A business intelligence tool that scrapes a target site, runs a nine-stage analysis pipeline with LLMs, and emails PDF + DOCX reports, with per-stage quality scoring, SSE progress, and an offline eval harness that fails the build on regression.
How it knows: Every stage carries two independent quality scores (a deterministic completeness contract and an LLM critic), and a seven-fixture eval harness, one of them a government site included as a hallucination canary, fails the build on regression.
Nicholas Smith - all work