About

I study what post-training actually changes in small open-weight language models, and what it quietly breaks.

I co-founded Experimental Machines, which benchmarks inference across hardware, and Experimental Intelligence, which does low-level machine learning: two small open research efforts. My current work is two frozen studies. OpenGrad Study 001 post-trained Qwen3.5-2B for tool-call policy: call F1 rose from 0.6264 to 0.7548 on a pre-registered internal partition, and the promoted model refused every one of 1,319 zero-shot GSM8K questions, a regression the promotion gate could not see. Small-Mind Study 001 asked how much of a ~2B model's long-horizon memory gap retrieval and LoRA post-training can close; abstention on unanswerable questions rose from 13.75% to 71.25%, while absolute recall stayed under 20%.

I build my own research infrastructure too. OpenPapers is a provenance-first MCP server for finding and checking the sources behind a claim, and both studies publish a claim audit and errata alongside their frozen reports.

I study Computer Science at the Technological Institute of the Philippines.

Everything on this site links to a repository, a result file, or a write-up.