Vardhan Mistry

Developed DirtyGAN (CycleGAN+VAE) from scratch to generate realistic soiling data for automotive cameras. Built a 20k+ hybrid dataset (synthetic + real) and trained segmentation models (DeepLabv3+, SegFormer, PIDNet) in PyTorch, improving IoU/Dice by 6% and halving false positives. Delivered a real-time prototype with detection-to-cleaning trigger logic, cutting water/energy use by 50% while ensuring robust ADAS performance.

Built a Python-based procedural soiling generation pipeline producing unlimited synthetic images with pixel-accurate masks for scalable ADAS training. Converted academic methods into a flexible, parameterized framework, expanding datasets 80% faster than manual collection. Benchmarked vs Stable Diffusion 2.1, where the pipeline achieved ~70% higher realism/consistency for lens soiling, directly boosting computer-vision model robustness.

Coordinated cross-functional teams and vendors for steel structure and Fanuc 400i robot installations, ensuring timely execution, material availability, and adherence to safety standards; reported daily progress using Power BI. Managed SAP MM workflows for material handling, PR/PO creation, and vendor billing, while performing on-ground supervision and process monitoring to maintain smooth project operations.

Computer Vision, Robotics, Autonomous Driving, Automation

Majors: • Computer Application in Mechanical Engineering • Computer Techniques and Programming Language • Control System • Electrical and Electronics • Industrial Engineering