LaxForgeTechnology
Built on real lacrosse.
Not stock animations.
LaxForge runs a custom-trained AI shot classifier on your phone. Trained on 902 quality-filtered shots from college and pro creators across YouTube and TikTok — labeled, validated by hand, and filtered by a Gemini QA gate. It reads seventeen biomechanical features from your pose, the stick when visible, and the LaxPod sensor, then tells you what kind of shot you took and where the form broke down.
01 · The corpus
Real shooters. Real angles. Real noise.
Most lacrosse “AI” demos are recorded in a controlled lab. Ours isn't. We pulled video from creators across three platforms, covering broadcast angles, hand-held side angles, indoor turf, outdoor field, and mid-game footage. Each shot was labeled with release timing, phase boundaries, and 14 pose landmarks — including the stick tip when visible.
Long-form drills, broadcast clips, full-game footage. The bulk of the corpus.
Reels from college + pro creators. Tight angles, vertical framing.
Quick form drills + in-game highlights. 73% quality-pass rate — the cleanest source in the corpus.
02 · Shot types
Six shots, with the honesty about which ones are solid.
The model classifies every shot you take into one of six types. Held-out test accuracy below — the same numbers we look at internally, no rounding up. Minority-class shots are gated by corpus size, not model capability. We're actively collecting more bounce / BTB / quickstick footage to fix that.
The traditional release — stick high above the head, body squared.
Three-quarter release with the stick parallel to the ground at release.
Low scoop release — common around the crease and in tight quarters.
The stick passes behind the body before release. Tough to pick up cleanly.
Rapid release off a feed — no full crank, almost a redirect.
Low-trajectory shot intended to skip in front of the goalie.
03 · What it measures
Seventeen features. The same ones lacrosse coaches care about.
The model doesn't look at raw video. It looks at seventeen specific measurements — ten from your body pose (including the stick orientation when it's visible), seven from the LaxPod sensor — that map to the things coaches already grade. Open-sourced math: BIOMECHANICS_SPEC.md ↗
- Release angleHow vertical your stick arm is at release
- Elbow flexionHow extended the elbow is at release
- Shoulder external rotationHow cocked the throwing shoulder is
- Hip rotation (crank-back)Hip-shoulder separation at peak crank
- Hip rotation (release)How much the hips have rotated through
- Trunk tiltForward lean at the moment of release
- Stance widthFoot spread relative to hip width
- Front foot angleHow open or closed the front foot is
- Weight transferHow much weight has shifted onto the front foot
- Stick angleStick orientation at release — when visible (spec v0.4+)
- Shot durationTotal time from cock-back to release
- Wrist-snap timingTime from acceleration onset to release peak
- Peak accelerationMaximum g-force during the swing
- Wrist-snap ratePeak angular velocity of the wrist
- Stick pitch at releaseStick-head angle the moment the ball leaves
- Pre-release dipLowest stick pitch during cock-back
- Stick rotation at releaseYaw rate at the moment of release
No LaxPod? The pose-only path still classifies your shot — the seven sensor features are NaN-padded and the gradient-boosted model handles missing values natively. The pod just sharpens the picture.
04 · Accuracy
Numbers from the test set, not the marketing department.
159 shots set aside before training, never seen by the model until evaluation. Every number you saw on the previous tile comes from this held-out test. We surface a per-class confusion matrix in our internal admin so we can spot when a shot type is regressing — same numbers, no smoothing.
We refuse to ship a model that pretends to know what it doesn't. When a class doesn't have enough training data yet, we tell you (look at the bounce / BTB / quickstick cards above). When the model is unsure, the app falls back to a rule-based classifier rather than guessing.
05 · Where it runs
On your phone. Your video stays yours.
The classifier is a 686 KB ONNX model bundled with the app. Pose detection runs on Google ML Kit (also on-device). No video frame ever leaves your phone unless you opt in to a cloud session for replay on the web. Sensor data is yours — export or delete anytime.
See it on your own film.
Open the app, point your phone at the shooter, and the model scores every rep in real time. Add the LaxPod when you want sensor-grade mechanics on top.