04Writing
Writing
Engineering notes from real projects — production LLM systems, coding agents, and computer vision.
Grade the grader: an eval harness for an LLM IELTS examiner
My last post claimed a rubric-as-contract makes LLM grading reproducible. This one tests that claim on real production data — and finds the grader is a distribution, the fallback model is +0.34 band more lenient, and top scores drifted almost a full band in three weeks.
Seed-first AI: serving LLM explanations at near-zero cost
How ACT Learner serves grounded AI explanations to real users while calling the live model on only a fraction of requests — and why the prompt is the cache key.
One library, every agent: syncing AI-agent skills across 15+ tools
Coding agents each store skills in their own folder, their own way. Skills Manager normalises all of them behind one central library — the architecture, and why it's a Rust filesystem program wearing a GUI.
From speech to band score: grading IELTS speaking with an LLM
IELTS speaking is a subjective exam graded by a human on four criteria. EasyIELTS does it with an LLM — here's everything you have to pin down to make a fuzzy human judgement reproducible, and honest.