Extreme Vision built a unified AI feature data platform on MatrixOne, cutting model development cycles by 30% and boosting feature reuse by 40%.
Extreme Vision, founded in June 2015, is a computer vision algorithm provider operating China's first computer vision algorithm platform, bringing together over 400,000 developers worldwide and serving more than 3,000 enterprises and research institutions.
Extreme Vision builds its business on machine-learning algorithms, and feature engineering — turning raw data into feature vectors through cleaning, extraction and selection — is the step that most directly determines how well a model performs, and the one that consumes the most time.
Within that work, storing and managing feature data is a discipline of its own. Conventionally AI engineers keep features as files and load them into training or inference through code, and Extreme Vision worked the same way.
Over years of development the company accumulated a large body of vertical-domain algorithms, and the feature data behind them grew vast and scattered — most of it living only inside an individual engineer's own workflow.
Meanwhile the scenarios themselves began to overlap and merge. A good deal of that feature data was reasonably general, but because it was not managed anywhere, every new model meant repeating the whole feature-engineering sequence from the start — a real cost in engineering time and compute.
Extreme Vision set out to build one unified feature store platform, to raise both the efficiency and the reusability of feature engineering. Three capabilities defined it.
It had to hold data of every format and modality. Feature-engineering algorithms differ widely and their outputs do too, so the platform needed strictly-schema'd structured CSV, flexibly-schema'd JSON, text, images, and the vectors produced by embedding algorithms — together.
It had to support efficient sharing and secondary processing. Features are built by different engineers and then reused in other projects, so feature data needed to land in the store and be published to colleagues quickly, with those colleagues able to derive new feature sets on top of it.
And it had to transform data. Different algorithms have different requirements for training and inference input, and even within one project features frequently need reshaping, so transformation had to be a platform capability rather than a script somebody keeps locally.
With the feature platform built on MatrixOne, engineers no longer stand up a separate feature-engineering effort for every model.
Pre-computing and storing features cut the model development cycle by 30%, from an average of four weeks to 2.8 weeks. Feature reuse removed 40% of the feature-engineering pipeline work per model iteration, as duplicated pipelines were identified and replaced by existing feature engineering or ready-made feature data.
Sharing features across the business also made it markedly more flexible: when a requirement arrives in a different department or project, cross-team and cross-project access to feature data raised overall data utilization by 40%.
A feature store is a common requirement in machine learning, and this build offers the wider industry a working reference for what one should look like.