I build AI agents that run unattended, and the checks that catch them when they lie.
Portfolio: denzil10.github.io
- Agent fleet, built in public: live numbers from an unattended agent fleet, its decision log and real examples.
- FabBench: a benchmark of my own agents claiming success they did not have: 30 real claims with their evidence, five judge prompts and the model runs behind its results table.
- dev-tribe: a GitHub onboarding program where newcomers learn open source by shipping real pull requests.
- Tools: free tools I build in public, like Snack Bouncer, an Android app that blocks the pay button while junk food sits in the cart. Tell me what to build next.
- agent-setup: the skills, browser tool and guard hook that keep my Claude Code and Codex agents in line while unattended, each with the incident behind it.
Writing:
- My LLM judges passed a false 'verified live' 11 times in 12. My prompt fixes moved the miss.
- My coding agent's guard said no 17 times in a month. It was wrong 10 of them.
- My agents got more reliable when I gave them less to do
- 88% of my coding agent's usage goes on reading
- Letting agents drive a real browser on the Mac I'm typing on
- A prompt is not a control: two autopilots that forged their own evidence
Open to AI engineer and forward deployed engineer roles: hybrid in Navi Mumbai, or relocating within India or abroad. CV · LinkedIn · X @denzilnelsonai


