AIM — AI-Powered Football Coach
AIM is an AI football coaching platform. The product analyzes football drill footage using computer vision to provide tactical insights, player performance metrics, and coaching feedback in real time.
When I joined, AIM had a rule-based system that was brittle, inaccurate, and expensive to run. The task was clear: replace it with production-grade ML, and make it fast and affordable enough to scale.
What I Inherited
Fragile Rule-Based System
Hand-coded rules broke on edge cases. Outputs were barely usable, requiring constant manual correction. No path to improvement without fundamental redesign.
$2,500/mo Training Costs
Models trained on A100 instances at premium cloud rates. Unsustainable for a startup. Every retraining cycle burned through budget.
Licensing Risk
Models with restrictive licenses (AGPL, non-commercial clauses) created legal exposure for a commercial product.
10 FPS Inference
Processing at 10 FPS on 720p video — far too slow for real-time coaching. Analyzing a full drill took far too long for a responsive tool.
What I Built
I replaced the entire system from scratch. Custom models, optimized inference, efficient training pipelines, and a licensing audit of every dependency.
Custom Object Detection
Replaced the rule-based player/ball detection with RF-DETR (Real-Time Detection Transformer). Custom-trained on football-specific data with aggressive augmentation. RF-DETR's architecture gave us better accuracy than YOLO alternatives while maintaining speed.
Pose Estimation Pipeline
Integrated RTMPose for real-time player pose estimation. This feeds into movement analysis, fatigue detection, and tactical formation tracking. Optimized the model for the specific poses and camera angles common in football footage.
TensorRT Optimization
Converted all models to TensorRT engines with FP16 precision. Custom CUDA memory management to eliminate allocation overhead. Batch processing with optimized memory layouts. Result: 1 minute of 720p video analyzed in 7 seconds on a single L4 GPU.
Training Cost Elimination
Designed efficient training pipelines that run on free-tier L4 GPUs instead of expensive A100 instances. Smart data loading, gradient accumulation, and mixed-precision training. Monthly training cost: $0.
Custom Neural Networks for Drill Analysis
Built custom neural networks for analyzing individual player drill performance. Movement pattern recognition, technique scoring, and spatial analysis of player positioning throughout each drill.
Apache 2.0 Migration
Audited every model and dependency for licensing. Replaced all restrictively-licensed components with Apache 2.0 alternatives. Zero licensing risk for commercial deployment.
What Powers It
Models
- RF-DETR (Object Detection)
- RTMPose (Pose Estimation)
- Custom NNs (Drill Analysis)
- TabPFN (Tabular Predictions)
Inference
- TensorRT (FP16)
- CUDA (Memory Management)
- ONNX Runtime
- Triton (Inference Server)
Infrastructure
- GCP (L4 GPUs)
- Docker
- Python / FastAPI
- Weights & Biases
Data
- Custom Annotation Suite
- Roboflow
- Video Processing Pipeline
- Synthetic Augmentation
What Changed
Beyond the Code
I didn't just write ML code. I drove key strategic decisions that shaped the product's direction:
- Model selection — chose RF-DETR over YOLO based on accuracy/speed tradeoffs specific to our use case
- Architecture decisions — designed the full pipeline from data annotation to inference
- Cost strategy — proved we could train on L4s instead of A100s, saving thousands per month
- Licensing audit — identified and replaced every restrictively-licensed dependency
- AI strategy — scoped new features and advised on what's feasible vs. hype
- Code review — reviewed other engineers' ML contributions for quality and correctness
- Stakeholder communication — translated technical progress into business language for non-technical stakeholders
Need something like this?
I build production ML systems that are fast, affordable, and legally clean. If you have a computer vision or edge deployment challenge, let's talk.