The field guide
Running vision AIin production.
Running vision AI in production.
Five chapters, from the questions to ask before you label a single frame to a model holding its accuracy in year two.
- Chapters
- 5
- Lessons
- 14
- Reading
- 100 min
Looking for the product itself? Read the docs
// Chapter 01
Before you label
Most vision projects fail before a single box is drawn, because nobody agreed what the model is supposed to answer.
What a vision model can and cannot see
The honest boundary. What is a detection problem, what is a measurement problem, and what is not a vision problem at all.
guide 9 min
guide9 minBoxes, polygons, or masks: pick by the question you are asking
Annotation type is not a quality setting. It follows from the question, and picking wrong costs you the whole dataset.
field note 6 min
field note6 minHow much footage you actually need
Why the answer depends on how varied your world is rather than how big it is.
guide 6 min
guide6 min
// Chapter 02
Building the dataset
Labeling is the part everyone budgets for and still underestimates. The goal is not speed, it is a dataset you can defend.
First labels in five minutes
Upload footage, describe what matters, get a labeled set back. The shortest path from zero.
reference 5 min
reference5 minAuto-labeling from a sentence
How a plain-language prompt becomes a class schema, and where a human still has to look.
field note 7 min
field note7 minPrompts, classes, and the verification pass
Writing prompts that produce consistent classes, and the review step that catches what the model was confident about and wrong.
reference 7 min
reference7 min
// Chapter 03
Getting it into the world
Where a model runs decides your latency, your bandwidth bill, and who can see your footage.
// Chapter 04
Day 90
This is the chapter the rest of the industry does not write. Your model was accurate at launch. The world it watches has moved since.
What model drift actually is
Not a bug and not decay. Four distinct failure modes with different signatures, and how to tell which one you have.
guide 9 min
guide9 minWhat a confidence score actually means
Why 0.9 does not mean ninety percent, what a threshold actually chooses, and how to spend the score where it earns the most.
guide 7 min
guide7 minHow to catch model drift before your customers do
The signals that move before accuracy does, and the ones that only look like drift.
field note 7 min
field note7 minAttach a model to a feed and describe the alert
Turning a trained model into an operation: what to watch, what to escalate, and what to ignore.
reference 6 min
reference6 min
// Chapter 05
Closing the loop
Retraining is not a rebuild. Done properly, every correction an operator makes is a deposit into the next version.
Retraining without starting over
How corrections become training signal, what to do about the old labels, and how to know the new model is actually better.
guide 8 min
guide8 minCorrosion detection on 500kV insulators, end to end
One loop, start to finish: footage in, labels, model, alerts, drift, retrain.
field note 10 min
field note10 min