The problem
After training the models behind SwishAI and RotoAI, I had learned the lesson every Computer Vision practitioner learns the hard way: the bottleneck is almost never the algorithm. It's the data, duplicated images, overlapping labels, class names that disagree between sources, splits that don't hold up. And the tooling to fix any of that was either cloud-based, paid, or a pile of one-off Python scripts.
I wanted one place, running entirely on my machine, where I could look a dataset in the eye before trusting it with 48 hours of training. No uploads, no subscriptions, no data leaving the computer, which also makes it viable for the sensitive industrial datasets I encounter in my day job.
Four modules, one loop
Dataset Engine is organised as four modules that close the full data-quality loop, from seeing your data, to measuring it, to combining it, to actively improving the model trained on it.
The part I'm proudest of
The Improver closes the active-learning loop that usually requires an MLOps platform: test the model where it actually fails, correct exactly those frames in a Photoshop-style editor built on Konva, and export them straight into the next fine-tuning run. It turns “my model misses the ball in backlight” from a shrug into a 15-minute fix.
Under the hood it's a FastAPI backend doing the heavy lifting, Ultralytics inference, OpenCV frame extraction, YAML dataset management, with a React + Vite frontend fast enough to browse thousands of local images without ever feeling like a web app.
