Show HN: Fighting Medical LLM Hallucinations with a Grounded RAG System

Hi HN,

We've been frustrated with how confidently LLMs hallucinate—a dangerous flaw in high-stakes domains like health and medicine. The standard "I am not an expert" disclaimer feels insufficient since we all ignore those statements.

Our approach is a RAG/agentic system built to solve this. It runs on ~40M+ scientific papers, but goes beyond simple retrieval. A multi-agent workflow decomposes queries, cross-references claims against multiple sources, and synthesizes answers, ensuring every key statement is cited directly from the literature. Beyond the literature, our agent system has tools to access the internet, databases, and social platforms, with dedicated review agents to ensure proper citation and reduce hallucinations.

This is just the start. Our long-term goal is building health superintelligence by integrating multiscale data—from the genomic and cellular level all the way up to clinical studies in humans. To achieve this, we're exploring SFT, RL, and self-improvement techniques like GEPA to create models that can evolve their own scientific reasoning and to pioneer new standards for accuracy/hallucination mitigation. We plan to rigorously benchmark our work and share the data publicly.

We'd love specific feedback on:

Our RAG/agentic architecture—what failure modes are we missing?

On building superintelligence—beyond SFT/RL/GEPA, what other techniques should we be exploring for a model to truly understand multiscale biology/health/medicine?

Evaluation—what are the best benchmarks for medical/health AI trustworthiness today?

The site itself—any thoughts on the UI/UX, quality of the responses, or other features?

You can see the current system here: https://www.my-openhealth.com/