AI systems engineer. Building Nisayon.
I study how to diagnose failures, test corrections, and decide what the evidence supports.
Nisayon · Repository · Interface · Contributions
Experiments for diagnosing robot-policy integration failures and testing corrections on fresh conditions. The Python SDK retains replayable evidence; savings in total engineering work remain unproven.
Atlas · Private development
An agent execution and evidence system studying when to execute, verify, repair or stop. Comparative benefit remains unproven.
Telos · Repository · Interface
Research on coding-agent patches that pass graded tests but fail additional checks. Findings remain limited to the tasks and reference implementations studied.
Sentinel · Repository · Interface
Runtime monitoring and recovery around frozen driving planners. Closed-loop studies report benchmark gains, intervention costs and null transfer.
Reiyah · Repository · Interface
An offline engine for comparing perception systems under incomplete or disputed reference evidence. Returns checked loss bounds; reduced total effort remains unproven.
Inbar · Repository · Interface
Research on diagnosing physical causes through tests that distinguish competing explanations. No physical result has been established.
Odeya · Repository · Interface
Architecture for a research engine linking questions, evidence and separately assigned verification. Contracts and fixtures exist; the engine is not built.
Aweb · Earlier work · Website
Governed AI-agent orchestration, connecting tools, durable workflows and operator approvals.