• Built a multi-vehicle tracking REST API using FastAPI and YOLOv8n, maintaining persistent object identities across occlusions (up to 5 missed frames) by implementing SORT (Simple Online and Realtime Tracking).
• Improved real-time cloud processing speed on Hugging Face Spaces by bypassing YOLO detection on 50% of frames via Kalman predictions, utilizing FP16 inference, and encoding WebM streams.
• Conducted workshops and classes on Machine and Deep Learning for 100+ BE/BCA/Diploma students

• Built a multi-agent AI system using Llama vision language models and Visual RAG to automate surgical suturing evaluation. Curated a 514-image dataset, achieving an RMSE of 2.2/10 against surgeon annotations.
• Deployed a Python/PyQt5 computer vision application with multi-threaded OpenCV live feeds for real-time craniotomy scoring. Integrated a Grad-CAM pipeline to visually highlight spatial drilling performance.
• Configured 3 ESP32S3 controllers for a wearable device to study hand-eye coordination during endoscopy.

Participated in a 8-week program at IIT Delhi covering image processing, computer vision, and physically-based rendering.

• Implemented CNNs from scratch for visual recognition and constructed feature extraction pipelines with PyTorch.
• Built a multi-image house price prediction model that fuses visual features (facade appearance, neighborhood density) with structured regression — bridging visual perception and quantitative inference.
• Explored photorealistic 3D rendering through ray-tracing algorithms and light transport simulation techniques.