Neuron Vision is a multi-agent system designed for real electronics manufacturing quality control.

Architecture

  • Triage Agent analyzes the PCB image and identifies high-risk zones such as BGA packages, fine-pitch ICs, and connectors.
  • Parallel Inspection Stage runs three specialized agents concurrently:
    • Solder Inspector: bridges, cold joints, insufficient solder, excess solder
    • Component Inspector: missing parts, misalignment, orientation, shifted components
    • Marking Inspector: silkscreen, QR codes, polarity marks, traceability labels
  • Chief Inspector synthesizes structured reports from all inspectors and issues a final verdict using only validated Pydantic v2 models.

Key Technical Strengths

  • True parallel specialist execution with measurable latency improvement over sequential inspection.
  • Strict structured outputs via Pydantic v2. Chief Inspector never makes a decision from raw free-form text.
  • Full production observability with Arize Phoenix/OpenTelemetry, including parent inspection spans and per-agent waterfall spans.
  • Deployed on Google Cloud Run and powered by Vertex AI Gemini 2.5 Pro.
  • Demo mode included for reliable judging without requiring live GCP credentials.

The project addresses a real manufacturing pain point: PCB defects are expensive, AOI machines are costly, and manual inspection is inconsistent. Neuron Vision demonstrates how specialized multi-agent systems with strong observability can deliver reliable, auditable decisions in physical-world applications.

Built With

  • ai
  • arize
  • asyncio
  • cloud
  • gemini
  • google
  • openinference
  • opentelemetry
  • phoenix
  • pydantic
  • python
  • run
  • streamlit
  • vertex
Share this project:

Updates