1. Real-time OCR (optical character recognition) implementation to detect port crane ID’s using EAST (Efficient and Accurate Scene Text detection) text detector, to track the cranes based on the checksum value. Created GUI on Tkinter having a port camera, videos and image interface for OCR detection.

2. Implemented automatic image orientation correction using RotNet model.

3. Did image preprocessing to improve OCR detection accuracy and compared the results with available OCR websites. Achieved an average accuracy of 80%.

Client: Suzuki XF351, JL, and Honda IC

1. Worked as a Software Developer using Cubesuite++ software for the creation of different modules, their drivers and scheduler using the RL78 (R5F10CME) microcontroller for Suzuki Instrument Cluster. The modules include that of ADC, Button, EEPROM, Fuel, Ignition, Power on Cycle, LCD, Odometer, Tripmeter, Speed, Oil Change Indicator, Serial and Interrupt Handler.

2. Implemented UART communication in Suzuki Cluster different from existing I2C communication mode.

3. Unit testing in VectorCAST environment.

Client: Nissan Short Range Radar

1. Deep learning for applications such as traffic sign recognition, pedestrian detection, lane change assist for the autonomous driving systems.

2. Building and refining neural network architectures like Le-Net 5 to continuously improve the accuracy of the recognition and detection systems.

3. Worked also for Requirement based testing using CAPL scripting in CANoe environment for the software of a Nissan project (SRR320NN) for generation 3, 5 and 6.

4. Worked on Toyota ARS (Advanced Radar System) project and modify the existing code as per different platforms in order to deploy different functionalities based on the destination country where the product would be eventually delivered.

5. Experience in working in a Japanese team with cross-functional interaction.

1. Ran experiments on dialog acts classifiers like BERT (Bidirectional Encoder Representations from Transformers), CNN (Convolutional Neural Networks) and Bi-directional LSTM (Long Short-Term Memory) using Tobacco depositions to get the predicted dialog acts for question-answer pairs.

2. Compared the three models using precision, recall, and F1-score criteria. BERT turned out to be better with an average F1 score of 84%.