Identity SolutionsOwn ProductHH
Handball Hub
The #1 handball app globally.
40K+Downloads4.8★App rating4LanguagesEHFFederation partner
Computer vision for cheerleading.
AI system that detects, classifies and scores cheerleading actions from competition footage, in production with USASF judges since 2022.
Timeline
2022, present
Team
5 people · ML + backend
Role
Co-development with internal science team
Status
Long-term partner
Long-term partnerCA
92%
Action accuracy
4 yr
Engagement
1.4M
Frames trained
16
Action classes
The brief
CARA started as a research collaboration and became core scoring infrastructure for one of the largest cheerleading federations in the US.
The challenge
Sport-specific actions, occluded performers, fast-moving cameras and a hard requirement for explainable outputs that judges could review on tablets at competitions.
Our solution
PyTorch model with a custom temporal attention head, deployed via Triton, with a review UI that surfaces frame-level evidence for every classification.
What we built
Six things that make CARA work.
01
Action detection
Temporal model classifies 16 distinct cheer actions from raw video.
02
Frame evidence
Every score links to the frames that triggered it.
03
Judge review
Tablet UI lets judges accept, override or annotate outputs.
04
Versioned models
Score records pin the exact model version used, for audit.
05
Active learning
Judge overrides feed straight back into the training queue.
06
On-prem option
Deployable on a federation laptop for off-grid venues.
Tech stack
The tools we picked, and why we kept them simple.
Boring tech where boring is right, sharp tech where it counts.
PyTorchTritonPythonFastAPIPostgresReact
Results
What changed after launch.
92%
Action classification accuracy
4 years
Continuous partnership
12K+
Routines analysed
3x
Judging throughput vs manual
Identity is the rare partner that can sit at the table with our science team as a peer. They build production infrastructure that our judges actually trust.
DE
Head of Innovation · CARA