highlighting of the risks here. What is my risk register?

116:15 What kind of things is there? And very very important is the sourcing and verification. So this tab is a detailed layout of what could not be obtained what are the red flags? what inconsistencies were found during the process so that you as an individual human in the loop in control can go and validate and sort of play around with every information. So sort of I think you have a question in terms of the cellside report. So in this kind of a flow I've also put a input folder as a cell side. So my deep research agent is also using all existing style side report alongside all the inputs. So that way it is sort of already assimilated. Now these are the expert panel reports. So this is all the valuation. So it builds a deep update DCF or whatever you want to do.

117:03 it will do a football field analysis of the entire valuation from a different VCF beta OC every all parameters it will give that then the forensic analysis is again very detailed. It will come up with the score what are the different categories what kind of what was cleared which is a nearness pattern matches. Then the peer group analysis is a very detailed industry sector comparison sector overview. What are the gaps? What are the levers but is a benchmarking with all the companies and then technical analysis. It will see the momentum where it is in which stage. It will do a deep technical analysis. It will plot on the charts.

117:45 It will analyze everything. Then go into the capital allocation. This is built on the capital allocation framework in value picker which Donald has created. It's a very great sort of an asset to have. So it is already trying to see what where should it be put based on the valuation different sort of scenarios and then going to the expert panel. This is I think the very interesting agent I talked about where I'm talking to Kenneth Andrade, Prashant Jan, Samitak I put six personas for now. You can add any other investors you want how they are sort of evaluating on the different categories and what is their verdict. So you are basically virtually talking to them and see hey would they invest to this stock at this price and here's the verdict for them. So once all of it is there then you move to the actual initial coverage because that is my sellside report. So let me know if if that report is visible to all of you now. So this is not a long sellside report that 30 40 page reports we typically read. So all that backend work is for you to validate and analyze and and sort of brainstorm, refine give its inputs. Once all that is done, it will now create a very so in this company will not go to the company or the verdict. It's just for demonstration. It is giving a sale recommendation exactly the way sales and reports are structured. Market data, financials, valuation, what are the four pillars company at a glance going it's a very detailed sort of exactly the kind of initial coverage report you read. it has sort of explain everything and if you it will give the capital allocation the promoters foreign sake and everything and you can extract this as a PDF. This is not multitab for a reason because we like to read annual report initiate coverage reports as a PDF. So you can extract all of this, save it as a PDF, distribute, read, share. So the great part is it's very very detailed. I think very exhaustive and I had shared with Abhisheka and some of our analysts in that team. I think this is very comprehensive and what you can do is it saves a lot of time for you to read as individuals we want to read and not spend time in data gathering right. So as a investor that is our end goal. So I just want to show you that if you pursue on this path we just started today you can sort of reach to this kind of a mature multi-agentic orchestration with these agents where it can create this deep research effort and the initiate coverage portfolio you okay I think we are just nearing the end of the time now again this is just a summary of what we saw it covered deep research valuation six peer group everything it is giving a clear verdict where we part and the initial coverage came as a sale. Now you might agree or disagree with that and you can add things. Now my one ask from this group would be okay again this is just a summary the deep research the initial coverage what it sort of covered what could be the next agent because you saw I'll share this deck with you my one request would be that you can come up with recommendation what new agents we can add in this pipeline we can add that and my request to all in the audience would be if you can just spend 30 seconds scan the QR and give me a quick feedback so that would be very helpful for me so that like Vishinda mentioned we can come back with a better session next time. that be lovely. I think this QR should be easily accessible to all. Please request all to just spend 30 seconds. please give your quick feedback and I'm happy to take questions.

121:27 >> hi sir here. >> Go ahead. sir we do all these research on the company based we also have impact from global issues right so do we have any space where this particular company may get impacted because of global issues like crude oil >> yes like mentioned you are only by your sort of scope right so this is a fully automated and fully expandable sort of a structure. Whatever you want the game to this open source they can go into you can do geopolitical it is already taking in a lot of external headwinds and all into consideration in generating this report like we saw for the diffusion engineers case but wherever you have to play around with it right it is never perfect from day one but more you curate it will more be fine tuned for your need >> yeah okay yeah thanks >> Sure Rudra Anand here am I audible Yeah. Yeah. Please.

122:34 >> three questions. Oh, this is fantastic. First of all, thank you. three questions. So, one is you mentioned about the investor verdict, right? And you created those personas. How is that working? How how how is it getting into the psyche of those investors and giving a verdict? That is one. >> Yeah. >> Shall I ask all three? >> Okay. Let's let's take one by one. So, I think that agent took a long time to build, right? So what it does is when I give it a six persona so I gave it a very detailed prop that I want you to follow their recent podcast I wanted to follow their recent interviews. So it started building the personas. I think the six different personas is a very detailed files for each investor and like I say if you want to add new it will go through and do that detail. It will watch YouTube videos it will read their transcripts and build how what are the things they prioritize on right now. when they come to the SME panel for discussion it is consuming the deep research report already generated with those persona those prioritization are coming up with a verdict >> understood and does that also consume the consume tokens of our yeah okay >> of course like everything all the like these panels are very heavy like so I don't recommend running so your part should be like I showcase areas should be focused on the deep research once you have read the deep research If you want to probe deeper then only you should call the othersmemes otherwise not because all ofmemes are very token heavy and all run on opus.

124:05 >> So Rdra but then you you mentioned that this skill is already now built with you for example. So can this not be readily shared rather than everyone consuming token tokens to build those persona. >> Yeah. Yeah. We will see this agentic system I just showcase is still in the working progress mode. So I will plan to share like they will discuss how we can share this. We will plan to share it in the future. Fantastic. Thank you. the second thing is the report that finally came out right it has loads of data. Now how do I build trust on that data? Because what I have in my limited experience seen is that you know this data sometimes is also you know not not very accurate right it hallucinates it gives an appearance it's perfect but it is not so in your experience how do you do that right because there's so much data how can you valate every single data point there? So right right so there is no sort of replacement on say let's say human analysis right the purpose is not to but let's say this kind of engineic workflow can analyze 20 companies and you you like like three or four of them and in those you will maybe deep dive the overall thesis is roughly right we have put enough guard rails right wherever when you spending time developing this kind of a system at every point you keep it very factual very fact oriented ask it to explicitly say no I Oh no I can't find this right that is the one guardrail and second like I mentioned you can have it audited or real fact check another model so I fact check it with Gemini so there is an audit model in running in Gemini which reads and finds all flaws >> so they kind of fight with each other till Gemini is not able to find flaws I sort of fine tune it so that kind of structures you can use >> clear >> and you have to spend time with your first few like companies researching end to end to finetune Understood. Third and last point is because this is expensive in a way right because when you do for retail investors small portfolios how much can you spend on these various tools and other things. So have you tried doing this kind of things on any opensource open weight type models alternate cheaper models which provide it which provide a comparable level of output right maybe not as sophisticated what cloud can do cloud can do but that could be more costefficient for a retail investor.

126:20 Absolutely. I think it will be not easy for everyone. I think needs a certain amount of technical familiarity. So this session is meant for a retail investor today like who can just use one very easy to use platform right which is plot co-work right now there are sophistication there are some appendix slides which talk about those there are rag patterns and all which you can build there is no end to sophistication to this but it will not be easy for everyone to adopt those pathway there are a lot of complexities in open source model there are a lot of risks your data gets passed on especially with the Chinese models like GLMs and D6 of the world. Maybe not with Nanetrons of the world, but I think it's not for everyone to keep it short.

127:05 >> Okay. Thank you. >> Yeah. >> Yeah. Yeah. Just for the example which you shown typically let's say for a quad cloud pro subscription would how much data would it or how much tokens would it eat? Would it be a one day token equivalent? Just a rough guess. just a r of idea for the kind of an analysis you showed on that detailed company at the end >> it it would again it is not fine to like I said it's a work in progress now and if you run with allmemes might consume maybe a week's budget at least >> and without the the the persona thing you know just the company sticking to the facts and figures rather than the the thing See what I'm trying to get to is a potato that you know what 20% can get me to the 80 90% of it.

128:02 >> Right. Right. So I would say I mean it will evolve. Right. There's no one answer to it. Right. But I would say roughly maybe a week would need 60 to 70% of the weekly budget is required and there are ways to optimize it like you start with a baseline like we discussed in the ethan slide right this is the let's say the worst AI you are using because it will eventually get better so opus 5.5 is significantly less token consuming than opus 5 for example right it is able to do the same thing much more streamlined so I'm saying these costs and everything will go down as we go into the future So don't derail yourself that I only have to use it.

128:40 Start using it, start fine-tuning it and maybe you will find that best balance, right? What works for it and it will be different for every individual because every individual looks at the stocks or looked at the investment research in a very different way. Okay. Okay. take I think we only 10 minutes over 3 hours. Rudrau I think let's sort of conclude here today. >> Okay, let me just one slide Visheka I'll conclude there. So I hope everyone was able to get some time to please share the feedback. The feedback is very important please do so. I'll just move to the I promise. So this is one mustwatch. If people want to like average cloud so this is a really very very comprehensive 2hour session on YouTube pause it work alongside it show it watch it multiple times is very important and the skills we discussed are sh there's a multiple GitHub reports so I've just shared couple of examples but be aware of do not connect these skills with any live tools for example your trading platforms and all your brokers never do that because there is a high risk to it. Use the skills.

129:56 You can analyze any skills. If you search on GitHub, there's a tons of people uploading their skills. You can download them, use it in your own cloud, combine with your own work. But do not connect with live systems as of now. Okay. And there's a cheat sheet some curated. This is like I said, you can print this out how the four part recipe and some curated prompt both worth keeping handy. >> Yep. So with that, we conclude.

130:20 >> We'll get this P PT now. We will upload this. Yes, >> sir. Sir, two questions. Sorry. >> Mhm. So, can we get the link for this last? >> Yeah. >> Yeah. So, I I'll share all of it. don't worry. you know, all of that all the >> and if if we want to connect with you offline then that >> Yeah. Yeah. My handles are also there. You can drop me a note, right? You can we can connect >> and so we'll get the PPTs now.

130:49 >> Yeah. Yeah. We we'll share that. >> Rudra, one quick question. >> Do we need to drop a mail for that? >> No, no, nothing. >> No, no. In I think we can we can upload in the same link now. Where this >> either there or I'll share >> I will share. We will share. Don't worry. >> Quick question. Are you aware of anyone using cloud or any AI for trade execution? >> There are I think more like trade executions are more sophisticated. they don't use like a plot this way. I think there are enterprise system for from high frequency trading to other areas.

131:26 but individuals I'm not aware of but companies are doing it definitely. >> So I'll give you a simple example for an individual investor where let's say I go on a track for two weeks and I do not have any contact with internet and I want to set up a number of GT orders go till trigger triggered orders but there would be loads in number. So is there a way that I can tell clot that these are my target price and these are my stop loss >> that that you can set up in your broker systems and I'm sorry to cut you off you know these are all very individualistic questions you can take that offline and anyway we are over over the time >> sir can you just share the link the training the deep dive training again >> yeah yeah we'll share all the links the the presentation all of it. I'll share it. Don't worry. Okay. We'll send out the email to all of you. I'll also upload it on on Twitter, Substack, whatever. Right. So, you should be able to get it.

132:31 >> Great. Thank you everyone. Thank you so much for the >> Thank you. Thanks, Rudra. Thanks for a wonderful presentation. >> Thank you all. Bye-bye. >> Thank you.