Computer vision for production monitoring, workplace safety, and goods control
SOAI is a vision-based monitoring platform built for manufacturing and warehouse operators in Vietnam. The plugin architecture has two abstraction layers: AnalysisEngine for analysis and DeviceProvider for image sources. Each customer problem is assembled from tested vision modules and deployed with matching hardware: IP cameras, edge processing boxes, or direct integration into brand-name cameras where the infrastructure allows it.
From the shop floor to the report
Goods control
Situation
Packing and warehouse in/out reconciliation is hard to get right by hand: orders get mixed up, manual counts drift, and complaints have no evidence trail to check against.
What the system does
Extracts QR/barcodes directly from packing video and timestamps each order (pull up the exact clip for any complaint); counts dense SKUs in images with an object-detection model and exports annotated images for manual audit when needed; reconciles counted quantities against warehouse in/out records.
Device requirements
Fixed IP cameras at the packing station (ONVIF/RTSP), or existing warehouse cameras.
Where data flows
Images are processed on the local network; results (counts, timestamps, annotated images) are stored on-prem or on the customer’s own server.
Workplace safety
Situation
PPE violations and entries into hazardous zones are usually caught only after an incident, because manual monitoring can’t cover every shift and every area.
What the system does
Detects missing protective equipment (helmets, vests) and presence in restricted zones; custom checks, described in plain language with a handful of sample images, are added for a specific factory’s requirements without retraining a model from scratch. Every detection is scored for confidence and queued for a human operator to approve; nothing acts automatically.
Device requirements
IP cameras at hazardous areas or restricted entry points; edge processing boxes when on-site inference is required.
Where data flows
Inference can run at the edge to cut bandwidth and keep imagery inside the customer’s facility; decision logs are kept for later audit.
Operations monitoring
Situation
Shift managers need to pull up a specific window of time, such as an incident or a shift, but have no fast way to search through raw, scattered video files.
What the system does
Live view across multiple cameras, manual or scheduled snapshots, per-device recording lists, and precise clip extraction for the exact time range needed, for incident investigation or shift reporting.
Device requirements
All cameras already registered in the device management system.
Where data flows
Streams and recordings are stored internally; access is account-scoped, with protection against reading files outside the allowed path.
SOAI POC: six modules verified
Each module below has been tested and verified in POC, with the matching technical stack.
| Module | Capability | Technical stack |
|---|---|---|
| Packing video analysis | Extracts QR codes from packing video and timestamps the start/end of each order, so you can pull up the clip behind any complaint in seconds. | OpenCV · pyzbar · Celery |
| SKU image counting | Counts dense objects in under 3 seconds, exports annotated images, and records the model version used, so counts can be audited later. | YOLOv8 · SAHI |
| Binary classification | New checks are defined with plain language and a few sample images, no retraining required. A new inspection rule ships in a day. | VLM · few-shot |
| Device management | Central registration and connection testing for cameras/devices, with credentials encrypted at rest, run many camera sites from one place. | FastAPI · Fernet |
| Live view & snapshot | Live view across multiple cameras, manual or scheduled snapshots, for monitoring during a shift as well as after-the-fact review. | MJPEG · Celery Beat |
| Video recording & export | Per-device recording lists and precise clip extraction for a requested time range, for incident investigation and evidence. | FFmpeg |
Screenshots straight from the SOAI system
Watch the full workflow
SOAI system walkthrough video, using demonstration data.
From goods monitoring to brand-name cameras
Phase 1: Build the goods monitoring system
On-site monitoring and online packing analysis: the six modules above, verified in POC.
Phase 2: ACAP embedding into camera hardware
Embedding detection models directly into Axis and Hikvision cameras through ACAP is phase 2 of the SOAI roadmap. Not yet implemented; partner program requirements are still being confirmed with each vendor.
Phase 3: Extend monitoring scope
Extend to workplace safety, goods flow, and inventory/quantity control across the whole enterprise. Not yet implemented.
AI föreslår - operatören beslutar
Systemet upptäcker, bedömer sin egen tillförlitlighet och föreslår en åtgärd. Inget steg utförs på egen hand utan mänskligt godkännande.
Upptäckt: arbetare utan skyddshjälm · Linje 2
Kamera vid förpackningsområdet · bildruta 14:02:37
Granskningsåtgärd
Upptäckt: arbetare utan skyddshjälm · Linje 2
- Räknad SKU-mängd avviker från utleveranssedeln87%
- Person upptäckt i ett begränsat område utan skyddsutrustning78%
Talk technical about SOAI
Send a specific problem from your factory or warehouse. The SOAI engineering team replies with a deployment proposal, including modules and stack.
Talk technical about SOAI