SCENE 01 — RUNTIME ARCHITECTURE

Synapse-Graph — Architecture

How the AI interpretability pipeline flows from discovery to governance

🎨
Frontend
Next.js + React dashboard for synapse visualization
⚡
Neural Proxy
FastAPI orchestrating generation, tracing, OpenMetadata sync
🔍
Shadow Tracer
PyTorch hook-based tracer emitting AttentionTrace objects
🛡️
OpenMetadata
Governance plane storing topology, tags, and lineage
Component Flow Diagram
Frontend Neural Proxy Tracer OpenMetadata Dashboard Next.js + React FastAPI Runtime HeadMaskStore Shadow Tracer Gen Engine Topology & Tags API write tags

Key Flow: Discovery → Trace → Quarantine

  • Dashboard triggers discovery → Neural Proxy ranks attention heads by activation correlation
  • Shadow Tracer captures attention patterns → generates differential trace snapshots
  • User tags defective heads in OpenMetadata → HeadMaskStore updates → subsequent generations masked
  • This closes the loop: interpretability → governance → verifiable behavior change
← Intro Tech Stack → Status OpenMetadata
Architecture — Full View
Frontend Neural Proxy Tracer OpenMetadata Dashboard Next.js + React FastAPI Runtime HeadMaskStore Shadow Tracer Gen Engine Topology & Tags API write tags