Studio concept buildAI solutions
Manifest
A document-intake agent that turns shipping paperwork into structured records a person approves.
Client engagements are private under NDA. These systems are our own builds, shown to demonstrate the standard we deliver.
The problem it addresses
Freight teams retype the same shipment data all day. Bills of lading, commercial invoices and packing lists arrive as scans and email attachments, in mixed Arabic and English, and someone keys them into the operations system by hand. It is slow, it invites errors, and it does not scale with volume.
What the system does
- 01Accepts documents from email, WhatsApp or upload, and classifies each one by type.
- 02Extracts the fields that matter — parties, ports, containers, line items, totals — with a confidence score per field.
- 03Routes anything below the confidence bar to a review queue, where a person confirms or corrects in one screen; every correction feeds the evaluation set.
- 04Writes approved records to the operations system through its API, with an audit log of who approved what and when.
- 05Handles Arabic and English documents in the same pipeline.
Stack
- TypeScript
- Node.js
- Anthropic (Claude)
- PostgreSQL with pgvector
- AWS (S3, SQS)
- GitHub Actions
Architecture
A queue-backed pipeline: object storage in, OCR and model-extraction workers, PostgreSQL for records and the evaluation set, a review UI in front, one webhook out.
How we build this for you