Agentic AI · 2025
Chartwell Summary
Healthcare SaaS · internal product · Architect
A clinical summarisation engine that reads a patient's scattered record — problems, allergies, notes, external documents — and hands the physician a structured, cited summary before the appointment starts.
reduction in record-review time
notes summarised per month
cost per summary at steady state
The problem
Physicians were opening eight tabs to prepare for a fifteen-minute consultation. The record was complete but unreadable: duplicate problems from three source systems, PDFs with no structure, and no way to see what had changed since the last visit.
What I did
- Designed a deterministic extraction layer over HL7 FHIR resources before any model call, so the summary is grounded in codes rather than prose.
- Built a structured-output pipeline that returns cited fields — every line traces to a source resource the physician can open.
- Created an evaluation set from 1,200 de-identified charts scored by clinicians, wired into CI as a regression gate.
- Set per-tenant token budgets with model routing, keeping cost per summary under a cent at 1.4M notes a month.
Architecture
Java and Spring Boot services on EKS, orchestration over Kafka with idempotent consumers, a FHIR-shaped document store, retrieval from OpenSearch, and model calls through an internal gateway that enforces PII redaction, audit logging and spend caps. Datadog dashboards per tenant.
Request path
Similar problem on your side?