udarachamidu/medtrace-ai

Purpose and Operating Modes

MedTrace AI is a patient-record review application: it consolidates visits into timelines, medication and lab views, safety findings, and citation-backed record Q&A. It explicitly avoids diagnosis or treatment recommendations. The default reliable path is a synthetic, cached demo; a production mode connects the UI to Supabase and n8n AI workflows.

• frontend/src/App.tsx

• frontend/src/lib/data-adapter.ts

• backend/fixtures/synthetic-demo.json

Frontend Entry and Patient Workspace

The Vite/React client starts at frontend/src/main.tsx, routes between public/auth/patient views, and renders the main patient dashboard. data-adapter is the key boundary: it selects demo data versus configured remote data, insulating UI views from the source mode.

• frontend/src/main.tsx

• frontend/src/lib/router.tsx

• frontend/src/pages/PatientDashboardPage.tsx

Clinical Review Presentation

Dashboard components present the longitudinal timeline, medication/safety results, lab trends, and record-question interaction. The evidence drawer exposes the provenance required for a claim—source document, page, snippet, and confidence—rather than treating AI output as standalone fact.

• frontend/src/components/DashboardViews.tsx

• frontend/src/components/EvidenceDrawer.tsx

• frontend/src/components/SafetyNotice.tsx

Ingestion and AI Orchestration

n8n is the connected pipeline runtime. A patient-record workflow coordinates per-document processing, classification/extraction prompts, normalization, reconciliation, trends, safety checks, evidence verification, and persistence/query stages. Workflow templates are inactive until secrets, identity checks, and storage policies are configured.

• backend/n8n/workflows/WF-01-process-patient-record.json

• backend/n8n/workflows/WF-02-process-one-document.json

• backend/n8n/workflows/WF-09-ask-record-question.json

Clinical Processing Stages

Extracted records pass through medication and laboratory normalization, then longitudinal medication reconciliation and lab trend generation. Deterministic TypeScript safety rules evaluate allergies, duplicate therapies, dosage conflicts, and interactions; this provides a non-LLM safety boundary alongside prompt-driven extraction and verification.

• backend/n8n/workflows/WF-03-normalize-medication.json

• backend/n8n/workflows/WF-07-safety-rules.json

• backend/src/rules/engine.ts

Evidence, Confidence, and Prompt Contracts

Versioned prompts define LLM roles for document classification, clinical extraction, reconciliation, trends, evidence verification, and Q&A. Backend domain utilities model citations, confidence, dates, identity, ownership, and normalization so claims can remain attributable and uncertainty-aware.

• backend/prompts/clinical-extractor/v1.md

• backend/prompts/evidence-verifier/v1.md

• backend/src/citations.ts

Persistence, Security, and External Services

Supabase is the system of record and private document store, with migrations establishing schema/extensions, RLS, storage/RPC, vector-search support, and cross-patient integrity protections. The frontend uses only public Supabase/n8n configuration; n8n holds server-side Supabase credentials, webhook secrets, and OpenAI access. OpenAI is the intended LLM provider.

• backend/supabase/migrations/202607300001_extensions_and_schema.sql

• backend/supabase/migrations/202607300002_rls_storage_and_rpc.sql

• backend/supabase/migrations/202607300003_cross_patient_integrity.sql