Meridian

Inspiration

Healthcare quality improvement teams spend hours manually tracing whether positive mental health screenings receive timely follow-up assessments. The rules are buried in policy documents, the data lives in disconnected systems, and the audit trail is a spreadsheet. We wanted to build a system where the reconciliation logic is transparent, every decision traces back to the source evidence that defines it, and the knowledge corpus is explorable — not just searchable.

What it does

Meridian is a multi-tenant reconciliation platform that answers one critical question: Did each positive PHQ-9 screen get a follow-up C-SSRS assessment within two business days?

It ingests screening events and assessment events, runs a deterministic reconciliation engine that classifies each screen by eligibility and linkage outcome, and presents results through both a data table and a visual pairing view with timeline visualization. A "Company Brain" module connects the reconciliation decisions back to source documents — publications, procedures, and policy memos — so analysts can see why a decision was made, not just what it was.

How we built it

Backend: NestJS + Prisma + PostgreSQL with a pure-function reconciliation engine that computes eligibility, business-day windows (accounting for weekends and holidays), and linkage state for every screen.
Frontend: Next.js App Router + Tailwind + shadcn/Radix components, with a visual pairing view (screen↔assessment card pairs with color-coded connectors), donut chart summary dashboard, and business-day timeline.
Knowledge layer: HydraDB-backed document store with cross-collection unified search, provenance linking (decisions traced to source passages), graph context rendering showing how publications inform protocols, and a document upload UI.
Infrastructure: Turborepo monorepo, pnpm workspaces, Docker (Postgres + Redis + Mailpit), with shared types across API and web to catch drift at compile time.

Challenges we ran into

Field name drift between the reconciliation engine and the frontend DTOs — the engine uses matureEligible and absentLink while the UI expected mature and absent. This produced a page full of zeros that looked like missing data but was actually a mapping bug.
Multi-tenant data isolation made debugging harder: seed data lived in one organization while the logged-in user was navigating a different one, so everything appeared empty despite the database being full.
Business-day calculation across weekends and holidays (e.g., a Friday screen before Labor Day) required careful timezone-aware logic to get the two-day window right.

Accomplishments that we're proud of

The reconciliation engine is a pure deterministic function — same inputs, same outputs, every time. No AI in the decision loop.
Every decision can be traced back to source evidence via provenance linking, closing the loop between "what happened" and "why the rule says so."
The visual pairing view makes screen-to-assessment matches visible at a glance — blue connectors for within-window, amber for late, dashed grey for missing, with expandable business-day timelines.
Seven features shipped end-to-end (migration → service → DTO → API client → UI) in a single session.

What we learned

Building across the full stack in one pass (as the codebase conventions demand) catches integration bugs immediately — splitting backend and frontend into separate tasks is where field names drift apart.
Multi-tenant systems need seed data that matches the actual user's org membership, or demo day is a blank screen.
A "Company Brain" is fundamentally different from a search box — it's about showing connections between documents, not just listing results.

What's next for Meridian

Live reconciliation streaming — run the engine incrementally as new screens arrive instead of batch-triggering.
AI-assisted review — use Claude to summarize ambiguous decisions and suggest resolution paths, while keeping the deterministic engine as the source of truth.
Multi-protocol support — extend beyond PHQ-9/C-SSRS to arbitrary screen-assessment protocol pairs defined by each organization.
FHIR integration — ingest screening and assessment events directly from EHR systems via FHIR R4 resources.
Audit workflow — built-in review queues with role-based assignments, sign-off chains, and exportable audit packages for regulatory submissions.