I built the intelligence stack behind AIM’s AI football coach, from raw phone video and training data to the feedback a player finally sees.
AIM is a shipped AI football coaching product. A player records a drill on a phone; the platform checks the footage, reconstructs what happened, separates the drill into meaningful phases and repetitions, measures technique, scores the performance, and returns visual coaching feedback.
I joined while the product's first analysis path was still rule-based. My work grew into ownership of the full intelligence layer around the product: the data engine, perception stack, spatial and temporal reasoning, domain models, biomechanics, scoring, inference services, quality controls, and the visualization shown to users.
The first analysis path was rule-based. It proved the idea, but real footage has more variety than rules can cover, and separate outputs did not add up to one trustworthy account of the drill. The next step needed a shared data and model foundation.
Models trained on premium cloud hardware, so every retraining cycle had a real cost and iteration slowed down. Training had to fit a startup's budget.
Every model, weight file, and dependency had to ship under terms suitable for a commercial product, not only for research.
A one-minute drill took around six minutes to analyze. That latency did not fit a responsive coaching experience.
AIM is not one model. It is a chain of systems that must agree about the same player, ball, cones, drill structure, and scoring rules. I designed and built that chain so every stage produces a testable artifact for the next one.
Built a dedicated annotation and dataset platform instead of treating labels as disposable files. It covers the different kinds of labels the models learn from, with review workflows, batch tools, and model-assisted labeling so each round of training data is better than the last.
Added the checks that make downstream analysis reliable: whether the footage is usable, whether the player, ball and drill are in view, and whether the drill was set up as intended. The system understands the drill's layout, not just the objects in the frame.
Built the visual foundation for real-world phone footage: custom-trained models for the player, ball, cones and body pose, with tracking and cleanup that hold up when the footage is far from perfect.
Turned raw detections into consistent movement signals: where the player and ball are relative to the drill, how they move, and how the body moves with them. One shared layer feeds every downstream model, which keeps them consistent and easy to inspect.
Built specialized models that follow a drill over time: what the player is doing, where each repetition starts and ends, and what went wrong along the way. A shared training and release process turns experiments into models that can be deployed with confidence.
Connected body, ball and drill signals into technique measurements for every repetition, then into scores that show a player what drove the result and where to improve.
Connected every stage, from validation to the final report, into an asynchronous cloud pipeline, then optimized the whole path until one minute of 720p video took roughly seven seconds to analyze on a single cloud GPU.
Built the visual product layer that makes the analysis understandable: body overlays, player and ball trails, top-down views, live drill state, technique and score summaries, and graceful handling of uncertain or missing signals.
Brought training onto far cheaper hardware, added experiment tracking and release checks, and made sure every model and dependency is licensed for commercial use.
Analysis of one minute of 720p drill footage went from about six minutes to about seven seconds on a single cloud GPU. The same models now train on far cheaper hardware. The rule-based prototype became custom models trained and evaluated on annotated drill footage, and every shipped model is cleared for commercial use.
I was the hands-on technical owner for the intelligence layer. That meant making the model, product, and operational decisions fit together, not optimizing one component in isolation. In November 2025 I was promoted to AI Lead, overseeing model selection, feature scoping, and technical review across the project.
I build connected AI products like this: the data, models, domain logic, deployment, and user experience needed to turn an ambitious idea into something people can trust. To test a sports idea first, start with a one- to two-week feasibility sprint.