CircuitMindAI

Result and evidence

• Vision and voice PCB inspectionCapabilityCached for network loss; detection quality was not measured.

The case

Problem

An operator inspecting a board needs to see the copper fault and hear inspection guidance, including when the network drops. A detector that fails with connectivity loss is not dependable enough for the inspection floor.

Decision

Vision on the board and voice for the operator; results are cached locally so the floor can keep working when the network drops.

Constraint

The inspection has to survive a dropped network.

Example

Input: a board image and operator audio. Output: a fault overlay on the copper, with inspection guidance still audible when the network drops.

Considered and rejected

A server-only detector, rejected because it would fail when connectivity drops.

Architecture, on ECS Fargate

1. Next.jsUI
2. ExpressAPI

Beside the path

1. Bedrock Novavision and voice
2. OpenSearchindex

Around it

1. GitHub ActionsCI/CD to ECR

What was built

1. A multi-modal GenAI pipeline: Nova Pro handles vision-based fault detection, and Nova Sonic provides low-latency, two-way voice guidance for the technician.
2. A decoupled Next.js frontend with an Express REST API.
3. GitHub Actions builds and pushes Docker images to ECR and updates ECS Fargate on every push. An Application Load Balancer and Auto Scaling keep it available and cost-efficient.
4. A local cache that holds the inspection when the network drops.

Limitations

Detection quality (precision, latency, confusion) was never measured.

What I would change now

Measure precision, latency and confusion.