Care Coordinator

We endeavors to empower and educate people with evidence-based information to help them educate themselves about potential health conditions and better navigate the healthcare system.

Comment

Inspiration

The bottleneck in healthcare isn't diagnosis — it's that patient context never travels. Patients repeat their story to a scheduler, a nurse, a doctor, and it still doesn't arrive intact. Meanwhile ER-vs-urgent-care-vs-PCP decisions get made blind to both symptoms and benefits, ~80% of clinical trials miss enrollment timelines, and recruiting a single trial patient costs ~$6.5K. We built the layer that makes patient context portable, auditable, and machine-readable.

What it does

CareCoordinator is a multi-agent care navigator. A patient describes a health concern in plain language and four agents coordinate:

  • Safety screen — deterministic red-flag check on every message, before any model call. "Crushing chest pain" escalates straight to 911/ER with no LLM in the loop.
  • NP intake agent — asks 3–4 high-yield clarifying questions (skipping anything already in the patient's stored medical history), then builds a hypothesis-ranked differential grounded in live PubMed E-utilities calls — real, clickable PMIDs, retrieved at runtime.
  • Clinical trial navigator — queries the live ClinicalTrials.gov v2 API for recruiting studies and screens the patient's answers against real eligibility criteria. We surface research options; sponsors pay for qualified referrals, and matching is independent of compensation.
  • Insurance coordinator — reads the member's benefits (synthetic "Open Access Plus w/ HSA" plan) and produces deductible-aware out-of-pocket estimates and site-of-care guidance.

The care coordinator then produces two outputs: a compassionate patient plan, and our core artifact — care-context-handoff/v0.1, a machine-readable JSON handoff a clinician's agent can audit: generator model, assumptions, uncertainty notes, hallucination-risk disclosures, verifiable evidence, and the safety-screen result. Sessions stay open for grounded follow-up questions, and the whole encounter exports as a PDF for the patient's doctor.

How we built it

FastAPI orchestrator, four focused Claude (claude-sonnet-5) calls with strict JSON contracts, live tool use against NCBI E-utilities and ClinicalTrials.gov v2, and a dependency-free three-panel UI showing the patient conversation, the live agent activity feed, and the handoff JSON as it's generated.

Challenges we ran into

Thinking-model output handling (thinking blocks eating token budgets and breaking JSON contracts), and a real retrieval failure: the NP emitted "Alzheimer's disease (early symptomatic / mild cognitive impairment due to AD, biomarker-pending)" and ClinicalTrials.gov returned zero results. Our own handoff artifact's methodology audit is how we diagnosed it — we shipped a query-simplification fallback within minutes. The audit layer proved its value on day one, on our own system.

What's next

First sponsor pilot in one therapeutic area, real benefits data, HIPAA-eligible infrastructure, and opening the handoff protocol so EHR and provider agents can consume — and write back to — portable patient context.

Built With

Share this project:

Updates