1) Implemented high-performance low level compute custom pre-processing format conversion (transformation) kernel to improve the system latency performance. These kernels were developed for rapid INT8 neural network inference on the customer’s chip.2) Added a generalized padding support (both positive and negative padding) for a large chunk of transformation kernels. Worked on a variety of aspects including quantization, performance stats and optimization of the performance depending on the overall number of execution cycles.3) Created thorough documentation for the compute kernels and solutions created, which helped the team comprehend the arrangement.

1) A comparative study was performed to determine the strategies that can be employed to build an obstacle detection framework for autonomous driving using data collected by LIDAR sensors. 2) Additionally, work was done on obtaining denser data points from a lidar that provides parsed data points using an image super resolution method.

1) Worked on developing and training an Attention-based OCR system using synthetic data to build a real time robust OCR system that works on real time bank forms and OMR sheets.2) Worked on applying a Transformer based model that asses students knowledge based on the previous experience and predict correctness of future response. By performing multiple baseline experiments the state-of-the-art methods were outperformed.

1) Implemented multiple object detection in Aadhar card, Pan card, Driving license etc.2) Wrote a script that generates QR code and random Aadhar number that can be used as training data.3) Created a annotating tool using Flask in Python