From pixel differences to defect semantics
InspirAI is reshaping PCBA quality inspection. Our Mixture-of-Experts AI agent re-inspects your AOI output — not just seeing differences, but truly understanding defects. The result: dramatically fewer false calls, near-zero escapes, and lights-out quality on the lines you already run.
Goodbye to “only seeing.” Welcome to “truly understanding.”
The Problem
Traditional AOI has three pain points
Automated Optical Inspection catches differences, but it can't reason about them. Fixed rules and thresholds can't adapt to dynamic changes in incoming materials, lighting, and imaging — so quality teams pay for it three ways.
Programming
- Time-consuming setup — algorithms and thresholds rely on human experience.
- The “seesaw effect” — an endless trade-off between overkill and missed defects.
Labor Cost
- Costly operator re-checks — high false-positive rates force constant manual re-inspection.
- Rising labor costs from continuous recruiting and training.
Quality Risk
- Inconsistent re-checks that vary by operator experience and standard.
- Escaped defects that reach downstream processes and customers.
The Fix
An AI re-inspection layer that thinks before a human does
InspirAI sits between your AOI machines and your operators. Every AOI “NG” is re-judged by our AI agent first — passing the genuine OKs and forwarding only real suspects to a single, centralized review.
Today every AOI “NG” is hand-checked at each line (≈99% false positives). InspirAI re-judges them first — passing true OKs and forwarding only real suspects to one centralized review.
Every AOI station flags ~100 “NG” per batch. Operators re-inspect all of them by hand at each line — and roughly 99% are false positives.
The AI agent absorbs every AOI “NG”, auto-passes the true OKs, and sends only the ~10% real suspects to one centralized OP review — with no impact on cycle time.
The Product
InspirAI PCBA AI Agent
A perception-and-decision agent for the shop floor — compatible with mainstream AOI, installed without touching your line.
Value proposition
Operator false-negatives run 1,000–5,000 DPPM. InspirAI holds AI false-negatives below 100 DPPM.
>80% labor-cost reduction by centralizing operator re-inspection at a single station.
Form factor
- Mainstream AOI compatibilityWorks with the AOI vendors you already run.
- Non-invasive setupSimply plug in the network cable — no line modifications.
- Shop-floor applianceCentral inspection station, AI appliance, and real-time monitor over your shop-floor network.
Key metrics
- Cycle time
- No impact
- FN rate
- <100 DPPM
- Critical defects
- 0 FN
- FP reduction
- 90%+
- 3D AOI
- Supported
- Go-live
- 30 days
Shop-floor AI appliance
A single rack appliance serves many AOI stations over the network — X86/ARM/GPU/NPU compute, scaling from one machine to a multi-factory cluster.
3D AOI ready
Handles 3D solder-joint inspection — lifting, coplanarity, and volume — alongside 2D.
The Leap
From binary classification to an agentic “AI brain”
A single OK/NG classifier is a good start — but it's product-specific, caps out near 70% overkill reduction, and drifts as samples fall into the gray zone. An agent locates, classifies, and reasons about defects — and keeps learning.
| Dimension | Binary classifier | InspirAI Agent |
|---|---|---|
| Defect detection | NG vs OK only | >30 types of NG classes |
| FP (overkill) reduction | < 70% | > 90% |
| Model iteration | Manual iteration | Continuous self-iteration |
| Production adaptation | Fixed confidence threshold — can't adapt | Self-adapting, continuously optimized |
| Missed NG (FN) | 300–500 DPPM | < 100 DPPM |
How It Works
A Mixture-of-Experts agent, not a single classifier
CNN + Transformer fusion within an MoE framework — combining supervised labeling, self-supervised feature extraction, and unsupervised anomaly detection. Inspection flows through specialized experts, then feeds results back into the model so it sharpens every shift.
Adaptive edge platform
Quantifies operator OK/NG criteria and dynamically updates defect thresholds in-model. On a new product changeover, it tunes from “strict” to “optimized” within 1–2 hours.
Synthetic data generation
Diffusion-based generative AI with domain-specific LoRA fine-tuning synthesizes controllable PCBA defects — feeding the models abundant, diverse training data for higher precision and robustness.
Expert-free MLOps
A guided data → label → train → validate → deploy workflow that frontline operators run themselves — continuously improving models with zero data-science overhead.
Why InspirAI
Built to drop onto the lines you already run
Works with your existing AOI — from a pure-software install on one machine all the way to a multi-factory cluster.
Superior performance
The MoE agent integrates positioning, classification, and detection — driving 90%+ overkill reduction and sub-100 DPPM escapes, where binary classifiers stall at ~70%.
Unified AI solution
An MLOps platform with a shared model repo and X86/ARM/GPU/NPU compute. Data connectors interface to any AOI vendor — plug-and-play over a single network cable.
Expert-free maintenance
Zero-threshold management and O&M. A guided data → label → train → validate → deploy workflow is run by frontline operators — no data scientists required.
Non-invasive integration
No physical modifications to your line. Install the AI Data Connector, plug in the network cable, and the agent re-inspects in parallel — with zero cycle-time impact.
Flexible configurations
One agent, four deployment footprints — scaled to your cost, migration, and availability needs.
SW Integration
Software only · one machine
- Cost
- Low
- Migration
- Easy
- Availability
- None
Single AOI Edge Box
Edge box · per AOI
- Cost
- Low
- Migration
- Easy
- Availability
- None
Multi-AOI Appliance
Appliance · many AOIs
- Cost
- Lower
- Migration
- Moderate
- Availability
- Medium
Multi-Factory Cluster
Cluster · many sites
- Cost
- Lowest
- Migration
- N/A
- Availability
- High
Proven Results
Deployed across customers, boards, and AOI equipment
Real production lines at leading electronics manufacturers — 2D and 3D, rigid and flex, from a handful of components to thousands.
Foxconn (Shenzhen)
Case #11–4 up medium-to-large boards, 200–600 mm, 1,000–6,000 components, 1–3 line changeovers per day.
Foxconn (Shenzhen)
Case #28–12 up high-density panels, single piece <100 mm, 80–300 components, 2-year product lifecycle.
S08 line (rails 1 & 2) reached lights-out manufacturing.
TPV (Xiamen)
Case #38–20 up, predominantly long-and-narrow boards, 20–50 components, changeover every 1–2 days.
Amphenol
Case #4100–200 up, low density (~20 components/piece). Cut downstream escapes from 500–1,000 DPPM.
Customer board images and identifying details are confidential and omitted. Figures reflect reported deployment results.
Put an AI agent on your AOI lines
Tell us about your boards and AOI setup. We'll scope a non-invasive pilot, prove the false-call reduction on your own production data, and have you live in 30 days.