.png&w=256&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
GenRisk
GenRisk is a multi-omic cardiometabolic risk reader. It takes a patient's genetic risk report and layers additional evidence on top — DNA methylation, gut microbiome, and structured clinical history — reading all four layers side by side instead of fusing them into a single, misleading score. The problem: three lab reports from three sources that no one reads together. The common move is to collapse them into one risk number, hiding which layer drives the risk and burying the evidence. What's different: every number on screen traces to a published study — hazard ratio, confidence interval, source, and outcome. Nothing is invented by the model. Layers are stratified by outcome, never merged; the base genetic risk is read alongside, never recalculated. When a factor is protective (e.g. light alcohol intake), the panel shows it honestly as a protective force, displaying "divergent forces" when risk and protection coexist. Architecture: a deterministic engine does the math (no LLM touches the numbers), grouping modifiers by outcome; a language model writes the cross-reading under strict grounding, citing sources or declaring "insufficient evidence." Retrieval uses local FAISS embeddings over a curated peer-reviewed corpus. AMD: the embedding pipeline auto-detects AMD GPU via ROCm (CPU fallback) and ships with a benchmark. On AMD Developer Cloud it scales to a larger medical-literature corpus. GenRisk is a proof of concept for an architecture — honest, provenance-first multi-omic integration — not a diagnostic product.
13 Jul 2026
