Zeeshan J.

Hi, I'm Zeeshan 👋 Give me something to build and I'll build it so it works for you 24/7. That's my promise. Send me an invitation or a message and I'll send back a personalised Loom video showing exactly how I'd approach your problem, before you spend anything. I'm a Computer Vision Engineer with 5+ years building production AI systems, not demos and not notebooks. One platform I built ran batch inference over a full year of the top 10 Urdu news channels and exposed the results through a queryable analytics database. My work is strongest where the footage is messy, real-time, or operationally constrained. WHAT I BUILD 1. Video & Object Detection Pipelines - Detection, tracking and segmentation (YOLO, YOLOv8, D-FINE, Detectron2) - Multi-object tracking and counting across full-length video - Video-to-metadata systems that feed analytics databases and automation - High-performance OpenCV + PyTorch pipelines, multi-GPU training (CUDA, cuDNN) 2. Pose Tracking & Motion Analysis - Pose estimation and landmark detection (MediaPipe, DWPose) - Sports video: my golf pipeline detects swings and putts through a pose-driven state machine over spine angle, foot stability and wrist velocity, a custom-trained YOLOv8 ball tracker, and audio-based impact validation, producing confidence-tagged clips - Frame-accurate event detection across long recordings 3. Real-Time & Live Inference - Live recognition with state validation on streaming input - Facial recognition, liveness detection and dwell-time analytics (FaceNet, DeepFace) - Latency and throughput tuning, deployment-aware model selection 4. Document AI & OCR - End-to-end extraction from scanned, multi-column and multilingual documents - Right-to-left and low-resource scripts: Urdu, Shahmukhi and Dhivehi (Thaana). On a client's Dhivehi OCR project, post-processing cut character error rate from 30% to 12% - TrOCR, EasyOCR, layout analysis, binarisation, skew correction, denoising - GAN-based augmentation for robustness on low-resource scripts - Searchable archives and semantic video search (Whisper, LangChain, FAISS) 5. Video Forensics & Privacy - Video forensic analysis: integrity and provenance checks, duplicate verification, optical-flow temporal analysis, pose tracking - Face and plate anonymization for images and video 6. How I Deliver - FastAPI / Flask services wrapped around the vision models - Clear milestones, working code, and a recorded walkthrough at handover BACKGROUND I've worked independently with international clients since 2023 and part time as a Visiting Lecturer at the University of Narowal. Previously Assistant Manager Research at KICS, UET Lahore (2021 to 2026), where I built production video analytics, OCR and generative video systems. Former Visiting Researcher at the University of Konstanz, Germany. Published researcher: LREC 2026 (accepted, co-author) and IEEE Access (under review). OPEN SOURCE Mixed-traffic vehicle detection: my own dashcam dataset, a fine-tuned YOLO26n model and a proximity risk app, all public. media-anonymizer: face and plate anonymization for images and video. STACK Python, OpenCV, PyTorch, TensorFlow, YOLO, Detectron2, MediaPipe, FastAPI, FAISS, LangChain, Whisper, ffmpeg, CUDA If you need a vision system that has to actually work, on real footage and at real volume, send me the problem. I'll tell you honestly whether I'm the right fit.