DataBridge AI
A production-oriented clinical data automation platform designed around one safety principle: AI can propose, deterministic code executes approved rules, humans review the risky decisions, and every committed value is auditable, reconciled, and reversible.
The Problem
Clinical teams often re-enter source data from files, lab portals, eSource pages, or spreadsheets into EDC forms. Simple automation is risky in this domain because wrong-subject entry, stale source data, silent field mismatches, and missing audit trails can compromise clinical data integrity.
- Manual copy-paste data entry is repetitive, slow, and hard to audit at scale.
- AI cannot safely write directly into clinical records without deterministic checks and human approval.
- EDC forms vary by vendor and study, so browser automation needs layered diagnostics and mismatch blocking.
- Every value needs provenance: source, transform, reviewer decision, committed value, and read-back result.
Solution Overview
Core Platform
- TypeScript pnpm monorepo with shared Zod schemas, clinical record states, RBAC roles, and domain types
- Deterministic transform package for units, dates, code lists, booleans, idempotency keys, and normalization
- Mapping registry with proposed, approved, deprecated, versioned, and effective-dated mappings
Audit & Reconciliation
- Hash-chained audit store with verification utilities
- Read-back reconciliation compares intended writes against committed EDC values
- Mismatch, missing-value, and stale-data cases route to quarantine and query generation
Review API & UI
- Fastify API with JWT authentication, RBAC middleware, review queue routes, audit routes, metrics, and PostgreSQL integration
- React review UI for login, queue triage, review detail, and audit-log workflows
- Confidence scoring combines validation, plausibility, historical accuracy, model confidence, risk class, and lock state
Chrome Extension
- Manifest V3 extension for Collect, Upload, Map, and Fill workflows
- Detects subject and visit context, scans fillable fields, previews mapped values, and blocks subject mismatches
- Supports Veeva-like, Medidata-like, Oracle-like, ClinSpark-like, localhost mock, and unknown EDC layouts
Architecture
DataBridge separates source ingestion, schema validation, mapping, deterministic transforms, human approval, EDC write adapters, read-back reconciliation, audit logging, and browser-assisted workflows. That separation keeps AI-adjacent automation useful without making it an uncontrolled clinical decision-maker.
- Core: TypeScript, pnpm workspaces, Zod, Vitest, RBAC, clinical state machine
- Backend: Fastify, PostgreSQL, Redis, JWT RS256, bcrypt, pino logging, Prometheus-style metrics
- Frontend: React 18, Vite, Tailwind CSS, review queue, review detail, audit log
- Extension: Chrome MV3 APIs, EDC field scanning, subject/visit detection, form filling, diagnostics
- Quality: 600+ automated tests and validation docs for IQ, OQ, PQ, and migration rollback
Screenshots
Challenges & Solutions
- Clinical safety: AI is limited to proposals and assistance; approved deterministic rules drive execution, and risky fields require human review.
- Auditability: Hash-chained events and reconciliation evidence explain where each value came from, how it changed, who approved it, and whether read-back matched.
- EDC variability: Layered subject, visit, and field extraction handles URL parameters, headings, breadcrumbs, labels, hidden inputs, data attributes, and mock/vendor-like layouts.
- Actionable failure states: Fill diagnostics show detected subject, visit, visible inputs, candidate fields, source selection, unmapped fields, and why filling is disabled.
Outcomes & Impact
- Demonstrates a safety-first pattern for applying AI-adjacent automation to regulated clinical workflows.
- Reduces manual EDC entry while preserving human review, provenance, audit evidence, and read-back reconciliation.
- Creates reusable contracts for future source connectors and EDC adapters.
- Combines systems design, frontend, backend, browser automation, data engineering, and applied AI safety in one product.