Vardhan Mistry

• Developed a real-time semantic segmentation framework for camera soiling detection • Designed training strategy combining real, synthetic, and clean data to improve generalisation • Developed mask-guided GAN-based pipeline (CycleGAN + VAE) for synthetic mud generation • Reduced false positive rate from 4.03% to 0.17% while maintaining Dice up to 0.94 • Benchmarked DeepLabV3+, SegFormer, and PIDNet • Deployed PIDNet via ONNX achieving 7 ms inference for real-time edge execution

• Developed physics-based raindrop simulation improving synthetic data realism for perception models • Automated RGB + mask dataset generation, reducing manual annotation effort by 100% • Benchmarked against Stable Diffusion 2.1 using perceptual metrics and human evaluation • Delivered actionable insights for algorithm improvement through detailed performance analysis reports.

• Developed OpenCV-based defect detection prototypes, reducing manual inspection time by 35% • Optimized camera calibration and lighting setup, reducing false-reject rates by 12% • Contributed to industrial automation deployment in a large-scale smelter project within a cross-functional team Also, • Integrated and commissioned a FANUC 400i industrial robotic system. • Developed Power BI dashboards for data-driven operational tracking and executive reporting.

Focus: Computer Vision, Robotics, Automation, ADAS

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