An AI-platform signal layer (L1.5) that turns trusted healthcare data into pre-computed signals — anomaly, cluster, classify, forecast, rank — that feed a GenAI agent as labeled fields, never as raw text in the context window. The Signal Console shows the full chain on three audience cases (ER triage, ops capacity, exec brief): L1 truth (dbt/warehouse) → L1.5 signals → L2 Gemini agent decision → L3 human override. The architecture insight — signals are computed before the agent, so it reasons on labels not noise — is the senior platform-engineering answer to context pollution.
Signals are machine intuition the agent consumes, not computes — the platform pattern Palantir AIP uses, shown end-to-end on real eval numbers.
Tech Stack
Features
Smoke Detector — anomaly
Z-score over (gender × condition × age-band) cohorts on LOS + admission alignment. Eval: P 0.79 / R 0.92 / F1 0.85, FPR 6% on a 250-case synthetic injection set.
Treasure Map — cluster
Pure-Python k-means on 5 patient-level features (z-standardized). Surfaces 535 complex high-utilizers (1.3%) out of 40,167 patients; silhouette 0.41.
Traffic Light — classify
ESI tier + NOW/SOON/WAIT with a safety overlay. ±1-tier accuracy 100%, bucket accuracy 94%, 0 safety-critical violations across the gold set.
Signals feed the agent as labels
Each signal is a labeled field ({anomaly_score, cluster, esi_tier}), pre-computed in L1.5. The L2 Gemini agent reasons on the labels — never recomputes them — avoiding context pollution.