AI Lead-Generation Pipeline

Outbound research is slow and repetitive. I built this pipeline to automate the parts that do not need a person, feeding the Shadow AI assessment service I helped design for the same client.

Given a target company, it runs structured research with the Claude API and live web search to gauge how exposed the business is to uncontrolled AI use, finds the relevant decision-makers, and drafts a personalised outreach sequence in the right voice. The output exports cleanly into the tools the sales motion already runs on.

It is built defensively: caching and cost guards so a large run never becomes an expensive surprise, and a staged architecture so each step can be inspected and corrected on its own. Sixteen command-line tools, backed by 427 tests across 21 files, all running against mocked API responses so the suite costs nothing to run and needs no keys. Measured cost is roughly fifty cents to a dollar per ten companies researched.

The test count is worth a note, because it is the kind of number that rots. The repository's own README says 320. That was true when it was written; `pytest --collect-only` now finds 427. Every figure on this site is defined once, with the command that produced it, for exactly this reason.

What I will not claim is that it generates meetings. It is an engineering result (throughput, cost control and data hygiene) and the commercial outcome depended on follow-up work that sat outside the pipeline.