🏥 Charge-Nurse Triage Assistant

Forward-Deployed Engineering — I land clinical AI inside a hospital's real systems (VPC, behind the firewall — not vendor SaaS), wire it into their EHR workflow, operate it in prod, and drive adoption. The model is the easy part; the engagement is the job.

● live · Cloud Run </> source API
1The engagement — how this lands at a customer · the artifacts an FDE actually delivers, not just a model
deliverables → customer-brief solution-design deployment-plan runbook postmortem handoff
2See it work — pick a patient · canned scenarios that hit the 497-case corpus (no free text — that misses the corpus)
↳ the agent's recommendation · rule-based ESI + BM25 retrieval over similar cases — grounded, no black box
Pick a patient above.
3The accountable human decides · the moment most AI demos skip — and where adoption lives or dies
Agent recommends: —
A charge nurse signs off. The agent never acts alone.
4When it breaks — graceful degradation · on-call flips mode via /admin/mode; the real forward-deployed deliverable (see runbook)
ai_assist
retrieval + rules + grounded rationale
stale_data
FHIR feed lagging → "verify vitals manually"
rules_fallback
vector store down → rule-only ESI, still safe
off
assistant silent → pure human triage
Live on Cloud Run (~$0 idle, scale-to-zero). Rule-based ESI + BM25 retrieval + grounded template rationale — makes no LLM calls today (consumes $0 of the Vertex credit; honest by design). The FDE value is the runbook, the acceptance-test CI gate, the postmortems, and this human-decision layer — not the retrieval algorithm.