How to measure a production classifier when there is far too much traffic to label by hand: a blind LLM labeller does the volume, a verified replay attributes the errors, honest statistics bound the number, and people review every disagreement.
How I used deterministic contracts and LLM-backed self-evals to make PostCraft generate grounded technical portfolio posts without inflating what the code actually supports.
How I designed Better ArchNews as a layered Python system that enriches technical articles with LLM intelligence, exposes them through MCP for token-efficient retrieval, and turns …
I built a lightweight pipeline to collect job descriptions, normalize them into markdown, analyze them with LangChain and a local LLM, and score them against my CV. The goal is not …