Projects
Local InferenceSep 2026 – now
Qwen3.8 27B on one RTX 3090 at about 146 tok/s, against about 31 for llama.cpp on the same card.
Live siteDebriefSep 2025 – now
A daily AI/ML briefing. Its triage step agrees with 119 of 120 hand-labeled stories.
Live siteTimbreMay – Sep 2026
Recommends books, films and games by how they feel. Paused, and there are no quality metrics yet.
Live site
Journal
Now
Updated
- Training a small model that checks claims about code against the files themselves. Frontier agents often act on assumptions they never check, which I think is part of why they get lost in large repos. The small model is a tool a larger one can call to ground an important assumption in evidence before it makes a critical decision.
- Finishing my CS degree at UMD and looking for a new-grad SWE or AI/ML role after December.
- Reading about automated trading on Kalshi. Research only, nothing built.