TactIntentNet
TactIntentNet bridges the gap between broadcast video and elite tactical analysis. Built for the AMD Developer Hackathon 2026, it transforms freely available StatsBomb 360 freeze-frame data into actionable tactical intelligence using a 3-layer GATv2 Graph Neural Network. It detects real-time opponent deviations via Gaussian Mixture Model fingerprinting and generates coaching alerts through a Qwen 2.5 1.5B LLM — all running on AMD Instinct MI300X. Elite tactical analysis is locked behind expensive proprietary tracking systems. Developed TactIntentNet, an open-source system to decode football tactical intent from broadcast video using causal graph neural networks on AMD MI300X. Democratizes tactical analysis by providing actionable tactical intelligence from freely available StatsBomb 360 freeze-frame data.
6
Tech Stack
5
Key Features
Technologies Used
Key Features
- ●Broadcast Only: Works with free video — no proprietary tracking hardware required.
- ●Causal Reasoning: Learns player-to-player influence weights, not just correlations.
- ●Real-Time Inference: 12ms inference latency at 80 fps on AMD MI300X.
- ●Counterfactuals: Drag any player to a new position, see intent shift instantly.
- ●LLM Coaching: Qwen 2.5 generates plain-English tactical alerts.