I build production computer vision systems in Python and PyTorch: self-hosted pipelines that combine object detection, OCR, and vision-language models, with no third-party API dependency and client data kept on-premise.
Document and drawing analysis. Detection, OCR, and vision-language pipelines that run unattended in production, including vector geometry parsing and tag extraction for technical drawings.
Model serving and reliability. GPU memory gating, per-request checkpointing, and a vLLM migration that cut a four-document batch from 6 minutes to 2.
Edge and aerial detection. YOLO training and benchmarking for real-time edge deployment and UAV vehicle detection.
First author of "Deep Learning Application for Vehicle Detection through Surveillance Drones", presented at IEEE ICECET 2024. It benchmarks YOLOv8, Faster R-CNN and SSD under domain shift and releases the Pak Vehicles UAV dataset.
YOLO · PaddleOCR · vLLM · Transformers · LoRA and ControlNet
Vehicle-Detection-using-YOLOv8-Faster-R-CNN-and-SSD: code for the ICECET 2024 paper.
Paddle_OCR_TextDetection_and_TextRecognition: PaddleOCR detection and recognition.
YOLOX: YOLOX training, validation and testing.
Open to computer vision roles and research collaborations.