Customs declarations, filled by an agent, checked by a person.

Role: Solo build, freelance
Client: Customs broker, Hong Kong
Stack: LangGraph · Gemini extraction · review app
Status: Delivered as source code

move across the tray · click to check a flag

done · 8/8 · 2 flagged

fig. 01 — a sample of eight fields, filed one by one; a raised flag waits for a person

The problem

Every shipment through Hong Kong needs a government declaration with dozens of exact fields. The broker's team read the shipping papers and typed each one by hand, about 30 to 40 minutes per declaration.

What I built

A 12-step LangGraph pipeline that reads the shipping PDFs, checks the courier's tracking page, does every money calculation in code instead of the model, validates each value against the government's reference lists and writes the file the system accepts. Anything it is not sure about gets flagged, so a person only checks those. I also built the review web app and the test set of 17 hand-labelled shipments.

Outcome

45.8 → 88%field accuracy on 17 hand-labelled shipments, first run to best run
~30 → 1 mintyping replaced by 30 to 60 seconds of processing, then a short check of the flagged fields
6 weekssolo, with 329 automated tests

Next projectPocketPulse →