Computer Vision Engineer: Real-Time CV Systems

A computer vision model that scores well on a clean benchmark dataset often breaks down against real-world conditions: inconsistent lighting, partial occlusion, camera noise, and network latency across multiple concurrent feeds. The engineering challenge isn't training a model that works once. It's building a pipeline that keeps working when the input is messy and continuous.

My CV work splits into two areas. The first is real-time streaming vision over live RTSP camera feeds: detecting and identifying subjects frame-by-frame without falling behind real time as more cameras are added. For identity matching, I generate 512-dimensional face embeddings with InsightFace Buffalo and index them in FAISS for cosine-similarity search, which scales to large enrolled populations while keeping per-frame lookups fast. The second area is offline and batch CV for domains like medical imaging: fine-tuning YOLOv9 via transfer learning on domain-specific annotated datasets, with data augmentation, class balancing, and hyperparameter search built into the training workflow before a model ships.

A deployed inference pipeline, either a real-time multi-camera identification service or a batch detection API, with the training and evaluation workflow documented so the model can be retrained as new data comes in. This is the architecture behind my CCTV-Based Attendance System: InsightFace embeddings and FAISS similarity search identifying employees in real time across 6 to 10 simultaneous camera feeds and 1,000+ enrolled identities at 95%+ accuracy, and behind my Wrist Abnormality Detection System, a YOLOv9 model fine-tuned on the GRAZPEDWRI-DX paediatric X-ray dataset and served through a Flask inference API.