Notes
Short-form. 300–1,500 words. Mix of AI/ML deep-dives, backend post-mortems, and paper summaries where the "what's still unclear" section is the point.
- I quantized a 72B LLM and then proved it didn't get dumber: benchmarking perplexity with llama.cpp2026-08-06
Qwen2.5-72B-Instruct compressed from 145GB to 60GB with Q6_K, then measured instead of trusted — the perplexity delta on WikiText-2 came out at +0.0057, five times smaller than the measurement uncertainty.
- I built a CI/CD pipeline that fixes itself — and remembers how2026-08-04
A Mac Studio, a local LLM, a custom YOLO model, and a vector database keeping my deployments healthy without me. Blue-green + dual-signal QA + fix matching with staleness tracking + a Telegram approval loop.
- Anti-bot defences in 2026: a practitioner's view of what holds up and what doesn't2026-04-20
The anti-bot stack has layered from IP reputation to fingerprinting to behavioural ML. Here's what each layer actually catches, where the defender economics still work, and what vision-capable agents are starting to change.
- DOM-first with vision fallback: architecture of a production browser agent2026-04-20
A two-phase design — cheap DOM locators first, vision-model fallback when the DOM doesn't answer — that made a browser-automation agent fast enough to run at scale and robust enough to handle canvas, WebGL, and selectorless UIs.