Document Data Extraction

A multi-tenant SaaS that turns uploaded invoices and receipts into structured, exportable data. Organizations sign up, invite teammates with roles, upload documents, and a background worker runs OCR-based extraction (date, vendor, total, line items). Results are reviewable in the UI and exportable as JSON or CSV. Every record is isolated by organization — a tenant can never read another tenant's data.

Built as a TypeScript Nx monorepo: a NestJS API, a standalone NestJS worker, and a Next.js (App Router) frontend, backed by PostgreSQL, Redis/BullMQ, and S3-compatible storage (MinIO locally).

Architecture

flowchart LR
  user([User]) --> web[web: Next.js]
  web -->|REST + JWT| api[api: NestJS]

  api -->|tenant-scoped reads/writes| db[(PostgreSQL)]
  api -->|store upload| s3[(MinIO / S3)]
  api -->|enqueue extraction job| redis[(Redis / BullMQ)]

  redis -->|consume job| worker[worker: NestJS context]
  worker -->|fetch object| s3
  worker -->|OCR| tess[Tesseract]
  worker -->|persist result + status| db

Loading

The API never does heavy work inline: it stores the upload, enqueues a job, and returns. The worker consumes the queue, fetches the object, runs the swappable Extractor (Tesseract today, an LLM implementation later), and writes the result back with a status transition (pending → processing → extracted | failed). The frontend polls document status to reflect progress.

apps/
  api      NestJS HTTP API — auth, organizations/invites/roles, documents, queue producer
  worker   Standalone Nest context — BullMQ consumer running the extraction pipeline
  web      Next.js App Router frontend
libs/
  shared-types   DTO / enum / queue contracts shared by api, worker, web
  domain         Tenant-scoped business rules (placeholder for growth)
  config         zod-validated environment loading
  database       Prisma client + module
  storage        S3-API storage service + module
  extraction     Extractor interface, Tesseract implementation, field parser

Quickstart

Prerequisites: Node 20+, npm, Docker.

npm install                       # also runs `prisma generate`
cp .env.example .env

docker compose up -d              # Postgres + Redis + MinIO (+ bucket bootstrap)
npm run db:migrate                # apply migrations
npm run db:seed                   # demo org, users, and documents

npm run serve:api                 # http://localhost:3000/api
npm run serve:worker              # consumes extraction jobs
npm run dev:web                   # http://localhost:4200

Health check: curl http://localhost:3000/api/health → {"status":"ok","db":"up"}.

Demo credentials

After npm run db:seed, sign in at the web app with:

• owner@demo.test / password123 (owner of Acme Receipts, pre-populated documents)
• member@demo.test / password123 (member of the same org)

Common commands

Tasks run through Nx via npm scripts (local Nx, no npx).

Task	Command
Serve api / worker	npm run serve:api · npm run serve:worker
Run web (dev)	npm run dev:web
Lint / typecheck / test (all)	npm run lint · npm run typecheck · npm test
Verify before commit	npm run verify
Build (all)	npm run build
Any other Nx target	npm run nx -- <target> <project> (e.g. npm run nx -- affected -t test)
One project's task	npm run nx -- test web

Conventions

Engineering conventions — code style, layering, central tenant isolation, the extractor interface, and the Prisma-vs-TypeORM rationale — live in CLAUDE.md. The /self-review, /architecture-check, and /verify skills under .claude/skills/ enforce them.

Future work

Out of scope for the MVP, intentionally deferred:

• Billing and subscription plans
• Real-time collaboration and notifications
• The LLM Extractor implementation (the interface and DI seam are ready)
• PDF rasterization for OCR (images are supported today)
• Real email delivery for invites (tokens are surfaced in the UI for now)
• Multi-language OCR and many additional document formats
• Webhooks and an audit log
• Training a custom extraction model