Building Tenslam MiniLM: 29.9M to 60.1M Parameters, 791M Tokens, and an Evaluation Bug
An engineering write-up on training two decoder-only Transformers, recovering an inference export, and fixing misleading validation.
Writing
Engineering notes, technical essays, and working ideas across AI, software, and science.
What I write about
This section is for applied technical writing, not content for content's sake. Expect notes on computer vision, local models, backend systems, AI tooling, product lessons, and occasional science-heavy curiosity.
If you want to follow new posts, the site has an RSS feed at /feed.xml.
From Pixels to Humanoid Motion: A Technical Pipeline for Video-to-Robot Learning
A technical pipeline for turning RGB video observations into executable humanoid robot behavior.
Why I’m Building Motion Intelligence
Why structured movement data matters, and why video is only the starting point.
Running Small LLMs Locally on Limited Hardware
What local model experiments taught me about speed, memory, and practical constraints.
Building a Real-Time AI Gym Trainer
Lessons from turning a mobile camera into a feedback loop for training.