Ascension is a mobile application that analyzes your climbing videos using computer vision and AI to give you concrete, objective feedback on your technique — the kind of feedback that usually requires a professional coach standing right next to you.

Film yourself on the wall. Upload the video. Get a biomechanical breakdown of your movement, joint-by-joint coaching advice, and an optional ghost overlay showing the optimal path for your morphology.

Professional climbing coaching is expensive, subjective, and hard to access. Most climbers train without real-time feedback and only correct bad habits when they plateau or get injured.

Ascension bridges that gap by making objective, AI-driven analysis available to any climber — from beginner to competition-level — at a fraction of the cost.

Video analysis Extracts a 33-point skeleton from each frame of your climbing video using pose estimation. Computes joint angles, center of gravity, and weight distribution throughout the move.

Hold detection Photograph the route before you climb. The AI detects and classifies holds (crimp, sloper, jug, pinch, pocket…). You can manually correct misclassifications before the analysis runs.

Coaching advice Combines skeleton data with hold positions to generate targeted, move-by-move feedback. Concrete, not generic — e.g. "hip too far from the wall on move 3" or "match hands before flagging right."

Ghost Climber (Premium) Generates the optimal movement path for your morphology using inverse kinematics, then renders it as a frame-by-frame overlay on your video. Compare your actual movement to the ideal line.

Personalized training programs Custom routines generated from your profile (level, goals, injury history) and updated based on your analysis history.

Ascension is built as a monorepo with three independently deployable services:

apps/
├── mobile/    Flutter app (iOS & Android)
├── server/    Rust / Axum REST API + WebSocket
└── ai/        Python AI workers (MediaPipe, PyTorch, OpenCV)

Data flow

User films → presigned URL → direct upload to object storage
→ API notifies queue → AI worker processes → JSON results stored
→ WebSocket notification → client renders overlay

Heavy AI workloads (pose estimation, hold detection, ghost path generation) are queued via RabbitMQ and processed asynchronously. The mobile client renders results from a lightweight JSON payload (~50 KB) rather than a re-encoded video — keeping latency low and egress costs minimal.

Performance targets

• API response time: p95 < 200 ms
• Analysis processing: < 30 s for a 30-second video
• Time to result (upload + processing): ~45 s

Active development. The architecture, API specification, database schema, and infrastructure guides are finalized. Core services are being implemented.