WJETT

Turn padel match video into evidence you can review and use.

Comment

WJETT is the padel analysis product I wanted for my own matches and for the friends I play with. You upload match video, WJETT processes every decoded frame, and the product turns the result into player and ball overlays, reviewable rallies, score correction, match analytics, and a history you can use to see what is changing.

The product is built as more than a visual prototype. The web app is React and TypeScript; the backend is Kotlin and Spring; PostgreSQL owns product state; and the video path uses NVIDIA DeepStream 9 with TensorRT engines. The runtime fails closed when its required model artifacts or decoded evidence are absent, rather than silently substituting synthetic inference.

WJETT existed before Build Week, but it was meaningfully extended during the submission period. Starting from commit e4690e44, the Build Week work includes a complete public and authenticated web redesign, expanded mobile surfaces, backend and storage hardening, runtime modularization, and a cross-stack static quality foundation.

I used Codex as the development environment and GPT-5.6 as the main engineering and orchestration model. GPT-5.6 Sol managed durable goals, split the work into bounded implementation lanes, integrated changes, and checked rendered and runtime evidence. Luna agents handled implementation-heavy and mechanical work; Terra agents acted as focused reviewers. Earlier detection work started with GPT-5.5 and was then improved and hardened with GPT-5.6. I remained responsible for the product direction, model contract, visual identity, quality gates, and every final trade-off.

WJETT is still a private pilot, not a claim of production-perfect sports analytics. The public URL explains the product; the video demonstrates the real local product stack and its current evidence honestly.

Testing instructions:

  1. Open https://wjett.42n.com.ar/ for the public product overview.
  2. Clone the private repository and run make qa-auth for the automated isolated authenticated smoke.
  3. Run make qa-auth-dev, then open http://127.0.0.1:18121/dashboard for an interactive test-auth workspace. This never weakens the normal Clerk stack.
  4. Full DeepStream analysis needs an NVIDIA GPU, DeepStream 9, and the private DS9 model artifacts documented in the root README. Those artifacts and private match footage are intentionally not committed.

Built With

Share this project:

Updates