• Created and deployed Aurva's AI-based context-aware data classification engine, identifying sensitive data across structured and unstructured databases including images with a F1 score of 0.96• Developed algorithms to calculate risk scores for data accessors and the data security posture• Designed and created Aurva's anomaly detection engine, analyzed malicious actors, access patterns, data transaction to identify compromised systems and vulnerabilities.

• Created a predictive Deep Learning and mathematical model to identify drive terrain in real-time using IMU data achieving an accuracy of 93.8 %.• Employed advanced feature engineering techniques to ensure the model's robustness to variations in IMU data, which are impacted by factors such as vehicle type, loaded/unloaded vehicle, and driving speed.• Upgraded Autoware, an open-source software solution, and added stable modules like localization, mapping, perception, path planning using HDL GraphSLAM, Extended Kalman filter, and NDT scan matching with data attributed from GNSS, LiDAR, and IMU. • Robustly tested these modules on entire Mohan estate area of Delhi.

• Extensively reviewed existing Multi-task learning, Muti Object tracking algorithms and trained them on multiple ReID datasets eg Market1501, CUB, CARS196, SOP using PyTorch. • Complex datasets with multiple unique IDs and label formats were carefully analyzed to achieve accurate object tracking and identification.• Used contrastive loss, dissimilarity space, center loss and Classical Machine learning like SVMs to improve a real time object reidentification algorithm and minimized ID switches by roughly 10%. • Accomplished increase in multi-object tracking accuracy or MOTA with 50% less dataset.

• Reviewed literature for existing Multi-task learning and Muti Object tracking algorithms and trained them on the KITTI MOT dataset which comprises of 8 different classes using PyTorch.• Comprising of 7000 unique ids, dataset and label format were analyzed and wrangled to track light vehicles.• Used contrastive loss, dissimilarity space and center loss to improve a real time object reidentificationalgorithm for pedestrians, minimizing ID switches by 14 and significantly increasing multi-object tracking accuracy or MOTA on very first smoke testing.

Developed and robustly tested a complete software stack for a Formula Student driverless car. Spearheaded a team of 40+ students that represented India in the Formula student driverless competition Italy 2022.7th overall international rank in Formula student driverless competition Italy 2022.• 2nd international rank in Business Plan Presentation and 7th international rank in Autonomous design report.

Performed robust ground removal and segmentation on 3D pointcloud, extracted from 2D laser scan of the cones present on racetrack.• Developed various Deep learning based novel pipelines to understand race track scenario and robustly implemented object detection algorithms for robust and real time traffic cone detection.• Robustly implemented extrinsic LiDAR-Camera calibration for projecting LiDAR data onto the image.• Ensured real-time, synchronized data extraction from multiple on-board sensors such as LiDAR, GPS, Camera.

Designed and launched a robust and effective path or trajectory planner for an autonomous car for formula student driverless. An apex curve that satisfies the mechanical constraints of the car and also provides the shortest path around the lap using the perception pipeline of the autonomous stack.Succeeded in implementing and designing an aggressive controller (lateral as well as longitudnal controller) that not only keep our car on the path planned and covers it within the shortest possible time while considering the mechanics and the safety of the car.Implemented a robust and precise SLAM algorithm to make an accurate map of the surrounding using a Lidar and Rgbd camera. Which is robust to outliers and noise. Sensor constarints and delays were also incorporated while implementation.

Exhaustively reviewed existing object detection algorithms and successfully deployed deep learning and computer vision-based pipeline for shipping container detection and converted them to TensorRT. Entire model was initially trained on google cloud platform.Successfully designed and tested computer vision and machine learning-based pipelines, modified and implemented various research papers for optical character recognition which is robust to lightning, weather conditions, texture and rusting effects. Created a synthetic dataset for training on a 3D designing software and automated the complete cycle to generate 1 M examples and then fine-tuned on standard dataset.Worked on data visualization, augmentation and various error check methods to derive meaningful information from the given dataset.

Developed Deep learning and Machine learning based novel techniques and implemented various perception pipelines for driving a car on a road in regular traffic and weather condition. Which is robust to dynamic lightning effect and weather conditions.Developed a NLP and CNN based novel techniques and algorithm for object detection, classification and text generation.

Developed a 3-D model (CAD) of a specialised chamber that would be situated on the multicopter and contain all the electronic equipments, devices and Telemetry necessary for an autonomous hexacopter to takeoff and complete the desired mission.