Muhammad J.

I build real-time computer vision that runs in production. Object detection and tracking with YOLO, optimised with TensorRT for NVIDIA Jetson and other edge hardware, or served in the cloud. Multi-camera video analytics, pose estimation, OCR and document AI, and face and person re-identification. Most computer vision projects look good in a demo and then fall apart in the real world, when the lighting changes, the camera moves, or the feed is noisy. I build for the opposite. My focus is systems that stay accurate in real conditions, run in real time, and actually ship. WHAT I BUILD Object detection and multi-object tracking with YOLO, RF-DETR, ByteTrack and BoT-SORT. Real-time inference optimised with TensorRT and ONNX for Jetson, Hailo and cloud GPUs. Multi-camera video analytics, including person tracking and re-identification across cameras. Industrial inspection and defect detection, measured in real units rather than pixels. Pose estimation and movement analysis, on a server or on-device. OCR and document AI pipelines, including handwriting and structured extraction. Full delivery, from the trained model to a FastAPI service, a Docker container, or a dashboard your team can actually use. SOME OF WHAT I HAVE SHIPPED A real-time drowning-detection system on NVIDIA Jetson, running on site at sub-100ms latency. A multi-camera person tracking and re-identification pipeline running in production on AWS. An inline inspection system on a bakery line, measuring 14 units per row in millimetres from two stitched cameras at around 26 fps. Four camera streams at roughly 30 fps each on a single Hailo-8 with no dropped frames. rubric.pk, a live document-AI product that reads handwritten answers with OCR and an LLM, with paying users. A smart-fridge retail billing system that tracks products and reads tags to build a live bill. X-ray void detection on electronic boards, and food-manufacturing inspection on constrained edge hardware. RAVE, an iOS app with on-device pose estimation, live on the App Store. HOW I WORK I start with your hardware, data and latency targets before writing any code, because those decide the approach. I pick the right model for the job instead of forcing one, benchmark accuracy and speed at each step, and hand over clean, documented code your team can maintain. When a detection is uncertain, I design the system to flag it for review rather than guess, which matters for anything used in the real world. COMMON QUESTIONS Can you work with our existing model or dataset? Yes. I can optimise, retrain or extend what you already have. Edge or cloud? Both. Jetson, Raspberry Pi and Hailo at the edge, AWS or GCP in the cloud. Can you take a proof of concept to production? That is one of my most common projects. Do you handle the API and app too? Yes. FastAPI services, Docker, and the web or mobile front end, so the model becomes a usable product. How many cameras can one device handle? Depends on the box and the model, but I benchmark it before anyone buys hardware. INDUSTRIES Retail and logistics, security and surveillance, manufacturing and quality control, healthcare and safety, sports analytics, education. I also run a small engineering team at Sovavis, so I can take on larger end-to-end builds as well as focused, well-defined tasks. If you need computer vision that works outside the lab, on real footage and real hardware, send me a message with a bit about your project and I will tell you honestly how I would approach it.