Writing
RSS- 30 Oct 2026 ·Machine learning · 9 min readI trained a 350M Python model from scratch on free computeOphyx is a 350M-parameter Python code model I built from scratch on Kaggle's free quota. How it works inside, every training stage, where it beats models its size and how far it is from frontier models.
- 14 Oct 2026 ·Developer tools · 4 min readI built a remote for the coding agents in my tmuxI built a remote for the coding agents already running in my tmux. Here's how it reads them and why there's no server.
- 5 Oct 2026 ·Machine learning · 13 min readParakeet Redux: is a 178 MB speech model useful in the real world?Moondream squeezed NVIDIA's Parakeet TDT 0.6B v3 into 178 MB with ternary weights. I tested it against its original, five other open models and three cloud APIs on a laptop and a Tesla T4: what it keeps, what it gives up, and where it's worth using.
- 28 Sept 2026 ·Machine learning · 6 min readFine-tuning a 0.8B model to beat hosted LLMs at one jobI fine-tuned Kev-0.8B, an open System-One model, to decide when a new fact makes an old one outdated. On held-out tests it beat every hosted LLM I tried, for my use case.
- 20 Aug 2026 ·Engineering · 7 min readBuilding a portfolio that feels fastThe engineering behind my portfolio: cached server-rendered pages, a small CMS with safe live previews, a browser terminal with six games, and analytics that count visits without tracking people.




