Ebin Babu Thomas Measures What AI Safeguards Actually Stop

Why it surfaced

Thomas publishes unusually concrete evidence: 434 of 434 simulated kill-point recoveries, 16 of 16 PDF conversions verified by render-back diffs, and a preregistered experiment whose reported effect survived after 65.8% of distress language was trained away. The portfolio treats limitations and failed tests as part of the work rather than burying them.

A personal AI-engineering portfolio documenting agent reliability, model-behaviour experiments, and open-source work with committed artifacts and failure reports.