Working on the development and testing of the Perception software stack for the autonomous vehicle:

• Created camera-based object detection and tracking module.
• Created camera-based traffic light detection and tracking module.
• Setup infrastructure for benchmarking several state-of-the-art 2D object detectors on autonomous driving datasets such as BDD100K and Indian Driving Dataset.
• Created ROS nodes for deploying these detectors on a Jetson AGX Xavier as part of the autonomous stack of the car using Torchscript and OpenCV with a frame rate of 22 FPS.
• Position and speed of detected objects were estimated by fusing Camera and LIDAR data using a Kalman Filter and calibration information.
• Created a Traffic Light color classifier using Gaussian Mixture Models.
• Working on an annotation tool that uses calibration information to facilitate multi-sensor annotation.

Worked on the development of Near Real-Time RAN Intelligence Controller (RIC)

• Brought up the Near Real-Time RAN Intelligence Controller(RIC) Kubernetes cluster.
• Established connection between Near Real-Time RIC cluster and RIC dashboard to facilitate deployment of xApps and creation of new policies.
• Wrote xApp components to communicate between xApp and the Near Real-Time RIC A1-Mediator.
• Used RIC Message Router APIs to facilitate movement of policies across the cluster components.
• Created Helm charts and Docker images for the xApps.

Worked on using deep neural networks for Natural Language Processing

• Extracted ‘action items’ and filtered out chit-chat from meeting transcripts.
• Trained Bidirectional GRU with attention mechanism for punctuation restoration model.
• Designed and trained a two-layer GRU model in Keras to classify an input sentence as relevant or not.
• Augmented training data by analyzing constituency tree and part-of-speech tags and inserting synonyms.
• Used Google Sentence Encoder to extract sentences with the same semantic meaning by using similarity metrics such as a cosine similarity for targeted training.