SuppDive AI: scan & evaluate supplements

Inspiration

A few months ago my mom noticed my hair thinning and like most people she recommended a supplement a friend swore by. When I asked, “how do you know it works?” the answer was basically: word of mouth. That moment made me realize how broken the supplement discovery process is. We lean on friends, family, influencers, and star ratings. Rarely on actual evidence. The space is noisy, full of hype and questionable reviews, and it’s hard to tell what’s safe, effective, or even necessary. I wanted to flip the script and make decisions start with science instead of anecdotes.

What it does

• SuppDive AI is an evidence-first supplement scanner and evaluator. Capture a photo or scan a barcode and it instantly parses the label (ingredients, dosages, brand) and generates an evidence-backed report. Behind the scenes, it: Scans any label instantly to extract ingredients, dosages, and product details. We have a special AI for this task.
• Checks each ingredient against peer-reviewed studies and clinical trials to surface efficacy, common side effects, and potential interactions.
• Evaluates the product for safety/quality signals (manufacturer reputation, FDA compliance cues, third-party testing, and more).
• Analyzes public reviews at scale, filtering likely fakes and comparing claims to clinical outcomes.
• Delivers a comprehensive report, clear verdicts, citations, and a plain-English summaries and insights so people can decide based on evidence, not hype. All this done by AI while ensuring transparency by showing the user the full scope of the investigation activity, making it easy to access and double check informations from the user end.

How we built it

We built SuppDive AI by decoupling it first, each component is at is own, for example the label scanning feature includes it's own integrations, LLMs, tools, and convenience. The evaluation feature includes many interconnected components to make the process smooth, factual, and accurate. For example, at the evaluation part, we have 4 types of AI "Agents". One that is the orchestrator, it's the one communicating and coordinating with other AI "Agents". The second Agent is working closely with scientific literature, it has custom tips and workflows to find relevant scientific studies/articles in the context of the current supplement that is being evaluated. We also built a special Agent that process content and detect "fake" reviews and claims by compiling information agains the scientific literature insights Agent 2 found, it uses a critical methodology that relies only on evidence, it can extract "highly relevant tokens" that will be used as evidence, help us cite and inform the report. That will put us in the last Agent, the report maker. We rely on the "Next Word Prediction" so much here, as this Agent will receive everything we did in the process from the state 0 including reasoning tokens, so it can generate the final report, a report that is designed to be accessible, factual and easy to grasp.

Challenges we ran into

• AI reliability: since the system uses AI at it's core, insuring a low rate of hallucinations was a must, we iterated so much on the system prompt and rules to make the process as factual as possible, including testing agains all the cutting-edge LLMs available today.
• Tools ecosystem: we knew that we need a sophisticated sourcing, so picking and integrating the right tools was also a challenging part, it took us so many time and dedication to double down on the winners and ditch the others.
• Realtime capabilities, one thing that we focused on that wasn't provided by the market, is to put the user in the loop of the evaluation process. We made sure that each information that is getting in the context window of collectively, SuppDive AI, is also being observed by the user, think of it as a human observer in the loop, we made everything easily accessible by clicks, even if the user is not in front of the app, Live Activities will make sure to deliver progress and facts to the user throughout the process.
• We needed resilience, because adding other features was tempting, the system was ready to scale, at the point of investigating a full Supplement Stack, doing a highly personalized research based on the user's Health Profile, managing supplements routines, managing supplements for multi profiles by one user, and many other features that we are really exciting about bringing it own. But we choose to gate most of it for now, letting our attention be directed at the core level, which is ensuring a high quality investigation process and an accurate factual report that will at least be near to how a human professional will do it.

Accomplishments that we're proud of

We are really proud of the problem we are solving. We did our research, we found out that: By the mid-2020s, taking dietary supplements has become a normalized behavior for the majority of U.S. adults, with many treating it as a staple part of their personal health maintenance.

• 77% of Americans view the supplement industry as trustworthy and 87% express confidence in the products they use. (NutraIngredients-USA.com)
• 64 % of supplement reviews are fake (Capital One Shopping 2025)
• Online Research is the Norm: The majority of supplement users turn to the Internet as their primary source of information before using or buying a product. One study concluded, “most [dietary supplement] users sought information prior to using them, [and] the majority did so by using the Internet,” whereas only a few consulted with a healthcare (providerjournals.sagepub.com).
• 69% of supplement users in 2024 valued a personalized regimen tailored to their (needsdrugstorenews.com). So this data points was clear about what we will offer, we are not interested in replacing healthcare professionals, and we really instruct users to do so, but we are providing a unified way to research supplements online, one platform, all sources, no more 10 tabs open, no more conflicting informations, no more hype-claims, only evidence.

What we learned

We learned that building AI-native mobile application is not an easy thing as all those yapping about "AI Wrappers" say, it needs a level of care especially if you want to make something useful, we are largely agree with the saying of "More than 80% of AI projects fail, marking twice the general failure rate of IT projects. This striking statistic is quoted by many expert organizations, including a global non-profit think tank, The RAND Corporation." And to conclude, one insight we found, is that wether an AI-native software will work or not depends largely on the reliability of the tools we give them, it's not a plateau of AI capabilities, it's a plateau of the ecosystem we put them in.

What's next for SuppDive AI: scan & evaluate supplements

We are waiting for users to come, with so many amazing features coming up as we stated above, we want to be the #1 Supplement App in the world, accompanying users throughout their supplementation process. We spent 1y and half building this project, and we are happy to continue doing so if we have users!

Built With

• antrhopic
• dsld
• figma
• firebase
• gemini
• golang
• openai
• postgresql
• pubmed
• swiftui