Daedalus - A 150M params LLM built for CPU inference.
My summer project: a small convolution-attention language model, the CPU inference trade-offs, and the things that didn’t work.
Writing & experiments
Posts
Notes on software, research, and the ideas I’m working through. These first posts take a closer look at three of my research projects, from the decisions behind them to the results and limitations.
Intent-Driven Dynamic Chunking: keeping answers together
Using predicted questions and dynamic programming to decide where a document should be split, before anyone searches it.
Acro Helper: contextual help without leaving the page
A browser extension that explains technical acronyms where they appear, and what a study with 25 participants taught me about the experience.