Docs

Xveris, end to end.

Xveris turns a shipping documentation inbox into a short, ranked list of work: it classifies every email, reads the SI and the draft BL in any format, compares the seven shipment fields, and escalates whatever it cannot decide, with the evidence attached.

Architecture

Sources. One interface serves the organisers' dataset, their evaluation server and Gmail (read-only OAuth), so the pipeline is identical for all three.
Pipeline. Classify (Claude + a rule engine, cross-checked) → read attachments (text, Excel, Word, PDF, vision for scans) → extract seven fields by meaning → compare deterministically → triage into a next action.
Store. PostgreSQL when a database URL is set, JSON files otherwise. Reviews and retries update one email in place.
App. Next.js serves the landing page, the dashboard, the inbox map, the email pages and the API, and runs whole-inbox jobs in the background.

Run it locally

npm install
cp .env.example .env.local     # add ANTHROPIC_API_KEY for AI + vision
npm run process                # read the sample inbox from the command line
npm run dev                    # http://localhost:3000
npm test                       # the pipeline tests, including all 520 emails

Self-evaluation

With the organisers' Docker server running, point Xveris at it and submit the run for the official scoreboard:

npm run process -- --source http://localhost:8080 --submit

The dashboard exports the same file from the left rail (“JSON”), keyed by email id, in the format ofsample_submission.json.

Source code

The repository link is published with the hackathon submission. It contains the full README, the pipeline, the tests and the deployment files (Dockerfile for Cloud Run, environment template, scheduled re-read endpoint).