Muhammad T.

I build Video Intelligence systems that turn raw camera footage into verified events, persistent object tracks, automated clips, structured data, and review-ready evidence. My work goes beyond running a detection model on individual frames. I build the complete Computer Vision workflow around the model: video ingestion, detection, tracking, temporal reasoning, false-positive reduction, event verification, performance measurement, deployment, and operator review. If your current system struggles with missed detections, duplicate events, unstable tracking IDs, difficult lighting, motion blur, occlusion, processing speed, or deployment, I can help turn it into a measurable and maintainable application. WHAT I CAN BUILD • Video object detection, segmentation, classification, and tracking • Automated event detection and timestamped clip extraction • Sports video analytics and highlight-generation pipelines • CCTV, dashcam, traffic, camera, and recorded-video analysis • Persistent multi-object tracking and trajectory analysis • Industrial visual inspection and anomaly localization • Real-time and offline video-processing applications • Raspberry Pi and edge-device Computer Vision systems • FastAPI inference services, background workers, and REST APIs • Human review dashboards for uncertain detections • Structured JSON, database, reporting, and audit workflows • Model evaluation, failure analysis, profiling, and optimization HOW I IMPROVE REAL-WORLD ACCURACY A single model is often insufficient for production footage. Depending on the problem, I combine: • YOLO or PyTorch-based detection • Object tracking across frames • Temporal and geometric validation • Trajectory and region-of-interest rules • Secondary verification models • Confidence calibration and abstention logic • Duplicate-event suppression • Human review for low-confidence results This approach helps reduce false positives and prevents the system from treating every uncertain prediction as a confirmed event. SELECTED PROJECT EVIDENCE BASKETBALL VIDEO INTELLIGENCE I delivered a paid Computer Vision system that analyzes complete basketball games, identifies likely made baskets, verifies difficult events, and automatically exports timestamped highlight clips. The workflow combines object detection, temporal and geometric reasoning, a verification stage, automated video slicing, and a review interface for remaining edge cases. The client provided repeat five-star feedback for speed, flexibility, communication, and delivery. EDGE ROAD PERCEPTION I built a modular road-video pipeline using YOLOv8n, ONNX Runtime, Kalman filtering, and Hungarian-IoU association. The system provides vehicle detections, persistent tracking IDs, lane and drivable-area estimation, annotated output video, frame-level JSONL records, and stage-by-stage performance reports. I also deployed and benchmarked the pipeline on a Raspberry Pi 4B, measuring detector latency, processing throughput, runtime behavior, and the effect of INT8 quantization instead of making unsupported performance claims. BOARDLENS PCB VISUAL INSPECTION I built a production-oriented proof of concept for PCB anomaly inspection using PatchCore, Anomalib, and PyTorch. BoardLens produces model-generated anomaly heatmaps, compares them with independent dataset ground-truth masks, and returns explainable PASS, FAIL, or REVIEW decisions. It includes a FastAPI inspection service, React dashboard, inspection history, analytics, automated tests, and Docker workflows. The current verified baseline is product-specific and is presented with documented measurements and limitations. MY ENGINEERING PROCESS 1. Review representative footage, images, and deployment constraints 2. Define measurable accuracy, latency, and throughput requirements 3. Build a reproducible baseline 4. Identify false positives, missed detections, and difficult cases 5. Add tracking, temporal logic, or verification where required 6. Measure performance on the actual target hardware 7. Package the system as an API, desktop application, worker, dashboard, or edge service 8. Deliver source code, documentation, tests, benchmarks, and reproducible setup instructions DELIVERABLES CAN INCLUDE • Trained or integrated Computer Vision models • Video-processing and inference pipelines • Annotated output videos and automated event clips • REST APIs and background-processing workers • Review dashboards and inspection interfaces • Docker configuration and installation instructions • Evaluation reports and failure-case analysis • Latency, throughput, memory, and hardware benchmarks • Maintainable source code and technical documentation CORE TECHNOLOGIES Python, OpenCV, PyTorch, YOLO, ONNX Runtime, Anomalib, PatchCore, NumPy, SciPy, FFmpeg, FastAPI, Docker, PostgreSQL, React, TypeScript, and Raspberry Pi. I am Top Rated on Upwork with a 100% Job Success Score and multiple five-star reviews for reliability, communication, quality, and on-time deliv