I build harnesses for coding agents, with recovery and termination rules, evaluation harnesses, and context kept within a fixed budget. My independent projects include a Python harness for Claude Code, Cursor, and Copilot, with 29 lifecycle hooks and a task CLI.
For the harness, I developed agent loops that run parallel agents in isolated worktrees and use reviewer agents to gate each lane. I analyze transcripts and failed tool calls, then turn recurring failures into hooks, tests, or gates.
I also built working LLM-agent demos, including a football playbook app that turns a play call into a diagram and a fitness app that scores generated workouts offline. At ING Turkey, I’m the sole administrator of four on-premise production platforms for regulatory reporting, and led installs, upgrades, and a 10+ TB migration.
