GlycoSwarm AI
A multi-agent clinical decision-support tool that screens diabetic patients' routine labs to flag early kidney, nerve, eye, and heart complication risk.
Overview
GlycoSwarm AI is a multi-agent clinical decision-support tool that screens diabetic patients' routine labs to flag early kidney, nerve, eye, and heart complication risk — months before symptoms appear — using specialist AI agents on real NHANES patient data.
A diabetic patient's labs can look stable — A1c not alarming — while subtle multi-marker patterns are already predicting kidney, nerve, eye, or heart damage forming underneath. By the time it's symptomatic, the damage is often irreversible. Today's clinical review is also siloed: one specialist, one chart, one metric at a time.
How it works
GlycoSwarm AI runs four independent clinical specialist agents — renal, neuropathy, retinal, and cardiovascular — over the same lab panel in parallel. Each specialist reasons out defensible early-warning cutoffs (deliberately more conservative than standard diagnostic thresholds) and writes and executes its own Python scoring code against the patient's real values, rather than calling a static lookup table. A synthesis agent then combines all four outputs, ranks which complication is most urgently developing, and recommends what to test or refer next.
Data and reliability
The pipeline runs on real de-identified CDC NHANES 2017–2018 patient data — not synthetic — and is honest under failure: if no LLM is reachable, the system reports "unavailable" rather than fabricating a clean result.
Technical stack
- LLM inference primarily on Gemma 4 26B via genuine on-GPU Ollama/ROCm inference on an AMD Instinct MI300X (AMD Developer Cloud)
- GLM 5.2 via Fireworks serverless as fallback
- Backend: FastAPI + LangGraph
- Frontend: Next.js
- Deployed live on Railway + Vercel
Background
Built for AMD Developer Hackathon: ACT II, Track 3 (Unicorn Track), by team Snowfall.