AI Developer Automation System

The Problem

Manual Sentry error triage consumed hours of developer time. Engineers had to read stack traces, identify root causes, locate affected files, and write fix PRs — entirely manually.

Architecture & Approach

Built a webhook-triggered pipeline where Claude Sonnet analyzes production errors and auto-drafts a Bitbucket pull request with root-cause analysis and a fix plan. Designed a multi-model cost-optimization strategy — cheaper Haiku for file discovery, more capable Sonnet for code generation — with structured tool use (typed JSON schemas) replacing fragile regex parsing. Security-first data handling: middleware strips .env files, settings.py, and credential patterns before any code reaches the LLM; all output is draft-only, with no auto-merge.