Architecture

INFERENCE BUILT FOR PHONE FOOTAGE

Raw practice film in, biomechanical telemetry out. The stack is engineered so a coach never does more than a two-tap upload, while GPU-accelerated inference returns branded reports inside the SLA window.

Ingestion & Pipeline

Video Storage

Direct S3 / Supabase Storage uploads via signed URLs — no coach-side tooling required.

Job Queue

Celery/Redis or AWS SQS handling asynchronous video processing at academy scale.

Pose Engine

Python microservice on MediaPipe / YOLOv8-pose tracking release time, shooting arc, elbow extension angle and lateral plant speed.

Backend & DB

Fastify (Node.js) or FastAPI paired with Supabase — PostgreSQL plus multi-tenant auth.

White-Label Payments

Stripe Connect (Express/Custom) for automated split fees: 1% platform fee plus SaaS billing.

NVIDIA Acceleration Stack

NVIDIA TensorRT

Converts PyTorch/ONNX pose models into high-performance inference engines: 3x–5x latency reduction, 60-second clips processed in under 2 seconds.

NVIDIA DeepStream SDK

Multi-stream pipeline handling hardware video decode, NvDCF tracking and coordinate rendering.

Phase 1 — Cloud Serverless

NVIDIA L4 / A10G GPUs on RunPod or AWS. Zero hardware on-site; pure smartphone/tablet upload at $0.50–$0.90/hr.

Phase 2 — Edge Smart Gym Hub

NVIDIA Jetson Orin Nano / AGX Orin leased to flagship academies at $499–$999/mo for real-time court feedback.

Estimated Infrastructure Cost

MVP / Pilot (0–10 academies)$150–$325 / mo
Scaled operations (50+ academies)$800–$1,900 / mo
Initial tooling / setup$1,500–$3,500 one-time