After ~9 years building safety-critical systems - triple-modular-redundancy UAV avionics, regulated medical-device SDKs, secure banking - I now apply that reliability discipline to AI agents, at the point where a model's output becomes a consequential action (send, pay, modify, disclose, tool call.
Current work - Agent Failure Microscope. A measurement lab for AI agents
at the executed-action boundary. Each decision-before-action is normalized as a
verifiable DecisionEvent - authority, argument-level provenance, deterministic
gates, world-state deltas, optional internal signals - to test which layer
actually reduces costly harm, and whether defense-in-depth layers fail
independently or together under adaptive attack. → repo
Open source - UK AISI inspect_evals. A scorer-validity audit of the
evaluation suite through an action-boundary lens: does proposed_action → typed args/authority → real tool/world-state effect → scorer verdict hold, or is a
substring/count proxy standing in? → pull requests · audit
BitGN Arena competitions - blind, deterministically-scored agent benchmarks. 70.3 / 100 on ECOM1 (agentic commerce); hardened a personal-assistant agent against prompt injection and boundary violations.
Background: 10+ yrs senior software engineer (systems, backend, iOS).
nefedov.d.d@gmail.com · LinkedIn · Telegram @nfdvd
Author of ModernRoboticsCpp — 500+ ★.