Stats Perform

Massive proprietary data asset (7.2 PB) focused on sports/Opta: This is not just big data for the sake of it — the size implies long-running, high-granularity time series (40+ years of global coverage), multi-modal sources (event logs, tracking, video) and extensive labels. That volume enables training specialized models that general public datasets simply can't reproduce.

Eight "foundation sports AI models": Instead of one general LLM/CV model, they appear to have invested in multiple domain-specialized foundation models (likely covering event prediction, trajectory forecasting, CV-based player detection/pose, natural-language generation for commentary, betting/prediction models, recommender/engagement models, and potentially anomaly/differential analysis). This multi-foundation approach is a technically distinct choice: vertically specialized backbones per subtask, then assembled into product modules.

Productized, composable inference surface (200+ software modules): Claiming 200+ modules implies a microservice/feature-store + model-as-a-service architecture that exposes small, reusable capabilities (e.g., xG calculation, pass chain detection, highlight identification, personalized feed generator). That leads to rapid recombination of capabilities for bespoke publisher/broadcaster integrations.

Tight integration of research + production: 'In-house researchers' plus large historical coverage suggests a closed loop where research outputs (new metrics, models) are quickly operationalized across modules — a nontrivial MLOps pipeline connecting experimentation, feature engineering on massive historical corpora, and low-latency production usage.

Real-time, latency-sensitive delivery to broadcast and betting partners: The commercial focus (broadcasters, bookmakers, live widgets) implies engineered solutions for streaming ingestion, near-real-time feature computation, and high-throughput low-latency inference — an implementation challenge beyond batch model hosting.