Any camera. Any factory. Defects detected, dispositioned by LLM, and acted on—physically—before the part leaves the line. No cloud required.
Most vision systems stop at detection. VIP closes the loop—LLM-based disposition triggers a physical action before the part moves on.
RTSP, USB, or GStreamer — VIP ingests any camera. ONNX models run inference at the edge, entirely offline. Defects are captured with image evidence and confidence scores.
Your cameras · no lock-inSpecialized LLM agents evaluate each defect in context—severity, model version, line history—and issue an automatic disposition: confirm, reject, or escalate for human review.
Sub-second · explainableDisposition triggers a physical action—robot sort, reject bin, line stop, or operator alert—at the point of production. No human needed. No cloud round-trip. No wait.
Physical action · no alerts-onlyThree interlocking capabilities that work even when your network doesn't.
Inference, disposition, and physical action happen on your hardware. SQLite outbox buffers 24+ hours offline. No cloud dependency for production decisions.
Register once, deploy everywhere. Manage dozens of edge stations as a unified fleet — model versions, health telemetry, and live/offline status in a single pane.
Specialized agents handle disposition, drift detection, active learning, and NL querying. An operator chat lets your team ask "why are defects spiking on Line 3?" in plain English.
Real-time visibility from camera to disposition—not just charts, but actionable state for every defect on every line.
Real-time quality inspection · Shift 2 · Line A & B
scratch (28) · dent (11) · contam. (8)
2,847 parts · 47 defective
AI confidence < 85% threshold
2 stations offline · syncing
Offline buffer healthy · C-01 queued 847 events · ~6h buffer remaining
syncingActual product UI · disposition states and offline buffer are live product features, not mockups
Click back from any disposition to the exact image frame, model version, confidence score, and operator action. Built for regulated industries and quality audits.
Scratch mark detected at 09:42:17 · Station A-12 · Frame #84291
conf: 0.97 · model: defect-detector-v3.2 · bbox: [142,88,210,156]
AI agent: "High confidence scratch, non-cosmetic area — reject." Triggered sort signal.
agent: auto_disposition · latency: 210ms · decision: REJECT
Image, metadata, and disposition log written to SQLite outbox. Synced to cloud at 09:43:01.
record_id: d4f8a2b1 · ↗ View evidence
Full inference and disposition with zero internet. On-prem, OT network, or air-gapped facility.
Images and metadata stay on your hardware. Cloud sync is opt-in. Your data, your control.
Every detection, disposition, and action logged with model version, timestamp, and operator ID.
ONNX models. Linux / Docker. No proprietary camera or compute hardware required.
See VIP running on your cameras, your parts, your line—in under 30 minutes.