04 / Computer Vision & Edge AI
In progressMultimodal Defect & Threat Inspection System
Target role · AI/ML Engineer
A multimodal inspection pipeline combining lightweight vision models, signal metadata, optimized inference, and anomaly records.
- Target: 42 FPS edge inference
- Target: 94.2% mAP across five categories
Problem & direction
Core problem: Visual inspection requires image evidence and operational metadata to be interpreted together.
An in-progress inspection system combining visual evidence with tabular or signal metadata, spanning model training through edge-optimized serving.
How it works
- 01
Prepare a defect or threat dataset and operational metadata
- 02
Fine-tune YOLOv8 or MobileNet with PyTorch and OpenCV
- 03
Fuse image embeddings with structured features
- 04
Quantize with ONNX Runtime or TensorFlow Lite and serve through FastAPI
Engineering details
Multimodal classification
The planned pipeline combines visual features with operational metadata rather than interpreting each image in isolation.
Edge-oriented inference
ONNX Runtime or TensorFlow Lite will optimize CPU inference. The 42 FPS and 94.2% mAP values are target benchmarks.
Structured detection API
FastAPI will return bounding boxes, confidence scores, and category labels through a typed response contract.
Auditable anomalies
Flagged detections will be stored in SQL so outputs can be reviewed, traced, and compared with later outcomes.