Jan 2026 → now · Associate Software Engineer, InternI keep a fleet of AI agents alive under load.
Reliability, code review and LLM cost on the internal agent platform used by 16+ engineering teams at a Series C data and AI governance company.
0+security review dispatches by the review service I built
~0repositories it reviews, on by default
0+isolated agent sandboxes started on a busy day
50%90%+platform run success, after root-causing 10 bugs
Where my work sits
MineReliability, capacity and admission
Instance caps, admission, cross-pod recovery, cold starts, evictions, and the numbers that say whether any of it works.
mine someone else's
March
April
May
June
July
August
September
October
Not mine
- The sandbox primitive itself, and the per-team agents that run on it.
- The scaffolding that gives each team its own agent deployment.
- The prompt-cache fix (I diagnosed it; a teammate shipped it).
- The org-wide cloud cost reduction and the support-ticket classifier.
04 · On my own timeAgentOS, the platform I built myself