work/Open source

squirrelscan

Agent architecture · Claude · Python

An AI audit agent that runs 200+ rules without hallucinating findings.

Problem

Website auditing means running a large structured rule set across HTML, performance, accessibility, SEO, then producing findings accurate enough to stake a client recommendation on. The challenge: layer AI intelligence over a deterministic rule engine without introducing hallucination into the output.

Approach

Built the AI layer as a structured agent, not a generative system. The model doesn't generate audit findings — it interprets deterministic rule outputs and synthesises them into prioritised, contextualised recommendations. Retrieval grounded in actual page data, not model memory. Every finding carries a source trace.

  • Rule engine (deterministic) cleanly separated from AI layer (interpretive).
  • Claude receives structured rule outputs, not raw HTML.
  • Output validation: every finding must map to a source signal.
  • Confidence scoring on synthesis-heavy outputs.

Result

Production audit agent running across thousands of domains. Findings accurate enough to publish without manual review — while flagging low-confidence outputs for human check rather than suppressing them.

Stack
Claude · Python · FastAPI · PostgreSQL

Take on a small number of engagements at a time. Mid-project and something's gone sideways is fine too.