Work/2026
Automated Tariff Classification
Every item crossing a border needs a tariff code. Getting it wrong is expensive. This reads the product information and proposes a code with its reasoning, so a specialist reviews instead of starting from scratch.
- Role
- Associate data scientist
- Client
- PwC, via TenX
- Built with
- LangChain · OpenAI · Agentic workflows
- less manual effort
- 50%
The problem
Tariff classification is judgement work over a large, hierarchical code list. It does not parallelise by hiring, because the expertise takes years to build.
The aim was never to remove the specialist. It was to stop them starting from a blank page on every item.
What I built
An agentic pipeline: rather than one prompt returning an answer, the system works in steps — narrowing the category, gathering what it needs, proposing a classification, and checking that proposal against validation logic before surfacing it.
That structure matters because the reasoning is the deliverable. A specialist needs to see why a code was proposed to accept or reject it quickly.
The outcome
Manual effort on the workflow fell by around half, and turnaround improved correspondingly.
Consistency improved too — the same item now gets the same treatment regardless of who is reviewing it.