
Offsides is a two-stage multimodal AI system that watches recent match footage, detects every player and ball using YOLO, annotates tactical overlays (formations, defensive lines, team compactness), and feeds these annotated frames alongside league statistics and market odds to Qwen3-VL 32B running on AMD MI300X via ROCm. The system identifies tactical signals invisible to traditional statistics: defensive shape deterioration, pressing intensity trends, and transition vulnerabilities. It compares its probability estimates against prediction market odds to flag potential mispricings. Built entirely on AMD infrastructure. The MI300X's 192GB unified HBM3 memory fits the 32B parameter model on a single device with no sharding, enabling real-time multimodal conversation at ~10 seconds per match analysis. The pipeline processed 280 UCL matches (6,500+ frames) and produced tactical assessments for 29 knockout-stage matchups. We backtested on 5 historical market upsets and the system correctly identified the winning outcome in 3/5 cases (60%) by detecting tactical edges the market discounted. For example, Dortmund's compact defensive shape against PSG's high line in the 2024 Champions League semi-final. Users interact via a Gradio interface: browse pre-computed analyses, ask follow-up tactical questions in natural language, or compare any two teams for a live VLM prediction.
10 May 2026