Muhammad Anas

<p>Building the real-time AI surveillance analytics service for the VIDIZMO enterprise video platform: live camera ingestion, GPU object detection, configurable alert rules, and continuous DVR recording.</p><p> Develop, test, and validate machine learning algorithms and computer vision features for the VIDIZMO Cognitive & Computer Vision platform, including custom neural network components, loss functions, optimizers, and evaluation metrics tailored to specific detection problems.</p><p> Architected the pipeline as a GPU/CPU split: detection, tracking, and geometry on GPU; alert decisions on CPU worker threads.</p><p> Operators edit surveillance rules live with no service restart, and one slow camera never blocks another.</p><p> Replaced an unreliable vendor direction-detection feature with a custom trajectory algorithm, adapted from published trajectory-analysis methods and tuned on real deployment footage.</p><p> Hardware testing showed the vendor feature false-flagged stationary objects and missed real movement; confirmed the defect with NVIDIA engineering before committing to a rewrite.</p><p> Built the rule engine as an extensible detector framework where each rule implements one uniform contract and registers independently, replacing hand-coded per-rule branching in shared processing code.</p><p> Delivered zone intrusion, loitering, line crossing, overcrowding, and direction rules with object-lifecycle tracking and edge-triggered alerts.</p><p> Designed the camera resilience subsystem: a three-state circuit breaker with exponential backoff and decorrelated jitter, retry scheduling, per-camera state tracking, and automatic quarantine of persistently failing cameras, so transient network faults recover without operator intervention.</p><p> Designed a leg-based multi-model inference architecture on NVIDIA DeepStream, extending its standard multi-model pattern to support chained detection stages alongside independent standalone legs with results funneled into a shared queue for per-leg rule validation and bounded per-branch queues so a slow model never affects the rest.</p><p> Diagnosed and resolved critical failures during pre-production testing, including a service-wide deadlock blocking new camera additions and silent stream stalls where cameras stopped delivering video without disconnecting, fixed with a frame-staleness watchdog.</p><p> Integrated the service with the platform backend over gRPC, RabbitMQ, and Kafka, defining the message contracts consumed by downstream .NET services and the operator dashboard.</p><p> Maintain 80+ unit and integration test suites.</p><p> Review and validate AI-generated outputs with ownership of final deliverables before release.</p>

<p>Engineered a multi-camera processing architecture that cut inference time from 7 minutes to 8-10 seconds across 177+ RTSP streams running 5 concurrent AI modules.</p><p> Achieved a ~50x throughput improvement through batch processing, queue-based multiprocessing, and a cycle-aware camera dispatch system.</p><p> Built a multi-camera ANPR system for Yunus Textile monitoring 40-60 trucks daily at 90%+ accuracy in field testing, supporting SOP compliance, weight tracking, gate-pass verification, and real-time WhatsApp alerting.</p><p> Deployed a 7-camera ANPR system for Taqat (Saudi Arabia) processing 60-70 vehicles daily using YOLOv8 and EfficientNet, reaching 90%+ accuracy in field testing on vehicle and material classification.</p><p> Delivered dynamic PPE and safety monitoring across 278+ cameras at 83% accuracy, using site-specific model configurations for helmet detection, safety vest detection, housekeeping, restricted-area monitoring, and supervisor presence, across Engro, Axens, and Candle.</p><p> Implemented blockage and intruder detection for high-security zones at 95%+ accuracy for Yunus Textile, replacing manual monitoring with automated surveillance.</p><p> Built a facial recognition system using MTCNN and FaceNet with 128-D embeddings for identity matching, reaching 90% accuracy in production for Yunus Textile.</p><p> Developed a modular visual inspection pipeline using YOLOv8 and AnomaLib across 10+ football panel types for Forward Sports (an Adidas supplier), achieving 80%+ accuracy on major defects and 75%+ on minor defects.</p><p> Built an LLM-based extraction pipeline reaching 97% accuracy on structured data extraction from multi-page real estate documents, processing 5-6 documents daily with automated page selection and contextual image filtering.</p><p> Built an LLM-based social media automation prototype for Facebook, combining AI-generated captions and visuals with scheduled auto-posting, content calendar generation, and automated ad campaign creation.</p>