Edge AI engineer working on AI safety. Creating monitoring methods and evaluations for robotics and agents.
- Seen vs. Followed (code). Do robotic foundation models act on text in the scene? Preliminary results on a DROID sorting task in Isaac Sim.
- I Spy Agent (code). Small decision models as black-box monitors for covert agent behaviour: sabotage hidden in files (LinuxArena) and collusion through covert table talk (NARCBench-Stego). Extends my Apart Research sprint report.
- Stance drift in multi-turn agent conversations and steering (code). An AI math tutor that must not give away answers, a student whose tone changes, and what a monitor can see. Continues my Apart Research project.
- mirage-probes-steering (code). Do probes and steering vectors catch a vision-language model that answers about an image it never saw?
- Sirens for Agents (code), parked. Resampling-derived, action-level honeypot discovery. Continues my Apart Research project.
- Layterms. A nonpartisan walk-through museum of AI. Try the game: Run the Lab!, a three-minute game where you run an AI lab.
- TibetianBuddhismTransciptionPipeline. Domain and capacity in low-resource Tibetan speech recognition: fine-tuning Meta's MMS.
- flywheel-grader-qc. A short project to learn Flywheel and quality control in medical imaging for eyes.
- zero_to_hero_notes. Notes and code following Andrej Karpathy's Neural Networks: Zero to Hero.