Declared runtime, framework, persistence, media, observability, testing, and delivery technologies
Guided views
Explore this system
Step through curated paths without changing the source diagram.
Beat
Next
ReadyChapter 01 / 01
Guided chapter
Diagram guideExplore this system
Inspecting compiled semantics
E ExportT ThemeS Style0 Reset+ Zoom in- Zoom outEsc Close
Find a node
⌕/
No matching nodes
Semantic passport
Verified source
Authored reach
Route probeChoose a start node
Pick two semantic nodes on the diagram
Choose the source, then the destination. Direction matters.
Semantic lensCompare system roles
Choose up to two semantic kinds. One reveals its real traffic; two compare only direct authored relationships.
Choose a kind to inspect its nodes and touching relationships.
Semantic radar
Building overview
Click nodeDrag to pan
Application stack
• React and MUI provide the staff interface; getDisplayMedia and MediaRecorder create local movement lab clips with explicit browser consent; MotionBERT-Lite, ONNX Runtime Web WASM and iterative refinement run in a cancellable browser worker
• NestJS hosts the modular monolith and realtime gateway
• Prisma maps explicit PostgreSQL business schemas
Media stack
• C++ and Zoom Meeting SDK own capture
• Node, FFmpeg, and ffprobe own Zoom participant finalization plus bounded allowlisted AddPipe normalization
• AddPipe output is H.264 320x180/15 fps yuv420p with AAC 128 kbit/s and faststart
• Private S3 stores media outside Redis and the API; authenticated playback resolves stable video ids only
Engineering stack
• OpenTelemetry standardizes traces and metrics
• dependency-cruiser protects module boundaries
• Archify artifacts and repository docs publish through CI to Outline
• Legacy YouTube import retains pinned yt-dlp 2026.6.9, Python 3, Node 22, bgutil 2.0.0, FFmpeg, ephemeral Redis jobs and private S3 cleanup. The movement lab UI uses browser tab capture and no longer invokes these import routes.
Optional cloud motion capture
• tools/motion-capture-modal deploys a private CPU HTTP gateway and an on-demand NVIDIA L4 worker using Python 3.12, PyTorch 2.11, SAM 3D Body DINOv3 and MediaPipe for person bounding boxes.
• Delivery persists MotionCaptureJob in PostgreSQL and signs direct private S3 uploads/results, including a separate versioned mesh ZIP preserving topology, per-frame vertices and camera metadata. New jobs require both skeleton and mesh artifacts before readiness; legacy jobs remain readable without a mesh. Modal proxy credentials remain backend-only. The browser adapts MHR70 skeleton points to the existing Three.js bones/muscles rig; browser MediaPipe/MotionBERT remains selectable.
Architecture diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • [/] views • P play story • T theme • E export