mAP@50 is not a promise: reading detection metrics before you ship
A detection model can score 95% mAP@50 and still fail the task it was built for. What the metric measures, what your pipeline actually needs, and why the gap matters.
Read the noteEngineering notes
Working notes on building AI systems that have to be right
Short, specific write-ups from the systems I build: how I evaluate document-AI pipelines, how I design review steps that people can actually use, and what I got wrong the first time.
Human-in-the-loop is a product decision, not a fallback
Routing low-confidence predictions to a human is the easy half. The hard half is designing the review step so it is faster than the manual process you replaced.
Read the noteWhat I changed after my first RAG system
Building a retrieval-augmented assistant for warehouse operations taught me that most of my problems were retrieval problems wearing a generation costume.
Read the note