All work
Agentic AI · Freelance

An agentic system that runs a procurement desk

One product input triggers autonomous supplier research, tariff compliance classification and quotation outreach — with a human approving anything that reaches a real supplier.

Client
Sambac Procurement Intelligence LLC — Alabama, United States
Role
AI Engineer (Freelance, Remote)
Timeline
Jul 2026 – Present
Status
In production
3 → 1Manual stages collapsed into a single workflow
ZeroOutbound emails sent without operator approval
GroundedClassification backed by official US trade data

The problem

Sourcing a product internationally is three separate research jobs wearing a trench coat. Someone has to find credible suppliers, then work out how the product is classified for export and what tariffs apply to it, then write to each supplier and chase quotations. Each stage is manual, each one depends on the last, and the whole chain restarts every time the product changes.

Sambac wanted that chain to run as one workflow, without giving up the judgement calls that matter — which suppliers are worth approaching, and what actually gets sent to them.

What I built

An agentic system on LangGraph that takes a single product input and drives the whole pipeline end to end. The graph decomposes the request into research tasks, runs them, and carries the results forward into the next stage rather than handing a human a pile of tabs to reconcile.

  • Supplier research across web search via Tavily and Serper, with results consolidated into a ranked shortlist
  • Compliance research that resolves the product to its Schedule B classification and checks the applicable tariff position
  • Quotation outreach that drafts and dispatches supplier emails once a human has signed off

Grounding it in real trade data

Export classification is exactly the kind of task a language model will answer confidently and wrongly. Rather than trusting the model, I integrated the US Trade.gov APIs so Schedule B classification and tariff compliance checks resolve against authoritative government data.

The model decides what to look up and how to interpret it. The numbers come from the source of record. That distinction is the difference between a demo and something a business can act on.

Keeping a human in the loop

Automation that emails suppliers on your behalf is automation that can embarrass you at scale. I designed a review layer where operators approve the supplier shortlist and every outbound quotation email before it sends.

The agent does the hours of research; the operator keeps the last word on anything client-facing. You keep the speed without handing over the relationship.

Stack

LangGraphPythonTavilySerperTrade.gov APIHuman-in-the-loop

Need something like this built?

I take on freelance briefs and part-time engineering work from Melbourne, Australia.