Collaborators
Open to research replication and OSS collaboration. These are personal projects.
Fifty RL steps restored output diversity across nine held-out tasks, no benchmark cost.
Prompt optimization mined ten verified fixes from the model; a LoRA on those ten generalized.
Linear probes recover claim-level correctness signal that structured confidence misses.
What a panel-of-experts scaffold does to RL-trained reasoning and search.
Train your own ChatGPT on Apple Silicon. MLX port of nanochat.
Adjust Night Shift intensity from the menu bar.
Claim-level correctness probes on Llama hidden activations.
Panel-of-experts scaffold vs native thinking on Qwen3-30B-A3B.
Multi-persona debate benchmark with traces, evidence, and a live demo.
RL environment for novel, evidence-grounded, falsifiable ideas.
Verifiers RLM environment for adaptive recursive search over long corpora.
Turn pasted text into files with one click.
Local-first MLX version of TTT-Discover AutoResearch.
Safari tab command center powered by Codex App Server.
Convert any Safari tab to clean Markdown in one click.
Reproducible LLM proof grading benchmark and API for math.
Ghostty tab renamer powered by Codex App Server.
Clipboard text transformer with case, encode, format, hash, and stats tools.
Read clipboard text, edit it, and save to any file format.
Per-volume disk space monitoring with color-coded usage bars.
See all listening TCP ports with one-click kill.
Monitor long-running processes and get notified on completion.
Manage Homebrew services. Start, stop, and restart with one click.
Gemma 4 on a 24GB MacBook: measured recipes, runtimes, and fallback paths.
One-notebook Gemma 4 RL on Apple Silicon.
Novelty-search-inspired autonomous search with run memory and agentic review.
Open to research replication and OSS collaboration. These are personal projects.