• Optimised mask segmentation tasks by replacing SAM and Grounding DINO models with lightweight SegFormer models,drastically reducing computational load and server costs while improving inference times.• Developed a tool using SegFormer & IDM-Vton to replace non-shoppable Pinterest outfits in app’s feed with outfits fromimages across e-commerce sites, increasing the exact match percentage of fully shoppable content from 56% to 89%.• Introduced an inpainting feature using BrushNet and PowerPaintV2, allowing users to edit images by dragging arectangle over the desired area, providing a prompt, and adding accessories, with an inference time under 10 seconds.

Introduced Temporal CNN models for accurately classifying stages of vegetable cooking and dish consistency, inducing robustness through variations in cooking conditions, and removed redundant hard coded lines.Modified existing rice model to implement edge detection and combined with a CLIP model to provide more context to solve the issue of residual water pockets beneath.Developed transformer-based LLM models with RAG to convert text-based recipes from the internet intomachine-understandable format which allowed users to save custom recipes.

As the Chief Operations Officer at Pravega Racing Driverless, the Formula Student Driverless team at VIT Vellore, I spearhead pioneering projects in autonomous vehicle technology for Formula Student competitions. My expertise encompasses object detection using neural networks, path planning, SLAM, stereo camera, IMU, loop closure, Kalman filtering, and sensor fusion. Committed to pushing the boundaries of autonomous racing, I'm dedicated to fostering innovation and preparing emerging engineers for the unique challenges of Formula Student Driverless competitions.

Standardized and preprocessed large datasets on children’s education and welfare using SQL Developer, enhancing data consistency and query efficiency.Analyzed performance trends and identified areas for improvement in state and district-level child welfare assistance.Created visual reports highlighting key trends and metrics, aiding data-driven decision-making for welfare initiatives.

• Developed and implemented a face detection and classification model using YOLOv4 for surveillance cameras, improving the accuracy and robustness of the existing system.• Expanded the dataset by incorporating video recordings from open source datasets and applied various augmentations to simulate real-world surveillance conditions, enhancing the model’s performance and robustness.• Optimized model training through techniques such as hyperparameter tuning and advanced data augmentation, leading to enhanced performance and reliability in real-world applications.