FreightAudit
Audits carrier freight invoices: an LLM reads each invoice, deterministic code checks every charge against the contract, and a person decides.
Live demo: https://freightaudit-bay.vercel.app (the sample data opens without signing in)
- Reads invoice PDFs, digital or scanned, one at a time or in batches of up to 500: header fields and every charge.
- Checks each charge in code against the carrier's rate card on the ship date and the shipment record from a TMS, and flags overcharges, charges outside the contract, charges billed twice and repeated invoice numbers.
- Holds every finding for a reviewer, who approves or disputes it in the web app; the audit trail records who decided what, and when.
The carriers are fictional and every invoice is synthetic.
Results
On 98 synthetic invoices with known ground truth (golden set v2); every metric with its 95% interval is in the eval card.
- Overcharges caught: 97.4% with Claude first (report), 94.9% with the free models only (report).
- Clean charges flagged: 0% with Claude first (report), 0% with the free models only (report).
- Cost per scanned invoice, paid-equivalent: $0.0105 with Claude first (report), $0.0020 with the free models only (report).
Built with TypeScript, Fastify, GraphQL Yoga, BullMQ on Redis, Postgres with Drizzle, Next.js, Claude Sonnet with free Gemini and Groq models behind it, OpenTelemetry, Playwright and Jest.
Run it locally
Needs Node 24, pnpm 12 and Docker.
pnpm install
cp .env.example .env # set API_KEYS; models stay off until AI_ENABLED=true and their keys are set
docker compose --env-file .env -f infra/docker-compose.yml up -d --wait
pnpm db:migrate && pnpm db:seed --scale=small
pnpm dev # web on http://127.0.0.1:3000, api on :4000
Checks: pnpm test, pnpm lint, pnpm typecheck, pnpm test:int and pnpm e2e (both need Docker).
Repository map
apps/ the web app (Next.js), the api (Fastify and GraphQL), the worker (BullMQ jobs) and a mock TMS.packages/ shared code: the domain rules and money math, the LLM provider layer, the extraction pipeline, the database, API contracts, settings and telemetry.evals/ the golden invoice sets, the eval harness and every eval report; start at the eval card.infra/ Docker Compose for local services and the demo, with the monitoring config.loadtest/ the k6 load test and its results.scripts/ the demo backend's start, stop and backup commands.docs/ third-party notices and screenshots of the app.
Licence
MIT, in LICENSE. Outside work used here, and its licences, are in docs/third-party.md.
