Superior AI Solutions ← Back to Website

Edge-based AI computer vision dashboard for inventory tracking and quality control inspection running on local hardware for small manufacturers

EdgeVision — Local AI Inspection
All processing on-device · No cloud · Real-time detection
4 cameras live 2 defects detected
Parts inspected today
1,847
Avg 12/min
Defect rate
2.3%
vs 4.1% last week
Items auto-counted
342
3 bins monitored
Alerts fired
2
Since 7:00 AM
Model accuracy
97.4%
Demo simulation only
Camera feeds — click to inspect
CAM-01 · QC Line A
28 fps
QC inspection line
Defect detection · active
2 defects
CAM-02 · Bin A storage
15 fps
Parts bin — shelf A
Inventory count · active
Normal
CAM-03 · Bin B storage
15 fps
Parts bin — shelf B
Inventory count · active
Low stock
CAM-04 · Receiving dock
20 fps
Receiving dock
Barcode + count · active
Idle
Live detection feed
On-device inference · ~18ms latency
09:44:12CAM-01 — DEFECT: surface crack — part #A-0841 — conf 96.2% — rejected
09:44:09CAM-02 — inventory count updated: bolt_m8 → 147 units (was 153)
09:44:05CAM-01 — part #A-0840 — PASS — no defects — conf 99.1%
09:43:58CAM-03 — LOW STOCK: washer_m10 at 18 units (par: 50) — reorder triggered
09:43:51CAM-01 — DEFECT: edge chip — part #A-0838 — conf 91.4% — rejected
09:43:44CAM-04 — delivery scan: PO-2241 received — 48 units logged to inventory
Live inventory — auto-counted by vision AI
Reorder queue
3 items
washer_m1018 left → order 200
hex_bolt_5/1631 left → order 150
rubber_seal_42mm22 left → order 100
Count accuracy
Bolt M8
99%
Washer M10
97%
Hex bolt 5/16
96%
Rubber seal 42mm
94%
PVC fitting 1in
98%
Defects today
43
of 1,847 inspected
Surface cracks
18
Most common
Edge chips
14
2nd most common
Auto-rejected
43
0 false positives flagged
Recent defect detections
Surface crack — part #A-0841 Rejected
Linear crack 4.2mm along primary face. Exceeds 3mm rejection threshold.
CAM-01 · QC line A09:44:12Conf: 96.2%
Edge chip — part #A-0838 Rejected
Chipped edge, north face. 2.1mm depth. Material loss outside tolerance.
CAM-01 · QC line A09:43:51Conf: 91.4%
Discoloration — part #A-0821 Review
Surface oxidation, 8mm patch. Below rejection threshold — flagged for manual review.
CAM-01 · QC line A09:31:04Conf: 74.8%
Active
YOLOv8n — defect
Detects cracks, chips, discoloration, warping on machined parts
mAP@5097.4%
Inference18ms
Classes6 defect types
Trained on1,240 images
Active
YOLOv8n — inventory
Counts and classifies parts in bins — bolts, washers, fittings, seals
mAP@5094.1%
Inference22ms
Classes12 part types
Trained on800 images
Ready to deploy
EfficientDet — lumber
Knot, check, and wane detection for sawmill / lumber grading use case
mAP@5091.0%
Inference31ms
Classes5 grade types
Trained on620 images
Training pipeline — on local hardware
1
Capture training images
200–500 photos per defect class · phone camera or USB webcam · real shop conditions
2
Label with Roboflow or Label Studio
Free browser tool · draw bounding boxes · export in YOLO format · 2–4 hours for 500 images
3
Train YOLOv8 nano (local or free Colab GPU)
yolo train model=yolov8n.pt data=defects.yaml epochs=100 · ~45 min on free GPU
4
Export to ONNX or TFLite · deploy to edge
Single .onnx file runs on Raspberry Pi, Jetson Nano, or Intel NUC · no GPU needed at inference
Cameras
QC line (fixed)Arducam USB · $35
Bin countingWide-angle USB · $25
High-res inspection12MP USB · $65
LightingRing LED · $18 ea
Consistent lighting is more important than camera resolution. Always add diffused LEDs.
Recommended compute
NVIDIA Jetson Orin Nano
Price~$250
AI performance40 TOPS
Cameras supported4–6 USB/CSI
Inference speed<20ms / frame
Runs all 4 cameras simultaneously. Fan-cooled, industrial temp range.
Budget alternative
Raspberry Pi 5 + Hailo-8L
Price~$130 combined
AI performance13 TOPS
Cameras1–2 simultaneously
Best for1–2 camera installs
Hailo-8L HAT plugs directly onto Pi 5. Good entry point for a single QC station.
Full software stack — all open-source
Ultralytics YOLOv8
Object detection + classification. Train on your parts in hours. Export to ONNX for edge deployment.
Free
OpenCV
Camera capture, image preprocessing, contour detection for precise measurements.
Free
Label Studio
Self-hosted image labeling tool. Runs in browser on local machine. No data leaves the shop.
Free
SQLite + FastAPI
Local database stores detection events, inventory counts, and defect images. REST API for dashboard.
Free
Grafana
Local dashboard for defect trends, throughput, and inventory levels. Visible on any shop floor tablet.
Free
Twilio / local buzzer
SMS alert on critical defect spike or stock-out. Falls back to wired alarm if no cell signal.
~$5/mo