build-small-hackathon/metabolic-forensics

🩸 Metabolic Forensics

*An N-of-1 biosignal evidence engine — forensics, not coaching.*

Point it at your own wearable history (CGM glucose, HRV, recovery, sleep RHR, steps, morning alertness) and ask forensic questions:

• —"What reliably precedes my glucose spikes?"
• —"What's different about my bad-recovery mornings?"

The system does the hard part deterministically — temporal alignment, event-window search, personal baselines, and counterexample retrieval (does the pattern actually hold, or are there days it breaks?). A small local model only narrates the evidence it's handed. It cannot invent findings.

Every answer follows the same honest shape:

observed association → counterexamples → one next experiment to run

Never a diagnosis. This is associations in your data, not medical advice.

Why this isn't "just prompting Gemma"

The model is a commodity ≤32B model. The moat is the evidence pipeline: event detection, window alignment, baseline/variance estimation, and counterexample mining. The LLM is handed structured evidence and narrates it — it is structurally prevented from hallucinating a pattern that the data doesn't support.

Status

🚧 Skeleton running on synthetic demo data. The real engine + local model (llama.cpp, off-grid) plug in at the marked points in app.py.

Built for the Build Small Hackathon.