Field notes
Writing on building computer-vision systems that run reliably in production.
- 3 min read
Why your vision model gets worse over time
A detector that worked on launch day can quietly degrade for months before anyone notices. The model didn't change — the world did. Here's what drift looks like on a real camera, and how to catch it before it costs you.
- MLOps
- production
- monitoring
- 3 min read
When computer vision is the wrong tool
Vision is the right answer less often than the hype suggests. A few honest tests for whether your problem actually needs a camera and a model — before you spend a budget finding out it didn't.
- computer vision
- scoping
- honest
- 3 min read
What a computer-vision proof of concept should actually prove
A good PoC isn't a demo that works once on a clean clip. It's a cheap, fast answer to one question: will this survive your real conditions? Here's how to scope one so it tells you something true.
- computer vision
- scoping
- deployment
- 2 min read
Your cameras are already a sensor network
Most industrial sites have spent years installing CCTV for security. The same feeds can answer operational questions too — if you treat the camera as a sensor, not just a recorder.
- computer vision
- industrial
- deployment
- 2 min read
Latency budgets: the number that decides your video pipeline
Before you pick a model, pick a latency budget. How many milliseconds you have per frame quietly determines almost every other decision in a real-time vision system.
- real-time
- video pipelines
- GPU inference
- 2 min read
Why computer vision fails in production (and how to ship it anyway)
A model that scores 95% on your test set can still be useless on a live camera feed. The gap between a notebook and a running system is where most CV projects quietly die.
- computer vision
- MLOps
- production