AegisVision AI v2.0 — 3-Tier CV, ByteTrack OEE & Document Intelligence Platform

Production Line / Retail Camera Feed

Inference Execution Provider

Conf Floor 0.25

Ambient Lux Shift 0%

SAM 2

Line Gate Verdict

REJECT_CRITICAL_HOLD

IPC-A-610H Class 3

Est. Scrap / OOS Savings

$4,200 / hr

CAPA SLA: 15 min

Pipeline Throughput & Latency

34.2 FPS (29.2 ms)

1 SAM-2 Zero-Shot Trigger

Illumination Drift (PSI)

0.0000 (STABLE)

CLAHE Gain: 1.84x

Optical Inspection Viewport (1280×960 Telecentric ROI)

Solid Box = Tier-1 YOLO26 + Detectron2 • Dashed Cyan = Tier-2 SAM 2 Zero-Shot Mask

14.5 µm / px

Detected Anomalies & Pixel-Exact Micrometer Metrology

ID / Class	Active Tier	Conf (Raw → Final)	Mask Area	Severity

Hybrid RAG Root-Cause & CAPA Work Order

SMT Convection Reflow Bridging Remediation

CAPA-2026

Cited Authority Standard: IPC-A-610H Class 3

Hybrid Dense+BM25 RRF Score: 0.9420

Physics-of-Failure Root Cause Diagnosis

Prescribed Line Corrective Actions

Active Learning Distillation Loop

Promote SAM-2 zero-shot masks to student YOLO26 training set

▶ Inspect Exported YOLO26 (.txt) & Detectron2 COCO Segmentation Payload ›

# YOLO26 Normalized Bounding Box Labels (class_id xc yc w h)

# Detectron2 Mask R-CNN Polygon Annotation JSON

▶ Inspect OpenCV 3×3 Planar Homography & CLAHE Calibration Matrix ›

Multi-Frame Conveyor Object Tracking

ByteTrack 2-Stage Association + 2D Kalman Filter & OEE

Prevents duplicate pneumatic reject triggers across consecutive 72 FPS frames using a constant-velocity 2D Kalman Filter and ByteTrack Stage-2 low-confidence glare recovery.

Duplicate Rejects Stopped

21 → 3 Unique

Glare Dip Recoveries

2 (0 IDSW)

Line OEE (A×P×Q)

89.08%

Track ID	Defect Class	Frames Hit	Next Kalman Box	Stage-2 Recov

Multimodal Document Intelligence

Engineering Blueprint & COA Tolerance Override Engine

Raw Engineering Change Order (ECO) / Certificate of Analysis (COA) Text:

Lot Disposition: HOLD_LOT_COA_NONCONFORMANCE

Interactive Client + Server OpenCV Inspector

Custom Image Upload, CLAHE Normalization & Sobel Edge ROI Detector

Upload any photo from your computer (or generate a synthetic test wafer/PCB) to compute live CLAHE contrast normalization, 8-bin luminance histograms, and high-gradient anomaly ROIs.

Upload Image File

Cyan Boxes = High Sobel Gradient Edge Anomaly Tiles

Raw → CLAHE Std

18.4 → 41.9

Mean Sobel Energy

24.6

Detected ROIs

3 Regions

ROI ID	Grid Box	Peak Sobel Energy	Est. Area	Conf

Model Optimization & Quantization Matrix (PyTorch 2.6 vs. ONNX Runtime vs. TensorRT 10)

Evaluated on 1,200 1280×960 industrial inspection frames on NVIDIA L4 GPU (24GB VRAM). Quantizing from FP32 to INT8 QDQ achieves a 3.96× throughput speedup and 74% model size reduction with less than 0.009 mAP@50 degradation.

Execution Provider	Precision	Model Size	VRAM	Latency p50 / p95	Throughput	mAP@50	mAP@50:95	Mask IoU

Google Cloud Run — Scale-to-Zero L4 GPU

$0.00 / hr Idle

Attaches a dedicated NVIDIA L4 GPU to a serverless Cloud Run container running onnxruntime-gpu and FastAPI. Scales down to min-instances=0 when idle to eliminate 24/7 GPU burn.

▶ Inspect GCP Cloud Run GPU Deployment Command

gcloud run deploy aegisvision-ai \
  --image us-central1-docker.pkg.dev/$PROJECT_ID/ai/aegisvision:latest \
  --region us-central1 --gpu 1 --gpu-type nvidia-l4 \
  --cpu 4 --memory 16Gi --min-instances 0 --max-instances 4

AWS ECS Fargate + SageMaker Serverless Inference

Terraform IaC Ready

Decouples the stateless FastAPI + LangGraph SOP RAG orchestrator (on ECS Fargate Spot) from quantized ONNX INT8 model execution (on Amazon SageMaker Serverless Inference backed by S3 Model Registry).

▶ Inspect AWS SageMaker Serverless Terraform Snippet (deploy/aws/main.tf)

resource "aws_sagemaker_endpoint_configuration" "aegisvision_serverless" {
  name = "aegisvision-onnx-int8-serverless"
  production_variants {
    variant_name = "AllTraffic"
    model_name   = aws_sagemaker_model.aegisvision_onnx.name
    serverless_config { max_concurrency = 10; memory_size_in_mb = 4096 }
  }
}