Developer Builds Two-Stage Local AI Pre-Commit Hook to Cut Secret Scanning Noise

A developer created a pre-commit hook that combines regex scanning with a locally run large language model to improve secret detection in code repositories. The tool first applies broad regex patterns to flag suspicious strings in staged files, then passes only those flagged hunks to a local LLM via Ollama for a final verdict. Because the model runs entirely on the developer's own machine, no code or diffs are sent to external cloud services. The system distinguishes real credentials from placeholders and test fixtures, reducing both false positives that frustrate developers and false negatives that let real secrets slip through. Most commits bypass the slower LLM stage entirely, keeping the hook fast for everyday use.