AthleteOS
AthleteOS is a production streaming platform that fuses real-time computer vision pose data with wearable biometrics to deliver intelligent, context-aware exercise coaching. It combines three distinct data streams — camera-based pose estimation, wearable sensor data, and historical training sessions — into a single coherent intelligence layer.
Problem
Athletes get feedback from two completely disconnected systems: vision-based form analysis and biometric wearables. Neither talks to the other, so coaching is always incomplete. Form looks right but heart rate is tangled, or biometrics look fine but movement mechanics are off. The data existed — it just wasn't fused.
Solution
Built a dual-topic Kafka ingestion pipeline that pulls MediaPipe BlazePose pose data and WHOOP API biometrics in parallel, joins the streams in Apache Flink, writes to Snowflake via dbt-modeled dimensional tables, and surfaces a RAG-powered coaching interface that answers natural language queries over structured biometric history.
Architecture
Dual-topic Kafka producers — one for MediaPipe BlazePose pose keypoints from camera feed, one for WHOOP API biometric data (HRV, strain, recovery score).
Apache Flink joins the two Kafka topics on session ID and timestamp, computing real-time form scores and biometric correlation metrics.
Snowflake stores raw and transformed data. dbt models apply dimensional modeling best practices — fact tables for sessions, dimension tables for exercises and athletes.
Great Expectations validation runs at ingestion and transformation boundaries, catching schema drift and data freshness issues before they reach downstream consumers.
pgvector stores embeddings of historical training sessions. LLM APIs (Anthropic/OpenAI) answer natural language queries — 'why was my deadlift form worse on Tuesday?' — with full biometric context.
FastAPI exposes the coaching interface and session data. Grafana/Prometheus dashboards monitor pipeline health, ingestion lag, and system performance in real time.
Highlights
- Dual-topic Kafka ingestion fusing computer vision and wearable data in real time.
- Apache Flink stream joins on session ID and timestamp for sub-second latency.
- RAG-powered natural language coaching over structured biometric history via pgvector.
- dbt dimensional modeling with Great Expectations validation at every pipeline boundary.
- Grafana/Prometheus observability stack for production pipeline health monitoring.
- Built with Claude Code — rapid API prototyping and architecture iteration.