The problem
Every engineering issue started with a long, manual investigation: figuring out what broke, and which algorithm in the codebase was responsible. That initial analysis took 5 to 10 days per ticket and ran on handoffs between developers, subject matter experts, and QA.
The approach
I led discovery, design, and delivery of a multi-agent AI system that does this triage autonomously. The system is built on retrieval augmented generation and the Model Context Protocol: agents retrieve the relevant context, reason over the issue, and pinpoint the root-cause algorithm in the codebase. I defined the requirements and coordinated delivery with the developers, SMEs, and QA who owned the old process, and shipped end to end within a 3 month window.
The outcome
- Initial issue analysis went from 5 to 10 days down to under 10 minutes per ticket.
- The system identifies the root-cause algorithm at roughly 70% accuracy.
- Manual handoffs in the triage process were eliminated.
Why it matters
This was the most autonomous AI system I shipped at HERE, and the clearest proof of a pattern I keep coming back to: find the slow, manual, expert-dependent step, and rebuild it as a system with the expert's judgment encoded in it.