Final project EC

0:01 More than one in eight American adults are currently on GLP1 and among women it's closer to one in seven. This are medications like ompic and wgoi prescribed for type 2 diabetes and weight loss. The drug works by suppressing appetite. So people eat less and lose weight fast. The problem is that the side effects that come with rapid weight loss. A large share of the weight loss can come from muscle instead of fat and the same appetite makes it hard to eat enough protein to hold on to the muscle. Many people also get nausea that makes eating and exercising much harder. Most of this is manageable actually with the right support, diet, exercise, behavioral shift. There's currently no system design for people who are on GLP1, however, and the supplement aisle only sells generic formula, but for no one in particular, and everyone's side effects and needs are actually very different. This is the gap that we're solving for.

1:01 So, moving on to the first portion of our demo, we made a really deliberate choice with the order here. The first thing that you reach is not a product is a free personalized plan. You answer 10 questions and get evidence-based guidance before we ever ask you to buy anything. So the value comes first and the product comes second. And only um product will be served to you if there is a fit or if there's a match based on the survey. So basically let's take a look at the survey. I'll pretend that I'm one of the GP1 user. Um so how long have you been on GLP1? I'll say under 3 months. Um clinical, what are your dosage? What what what are you in your dose journey? I'll say just started. What's your biggest challenge right now? Um low energy. What's your main goal on GLP1? Um manage side effects. How much protein do you eat per day? 70 to 100. What kind of movement do you do? Walking light cardio. How often do you exercise? I'll say one to two per week. What's your energy throughout the day? Crashes midday. And any dietary preference? I'll say none. And what's your age range? 25 to 34.

2:17 So when I'm finished, the system does three things in order. It retrieves the clinical claims most relevant to this specific profile. It writes a plan grounded only in those claims. Then a second pass scores how well grounded and how specific that plan is. Um here's the output. So it starts as a protocol result. Um basically it opens with what's likely happening in your body. The three free changes. Um so here the three changes are train away from your dose day window, hydrate with electrolytes and cover the everyday basics consistently.

2:53 Is I validated this two ways beyond the building scoring. I ran the quiz across a range of different profiles to see where the engine is confident and where it is weakened. I pressure test the thesis through interviews with clinicians and operators in GLP1 and consumer health space. Um, and only when the result is generated, we introduce our products framed as closing a specific gap um that's really personalized to the user. Um, so yeah, so here we have um the products that I've designed that's really catered to the GLP1 users. So I designed the packaging myself. The goal is ultimately develop more nuance products for different specific needs. But right now, Nomad is two pouches ground for everyday usage lived for work for workout days or for more active users. It could also be bundled together as a protocol about $75 a month. That's priced against roughly 370 that it would cost to buy every single ingredient um separately. So here is like a comparison of what the ingredients are and how much it would cost um separately and every claim traces back to the literature clinical evidence and conversations with clinician. Um at the bottom there's a front end people interact with a server that runs the engine and a database that stores every response. The interesting part is the engine and I make a few deliberate choices there. The first one is that I actually don't ask the model to remember clinical facts. That's exactly where this model make things up.

4:28 Instead, the system pulls a small set of vetted claims from a knowledge base and hands those to the model as the only evidence is allowed to use. Right now, it matches the claims to the person's profile by tax. At scale, that becomes proper vector search. The upside is that every claim stays traceable to a source which really matters when you're talking about someone's health and recommendation there. From there, the model gets those claims plus the person's profile and writes the plan based on those. The archetype, what's likely going on, three behavioral changes and a product suggestion. And I've told it two things firmly. One is don't introduce any numbers that is not um the actual evidence and two always lead with a free changes before the product. So the whole trust first idea is built into the system.

5:22 Also a second model pass that acts as a check. It scores how wellrounded and how specific the plan is and flags anything that's not supported. that's doing double duty is guardrail against halluc hall hallucination and also is how I evaluate the system. Um I also build a fallback. The same pipeline has a simpler rulebased version that runs no setup at all. So if the model ever fails, you'll still get a result and a little label that tells you which one produced it. That's also what lets the demo run anywhere and with no infrastructure behind it. Lastly, I wanted to talk a little bit more about why I chose this topic. I care about this because it really sits where my interests meet. Women's health and a market that keeps underserving women.

6:07 GLP1s are also one of the biggest shifts in medicine right now. Most of the people on them are women and they're being handed a really powerful drug with almost no support around it. That gap is what pull me in. I want to provide more personalized support and high quality product for GLP1 users. Um and lastly, I want to talk about um what's next and what um improvements can be done and a few directions that I'm excited to take this project to. First of all, the clinical layer. Right now, the engine pulls from a really small knowledge base. The real version replaces that with a fully cited research corpus and proper vector search. So, every recommendation is traceable to the verified source.

6:51 Second, richer intake. I would add a photo upload that reads the label on whatever supplement the patient's already taking. Maybe a blood panel and the associated results with it. Second reacher intake, I would add a photo upload that reads the label on whatever supplements the patient's already taking. Or it could be a health record or whatever the primary doctor prescribes to the patient that patients already taking today. So the plan accounts for what the patients already has and flex for what's redundant or missing instead of asking patient to type it out or to ask you know a million questions to to arrive there. Third, I want to make it feel like a real brand.

7:30 Use AI energy AI generated product video and the kind of user style content that makes a launch seem more credible. All designed around the archetype the the quiz already produces. And longer term, I think the most interesting piece here is the quiz isn't just a recommendation tool. It also is a data engine. So we'll collect a lot of the data based on the responses and it's a real signal on what the population actually needs. With enough of them, I can see which segments are underserved, which problems show up together and where the real product opportunities are for the future.