Plain-language prompts
Describe what to look for in a sentence. Change the question at any time without retraining a model.
PAI-VISION
Private Vision Intelligence
Your cameras, understood in plain language. On-site.
Connect a webcam or an existing IP camera, describe what you want to know in plain language, and a vision-language model running on your appliance analyses the live feed in real time. Spot idle time on a production line, unsafe behaviour or activity after hours, without training a custom model and without sending a single frame to the cloud.
- Live analysis of webcams and IP cameras
- Plain-language prompts, no model training
- Video never leaves your site
01The challenge
Cameras record everything. Almost nobody watches.
Factories, warehouses and stores already have cameras, but footage is only reviewed after something goes wrong. Traditional computer vision needs a custom model for every question. A vision-language model understands scenes described in everyday language, so one system can answer many operational questions, live.
02Capabilities
What it does.
Live camera streams
Browser webcams over WebRTC and IP cameras over RTSP feed the model continuously.
Runs on your appliance
Open vision-language models are served locally, so video and results stay on-site.
Real-time responses
Observations stream back within seconds, with latency and GPU utilisation shown alongside.
Structured events
Findings can be logged, summarised by shift and pushed to dashboards or alerts.
Human review
Observations are indicators for supervisors to review, not automated decisions about people.
03How it works
From question to answer, inside your walls.
NVIDIA Live VLM WebUI, an open-source real-time interface for vision-language models, with open models served on the appliance. Logic Sonata adds camera integration, prompt design, event logging, dashboards and governance.
- 01CamerasWebcam or IP camera (RTSP)
- 02Frame samplingContinuous, on-site
- 03Vision-language modelRuns on the appliance
- 04ObservationsStructured events
- 05SupervisorsDashboards and alerts
- 01
Connect cameras
Add webcams or existing IP cameras. No new camera hardware is needed for most sites.
- 02
Describe what matters
Write prompts such as "Is the operator at station 4 working, waiting or away?"
- 03
Analyse live on-site
The appliance samples frames continuously and a vision-language model interprets each scene.
- 04
Act on the findings
Supervisors see live observations and shift summaries, and receive alerts for agreed events.
- 05
Refine and extend
Adjust prompts, add cameras and new questions as your team learns what is useful.
04Time and motion, in plain language
Ask the line what is happening.
Instead of training a model per question, supervisors describe what they care about. The model answers continuously, and results roll up into shift summaries.
Prompt
Is the operator at station 4 sewing, waiting for work or away from the station?
Prompt
Are bundles piling up between stations 3 and 4? Describe the queue.
Prompt
Is anyone in the marked forklift lane without a high-visibility vest?
Prompt
After 22:00, report any person or vehicle near the loading bay door.
05Use cases
Where it pays off first.
Time and motion on the line
Detect idle stations, waiting for materials and unnecessary movement to balance lines and remove bottlenecks.
Quality spot checks
A second pair of eyes for visible defects such as stains, broken stitches or shade variation.
Safety and PPE
Flag missing helmets or vests, blocked exits and people too close to moving equipment.
Security monitoring
Activity after hours, loitering near restricted doors and unattended items, described in plain language.
Warehouse and logistics
Dock occupancy, loading progress and unsafe pallet stacking, shift by shift.
Retail operations
Queue length, empty shelves and fitting-room waits, so staff go where they are needed.
06Security and governance
Controls built in, not bolted on.
Every Logic Sonata system runs on hardware you control, with access, audit and data boundaries designed in from the start.
- Video and results processed and stored on-site only
- Configurable retention, with no footage kept by default
- Focus on processes and safety, not individual scoring
- Access to live views and history limited by role
- Deployment guidance aligned with local data protection laws
- Signage and worker consultation recommended before go-live
07FAQ
Questions about PAI-VISION.
Do we need new cameras?
Usually not. Webcams and most IP cameras that provide an RTSP stream can be connected. We check your cameras during the assessment.
Is this employee surveillance?
The system is designed to improve processes, safety and security. We recommend focusing prompts on stations and events rather than individuals, informing staff, and following local data protection law. Findings are for human review.
How accurate is it?
Vision-language models are strong at describing scenes but can make mistakes, so observations are treated as indicators. We tune prompts and camera placement during the pilot and measure accuracy against real footage.
08More from Logic Sonata
Other private AI systems.
Private Knowledge Assistant
A governed assistant that answers questions from your SOPs, specifications, policies and reports. Every answer cites its source and every user sees only what their role allows.
Private Coding Assistant
Code completion, repo-aware chat, review and agentic coding in the tools your developers already use, powered by open coding models on your own hardware.
Private AI Agent Platform
Self-hosted AI agents built on Hermes Agent or OpenClaw that research, write and act across your tools, running on private models inside a sandbox with approval for anything that matters.
Private Image Generation Studio
A ready-to-use image generation studio built on ComfyUI, preinstalled with curated models and templates, so designers create prints, concepts and campaign visuals without their work leaving the building.
Get started
See PAI-VISION on your own data.
Book a 30-minute consultation. We will show the system working and scope a pilot around your use case.