Full-stack AI · Clinical systems demo
An AI care companion that extends GLP-1 obesity care between appointments.
CareLoop pairs a patient-facing coach with clinician tools. Answers are grounded in reviewed guidance, personalized by longitudinal memory, gated by safety triage, and fully traced and evaluated.
Request pipeline
- 1Safety triage
Deterministic rules run before any model call. Emergencies bypass the LLM entirely.
- 2Retrieval
BM25 over a clinician-reviewed knowledge base, with patient-language synonym expansion.
- 3Memory read
Longitudinal facts (doses, symptoms, goals, barriers) personalize every answer.
- 4Generation
Claude streams a grounded answer that cites guidance ids. Abstains if nothing matches.
- 5Validation
Every citation is checked against what was actually retrieved.
- 6Memory write
Structured-output extraction adds new durable facts; newer facts supersede old ones.
AI patient coaching
Streaming chat grounded only in approved guidance, with inline citations and personalization from patient memory.
Patient memory
Facts extracted from conversation into a typed, longitudinal record that clinicians can read and the coach can use.
Clinician workflow
Escalation queue with acknowledgement, a memory timeline, and an LLM-generated pre-visit brief.
Observability
Per-request traces with a span waterfall, token usage, safety level, sources, and citation-validation errors.
Evals as release gates
Emergency recall, triage accuracy, retrieval hit@3, abstention, and memory extraction, run in CI.
Runs anywhere
Live mode uses Claude via the Anthropic SDK. Without a key, a deterministic responder keeps the hosted demo free.