Getting a computer vision model into production

Validation accuracy is a starting point, not a finish line. Production introduces lighting changes, new camera angles and inputs your training set never contained.

Version the data, not just the model

When results drift, the first question is what changed. Without dataset versioning you cannot answer it, and you end up retraining blind.

Monitor inputs, not only outputs

Track the distribution of what the model receives. Input drift shows up well before accuracy metrics degrade enough for anyone to file a complaint.

def log_inference(image, prediction, confidence):
    metrics.histogram("input.brightness", image.mean())
    metrics.histogram("pred.confidence", confidence)

    # Sample low-confidence cases for human review.
    if confidence < 0.6:
        review_queue.push(image, prediction)

That review queue becomes your next training set — the loop that keeps the model honest over time.