Wajid A.

I fix and ship computer vision systems that have to work on real cameras. Most recently I worked on a live multi-camera surveillance platform: optimizing its YOLO agents, building new ones, and taking license plate recognition from zero alerts to 99%-confidence reads in production. Rated 5★. If your model works in a notebook but fails on live RTSP streams, edge hardware, or messy real-world footage, that is the problem I solve. RECENT PRODUCTION WORK AI video surveillance platform (construction and commercial sites) Multi-tenant SaaS running 9 detection agents across concurrent RTSP streams: intrusion, loitering, occupancy, vehicle presence, entry-after-hours, late checkout, fire/smoke, license plates, and package tracking. → Debugged LPR end to end, from plate detection through PaddleOCR to snapshot rendering. Fixed silent GPU incompatibility, coordinate-space bugs, and config overrides. Plates now read at 93–99% confidence in production. → Built a package lifecycle agent (delivered, abandoned, picked up). Merged 7,000+ images and trained a custom YOLO11n to 97.8% mAP50. → Collected and labeled data and trained a custom fire/smoke model. → Built a vehicle color classifier (MobileNetV3) and per-class confidence filtering to cut misclassifications. → Resolved a production CUDA crash, with 38 regression tests added. Medical imaging: spine X-ray analysis 25-class vertebrae segmentation (C1–L5) with automated clinical measurements: disc height, segment angles, listhesis. Custom ConvNeXt encoder with attention, trained on real and synthetic X-rays. 0.92 Dice. Client review: “top-tier expertise… sub-pixel metric accuracy.” Edge AI for manufacturing and smart cities TensorRT FP16 defect detection on Jetson Xavier with PLC pass/fail integration. LiDAR-camera fusion for traffic monitoring (ROS, DeepStream). WHAT I DO → Detection, segmentation, tracking: YOLO, SAM2, ByteTrack, custom PyTorch → OCR and LPR: PaddleOCR, plate detection, multi-read consensus → Optimization: TensorRT FP16/INT8, ONNX, Jetson Nano, NX, Orin, Xavier → Production plumbing: FastAPI, RTSP, Docker, CI/CD, MQTT, S3, Supabase → Vision-language reasoning: Qwen2.5-VL, RAG over inspection history HOW I WORK I trace problems to the root cause instead of patching symptoms. I send short written updates and flag blockers early. You get code you own, documented and deployed. Good fit: debugging a CV system that underperforms in the field, deploying a model to edge or GPU cloud, or building a detection pipeline from camera to dashboard.