ML Systems / Agentic AI · Sports · production
Uplift
AI motion capture and coaching from phone video.
Production sports performance system that turns iPhone/iPad video into full 3D kinematics, movement Signals, and an AI Coach that translates biomechanics into actionable training recommendations.
Problem
Marker-based motion labs are expensive and inaccessible. Coaches need biomechanical ins…
System
A CV → biomechanics → retrieval-backed coaching pipeline: mobile capture, 3D reconstruc…
Role
AI Architecture & ML Systems
Status
production
Outcome
1–2 iOS cameras
My role
AI Architecture & ML Systems
- AI Architecture
- ML Systems
- Retrieval
- Evaluation
- Productionization
Period / 2024–Present
Phone-to-3D movement intelligence
Plays once · tap to replay
Problem
Marker-based motion labs are expensive and inaccessible. Coaches need biomechanical insight and personalized guidance without $50k capture rooms.
System
A CV → biomechanics → retrieval-backed coaching pipeline: mobile capture, 3D reconstruction, metric extraction, risk/performance Signals, and an AI Coach grounded in validated movement science.
Architecture
Mobile video uploads into a CV reconstruction service that emits 3D kinematics and metrics; Signals evaluate movement health; an AI Coach retrieves training knowledge to produce recommendations on web/mobile dashboards.
- iOS Capture → Cloud Ingest
- Cloud Ingest → CV Reconstruction
- CV Reconstruction → Biomechanics Metrics
- Biomechanics Metrics → Signals
- Signals → AI Coach
- Training Knowledge RAG → AI Coach
- AI Coach → Dashboard / App
- Signals → Dashboard / App
Execution path plays once on view · hover a node or tap the canvas to replay
System anatomy
Architecture
Capture → reconstruct → measure → coach.
- Mobile capture
- Cloud CV
- Signals layer
- Coach + dashboard
Engineering decisions
Decision
Phone-first capture over lab hardware
Constraint
Teams need assessments in gyms and fields, not only motion labs.
Approach
1–2 iOS devices feeding a cloud CV stack democratizes kinematics without wearables or multi-camera rooms.
Result
Production capture usable by pro, college, and youth organizations.
Decision
Metrics before language
Constraint
Generative coaching without biomechanical grounding hallucinates unsafe advice.
Approach
CV produces kinematics and Signals first; the AI Coach retrieves validated movement science before recommending drills.
Result
Coach outputs tied to measured movement, with guardrails against hallucinated prescriptions.
Decision
Signals as the coach interface
Constraint
Raw joint trajectories overwhelm athletes and staff.
Approach
An intermediate Signals layer translates 3D data into strengths, focus areas, and risk flags humans can act on.
Result
Actionable reports plus full 3D visualizer for deeper review.
Decision
Dual product paths, shared pipeline
Constraint
Pro capture (multi-cam) and consumer single-cam have different latency budgets.
Approach
Shared reconstruction and coaching services with mode-specific capture frontends keep one AI core.
Result
Capture for orgs and faster single-cam consumer feedback from one platform.
Reliability
- Guardrailed coaching recommendationsACTIVE
- Validated biomechanics metricsACTIVE
- Cloud analysis pipelineACTIVE
- Org-level athlete data controlsACTIVE
Evaluation
Outcome / Results
- Markerless 3D movement analysis from consumer devices instead of lab rigs.
- Signals that surface strengths, focus areas, and injury-risk patterns for coaches.
- AI Coach that turns complex biomechanics into personalized, guardrailed recommendations.
Stack
Product visuals
