1
I've been concerned about the lack of safety and efficacy data for most supplements on the market. While the supplement industry is booming, it's largely unregulated, and it's difficult to get reliable, evidence-based information. General-purpose LLMs often give generic advice and don't deeply engage with the medical literature.
That's why I built OpenHealth: https://www.my-openhealth.com/
It's an AI-powered platform designed to provide high-quality, personalized health information, with a strong focus on supplement safety. I recently wrote a blog post about the specific safety concerns that motivated me to build this project: https://www.my-openhealth.com/blog/supplement-safety-is-why-...
I wanted to share some of the technical details with the HN community, as I'd love your feedback on the approach.
The Tech Stack
At its core, OpenHealth uses a Retrieval-Augmented Generation (RAG) system that draws from over 38 million medical abstracts from PubMed and other scientific journals. Here's a breakdown of the key components:
• RAG and Paper Quality Ranking: We don't just retrieve information; we have a system to rank the quality of the papers. This helps us prioritize information from higher-quality studies and avoid relying on weaker evidence.
• Neural Search Across All Literature: Beyond our embedding database, we can access all the literature including preprints via neural search, which works pretty well. This gives us comprehensive coverage of the latest research.
• Fine-Tuned Models & Optimized Prompts: We use fine-tuned language models that are specialized for the medical domain. The prompts are crafted and optimized by a team of clinicians and scientists to ensure that the generated responses are medically sound and relevant.
• Context Engineering for Medical Data: We've put a lot of effort into context engineering to accurately extract and analyze medical data. This is crucial for understanding the nuances of medical literature.
• Drug and Supplement Interaction Database: We've engineered a comprehensive database for predicting interactions between drugs, supplements, and even lifestyle factors. This is a key feature for ensuring safety.
Between the prompt engineering and the grounding in extensive API semantic search, we get really good health responses. We've tested with clinics who prefer it to ChatGPT for medical queries.
The Goal
The ultimate vision is to build a "health superintelligence" – a system that can provide state-of-the-art guidance on medical, wellness, and longevity protocols, all grounded in scientific evidence. We're accomplishing this through SFT/RL on large biomedical RL envs. More to report on this in the future!
I'm launching this project and would be grateful for this community's feedback on the technical approach, the quality of the information provided, the UI/UX, and any other thoughts you might have.
Thanks for your support!
I've been frustrated with the quality of medical and health responses from general-purpose LLMs like ChatGPT and Gemini. They often provide generic, non-committal advice and don't seem to effectively search or cite the actual medical literature with fast responses.
So, I decided to build my own solution: OpenHealth.
https://www.my-openhealth.com/
It's an AI-powered platform designed specifically for high-quality health information. Under the hood, it uses a combination of Retrieval-Augmented Generation (RAG) over 38 million medical abstracts (from PubMed and other journals), paper quality ranking, fine-tuned models, and specialized prompting to generate responses that are grounded in scientific evidence. Importantly the prompts, analyses, and reports are crafted by a team of doctors and scientists to provide high quality health information that is vetted.
The goal is to provide users with direct access to insights from the literature, calculate personal health risks based on studies, and even analyze supplement ingredients/stacks etc.
The ultimate vision is to build "health superintelligence" — models that are state-of-the-art on medical, wellness, and longevity benchmarks.
I'm launching it today and would love to get this community's feedback on the approach, the quality of the responses, the UI/UX, and any other thoughts you have.
Thanks for checking it out!