OMNI-VISION ARCHITECTURE
█
Training a model is easy. Deploying it is hard. I architect end-to-end Computer Vision pipelines—from CVAT Data Labeling to YOLOv8 Training to FastAPI Production Deployment.
Slices high-res image into grid to maintain small-object visibility.
Merges detections using Non-Maximum Suppression (NMS).
"""
LIFECYCLE // DATA TO DEPLOYMENT
Rapid Annotation
Designing semi-automated labeling workflows using CVAT and auto-labeling scripts to speed up dataset creation by 500%.
Custom Architecture
Fine-tuning YOLOv8 and PyTorch models. Implementing Tiled Inference (SAHI) to detect microscopic defects on 4K resolution images.
Production API
Wrapping models in high-performance FastAPI backends with async queue management to handle concurrent video streams without crashing.
PROJECT OPTIC-GUARD
Manufacturing Defect Detection System
The client had 4K resolution images of assembly lines. Standard YOLO models failed because they downsized the image, making small defects invisible.
Implemented Tiled Inference—slicing the 4K image into 12 smaller "tiles" to maintain resolution.
Each tile is processed in parallel by the model.
A custom algorithm merges the detection coordinates back onto the original 4K frame, removing duplicate overlaps (NMS).
Real-time detection of <2mm defects with zero latency lag.
Real-time YOLOv8 Detection
ADVANCED CAPABILITIES
Detecting microscopic defects, scratches, or misalignments on assembly lines.
Tiled Inference (SAHI) for small object detection on 4K hardware.
Turning scanned PDFs and invoices into structured JSON data.
LayoutLM & PaddleOCR for complex table extraction (not just text reading).
Real-time monitoring for PPE (Hard Hats, Vests) or unauthorized zone intrusion.
Real-time RTSP stream processing with <50ms latency for immediate alerts.
Tracking foot traffic, dwell time heatmaps, and queue lengths.
Object Tracking (ByteTrack) to maintain identity across frames.
TECHNICAL FAQ
A: Negative. I utilize Transfer Learning (using pre-trained weights from YOLOv8/11). We can achieve high accuracy with as few as 50-100 labeled examples per class to start a prototype.
A: Affirmative. I optimize models using TensorRT and ONNX formats, allowing heavy neural networks to run smoothly on Jetson Nanos, Raspberry Pis, or local CPU servers without needing a $10,000 GPU cluster.
A: I build "Air-Gapped" capable containers. Your video feeds process locally (on-premise) or in a private cloud VPC. No data is sent to third-party APIs unless authorized.
TILED INFERENCE PIPELINE
Technical breakdown of the defect detection system
Powered By
