I build AI systems and test them like experiments. I'm based in New York, and looking for AI/ML engineering, research or AI training roles.

Contact: LinkedIn

What I work on

• Agent-based simulation and reinforcement learning: artificial societies whose people learn from experience, with every result tested by pre-registered, controlled experiments.
• LLM tooling: getting several language models to disagree usefully, on free API tiers.
• Research rigour in code: exact replays, frozen analysis plans, held-out data and sham controls.

Artificial Civilization Laboratory · Python, NumPy, SciPy An agent-based simulation of a small prehistoric people. Their nature (bodies, drives, memory, learning) is programmed; their culture is left to emerge, and to be measured.

• About 60 studies (controlled experiments, read-only audits, evidence reviews). Experiments are pre-registered and frozen before they run, with a discovery set, a held-out replication and a sham control; every result traces to a committed script.
• Reinforcement-learned habits (32 situations × 19 actions), and a bookkeeping method that splits every learned value into what was inherited, absorbed from a carer, learned, or copied in talk, exact to 10⁻¹⁵.
• Bit-exact replays guarded by CI (GitHub Actions, Windows), 56 tests including property-based tests that found a real flaw in the model's rules.
• A book written from the experiments' results: The World That Nobody Wrote, Part One (226 pages).

Unstuck · Python, Docker, Fly.io Asks several AI models the same question with conflicting instructions, merges what survives, then runs a drafter / critic / judge debate. Runs entirely on free API tiers.

Python · NumPy · SciPy · statistics (paired tests, bootstrap, multiple-comparison control) · reinforcement learning · LLM APIs · Docker · GitHub Actions · HTML/JS