Anshul M.

Computer Vision & ML Engineer | PyTorch, OpenCV, OCR, YOLO & Edge AI

My machine-learning work has covered bird detection around high-voltage equipment, industrial sensor anomalies, document images, and demand forecasting. I handle the data, model, evaluation, and deployment as one piece of work. A model can look good in a notebook and still fail once it sees poor lighting, noisy sensors, limited hardware, or changing real-world data. I measure false positives, missed detections, inference time, memory use, operating cost, and data quality before deciding whether the model is ready. For Power Grid Corporation of India Limited, I worked on an edge-AI bird-deterrence system for high-voltage infrastructure. A YOLO and OpenCV pipeline detects birds near transmission equipment, with ONNX Runtime running inference on the deployed device. The system then activates predator sounds and light patterns. Outdoor conditions, limited compute, and false alarms were part of the testing from the beginning. Anomix is an anomaly-detection platform for power plants and similar facilities. It ingests sensor readings, learns normal operating patterns, and flags unusual behavior in about two seconds. I worked on the model pipeline and the application used by plant teams to review anomalies before they turn into equipment failures or safety events. For a restaurant group, I used 279,143 transaction records to forecast hourly order volume, gross sales, and staffing demand across several locations. The work included data preparation, feature design, model comparison, and a dashboard that operations teams could use without reading the modeling code. I also take on: • Object detection, segmentation, and tracking with YOLO or PyTorch • Video analytics and inference on edge devices • OCR and document-image processing • Sensor anomaly detection and predictive maintenance • Demand forecasting and other time-series models • Model APIs, evaluation pipelines, and production deployment I keep datasets, training parameters, model versions, and evaluation results reproducible. When labels are limited, I help define the annotation process and test set before tuning begins. I can also review an existing model, reproduce its failures, and tell you whether the next step is better data, a different architecture, or a deployment change. Main tools: Python, PyTorch, TensorFlow, YOLO, OpenCV, ONNX, scikit-learn, pandas, NumPy, FastAPI, AWS, and Docker.