ai/ml engineer

A physical fly arena, a Doom experiment, and 93 moves of connectome chess. Replay the game and inspect the measurement failures.

A coding SLM training framework: attempt, test, repair, then train on the corrected spans. Verified repairs feed a difficulty-aware curriculum.

A weight-space ablation engine for studying LLM refusal directions, with evaluation tools to check the effects on model behavior.

Agents debate from distinct personas and vote on an answer. Session transcript hashes are recorded on Solana for an auditable record.

A multi-model council and proxy for frontier models, alongside deepfake detection and adversarial defense tools.

Graph-based LLM red teaming with multi-turn attacks, a state-machine evaluator, and a dashboard for inspecting results.

Three agents scan code, make edits, and review the changes before opening a PR. Cross-repository memory carries findings between runs.

An inference-backed agent that keeps coding agents working toward persistent goals. Model reviews, a claim ledger, and independent tests check whether the work is actually done.

Dataset intelligence with 55 catalogued datasets, 53 automated assessments, and 16 proxy-training outcomes, including six divergent runs retained as evidence.

A social feed for developers to share updates, publish releases, and follow builders. GitHub sign-in, public feeds, and live updates.