Defect alert — Line 3, Station 7. Colour anomaly: upper-right quadrant of component EV-204, size 8 mm, chromatic deviation +15% from target. Classification: borderline — requires evaluation. Component in inspection buffer, line production continuing.
AGENT · QUALITY SENTINEL MANUFACTURING
Production line defects recognised and described with precision.
Quality Sentinel Manufacturing identifies scrap patterns and defects on the production line by combining computer vision — cameras already active on the line — and LLM verbalization of the detected patterns. The quality manager receives the alert with the structured description of the defect, its location, and estimated severity.
Quality Sentinel at work.
Anomaly confirmed. New material batch, probably a mixing problem. Scrap and notify procurement.
Component EV-204 marked as scrapped, batch L2605-A flagged. Notification to procurement manager with batch reference. Defect record written: type, location, decision, operator, timestamp.
Why it exists.
Quality control on a production line today is split between manual sample inspection and computer vision systems for structured patterns — presence or absence of components, dimensions, alignments. Less-structured defects such as fine cracks, colour anomalies, and minor deformations are hard to capture with rigid rules.
How it works on the line.
Quality Sentinel combines two levels. The computer vision system already active on the customer's line detects the visual patterns. The LLM verbalization converts those patterns into structured descriptions — "colour anomaly located in the upper-right quadrant, size 8 mm, deviation from target" — which the operator can inspect in natural language.
The decision stays with the qualified operator.
For defects clearly outside spec: automatic hold of the component on the line with tracking in the MES. For borderline defects: alert to the quality manager for evaluation. Computer vision is not a built-in capability: the agent integrates with the system already active on the customer's line via an adapter delivered during the project.
Who it serves and where it applies.
Quality manager
Receives structured alerts with a text description of the defect, its location, and estimated severity. Reads the anomaly description instead of interpreting only numerical coordinates from a SCADA system.
Line L1 · vision 18:42
Defect detected scratch · cluster 4
Operator alert at booth
Line stop proposed to supervisor
Event recorded · audit
Production manager
Sees the rate of defective components that slip through inspection and reach later stages of the line fall. Automatic hold for defects clearly outside spec interrupts the flow before the problem propagates.
Line L3 · SKU-441 7-day fcst
Week 23 8,400 units
Week 24 9,100 units
MES · MRP fed
Variance vs plan: +6%
System quality manager
Has an inspectable trace of every defect detected, every decision made, and every component scrapped. Defect trend analysis — by type, by station, by material batch — runs directly on the audit registry with a standard SQL client.
fnol.receive 09:14:22 ALLOW
triage.classify 09:14:25 ALLOW
idd.check 09:14:31 WARN
liquidation.propose 09:15:02 ALLOW
SELECT * FROM audit_log WHERE claim_id = '2024-0847'
A concrete example.
Station 7 detects an anomalous pattern.
For a manufacturing company producing automotive components, the production line has cameras on five critical stations. The customer's computer vision system is already active. During the afternoon shift, station 7 detects an anomalous pattern on a component EV-204 from a new batch. The vision system passes the frame to the agent.
The LLM produces the structured defect description.
The agent processes the image with LLM verbalization. It produces the structured description: colour anomaly in the upper-right quadrant, size 8 mm, chromatic deviation of 15%. It classifies the severity as borderline — not clearly outside spec, but requires human evaluation. It places the component in the inspection buffer, the line continues.
The alert reaches the manager on the work channel.
The alert reaches the quality manager on the work channel with the text description and component reference. The manager evaluates: the batch is new, probably a material mixture out of spec. Confirms the scrap and notifies procurement. The agent writes the defect record and updates the MES.
Configuration and technical resources.
The Quality Sentinel rules are declarative. The customer's quality team and production engineering define in a readable format the expected defect patterns, the tolerance thresholds for each anomaly type, the severity classification rules (automatic hold vs alert for evaluation), and the alert format. The rules live in the customer's repository, versioned, validated at agent startup.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, topic-guardrail
- Native channels
- Slack, Telegram, WhatsApp, OpenAI-compatible HTTP
- Prerequisite: computer vision system
- the customer has an active computer vision system on the line (proprietary or commercial); the agent integrates with the existing system
- Computer vision system integration
- dedicated adapter delivered during the project
- MES integration
- dedicated adapter delivered during the project
- Memory
- persistent per instance, pgvector + PostgreSQL FTS on historical defect patterns
- Registry
- immutable, queryable with a standard SQL client
Frequently asked questions about the agent.
The typical pattern for Quality Sentinel is 16-22 weeks. Discovery and analysis of the existing line 3 weeks, defect rule and tolerance threshold configuration 4-6 weeks, integration with the computer vision system and MES 6-8 weeks, line calibration and validation 3-4 weeks.
Yes. Quality Sentinel does not replace the vision system: it integrates with the one already in use by the customer. The prerequisite is an active computer vision system on the critical line stations — proprietary or commercial (Cognex, Keyence, industrial proprietary systems). The integration adapter is delivered during the project.
When a new component is introduced on the line, the customer's quality team configures the new defect patterns and tolerance thresholds in the repository's declarative rules. The agent applies the new rules immediately. The adaptation curve for a new component is measured in days of configuration, not weeks of retraining.
The runtime audit registry traces every defect detected with type, location, severity, decision, operator, and timestamp. The quality team queries the registry with a standard SQL client for trend analysis: frequency by anomaly type, distribution by station, correlation with specific material batches.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how Quality Sentinel would configure to the customer's line. Which computer vision system is in use, which stations are critical, which defect types are the priority.