Introduction

A large part of my work is educating and enabling people—whether they are technical teams, leadership, or new hires. Below are lessons I try to communicate to every new hire or mentee.

  • Offload What You Can, Learn What You Must
    Avoid turning computer vision models into Swiss Army knives. Separating logic from pattern recognition helps keep CV systems simple and accurate.

  • Do the Math
    While platforms like Ultralytics or Roboflow allow non-technical users to create computer vision models, the best results usually come from users who have a semi-academic understanding of the mechanisms that drive performance and the metrics used to quantify accuracy. For example, did you know that mAP—the most common object detection metric—is highly sensitive to image size?

  • Expect Ambiguity. Build Systems That Account for Uncertainty
    ML introduces probabilistic behavior into otherwise rigid systems (databases and traditional software stacks). In the real world, knowing the likelihood of correctness for a given prediction is almost more important than having a highly accurate model.

  • Active Learning / Continuous Learning
    Whatever you want to call it, do it. Otherwise, expect failure.

  • Annotation Standards
    Subjective annotation instructions and inconsistent standards for your annotation team will always result in noisy labels—and a confused model.

  • Data Science Is a People Problem
    Creating groundbreaking technology is easier than mastering the process of technology adoption by your customer. Know the metrics that define success for your customer, and be purposeful in enabling them to get the most value out of the product.