top of page

Projects

Multi-agent Environment for OpenAI Lunar Lander 

  1. Developed a Multi-agent Lunar Lander environment using OpenAI as a base and modified the inbuilt PettingZoo libraries to customize the environment settings based on the requirements.​

  2. Utilized python libraries like PyTorch and TensorFlow to create a DQN RL algorithm for cooperation and competition between lunar lander agents. Demonstrated the successful landing of multiple agents using a UI created using the Python libraries Pygame and Box2D.

  3. Published a third-party environment in the PettingZoo forum for multi-agent reinforcement learning that is compatible with any other OpenAI APIs, developed for upcoming researchers interested in Multi-Agent Reinforcement Learning.

unnamed.png

Legal Named Entities Extraction (L-NER) 

  1. Developed two different deep learning transformers based on multi-label entity recognition on a dataset of Indian legal documents.

  2. Preprocessed the raw data containing 14,444 sentences into IOBE tagged sequence with the help of NLP toolkit and Spacy.

  3. Pretrained LegalBert transformer for generating feature embeddings and labeled entity-level sequences with LSTM networks

  4. Implemented BERT-based-cased transformers and decoded tag-labels using SimpleTransformers to generate a validation score of 90.7%.

NLP project_edited.jpg

SigML Hamnosys code Generator for Sign Language

  1. We initiated a benchmark for creating this language independent HamNoSys code because there is a gap in the literature regarding the Multilingual Sign Language recognition system called HamNoSys Unicode. On the internet, our suggested architecture is the first of its kind. 

  2. By taking pictures of us communicating using sign language, a benchmark dataset was produced to address this issue. To increase the precision of the suggested model, this was accomplished with the aid of tools like Mediapipe and OpenCV for image processing techniques like image blurring, sharpening, and segmentation. Many mathematical formulas have been developed to analyze the key Sign Language components that can be used in models. 

  3. On the first of the four categories, an accuracy of 89% was produced, and 60% on the other three. The proposal's scalability adds to its significance for the study of sign language in the future.

Anomaly Detection using Real time Video

  1. In order to create a Deep Learning-based solution for the Driver Attention Estimation, the posture and movement of a person were examined using the YOLO, Mediapipe, and OpenCV libraries.

  2. By identifying unusual pedestrian behavior with CCTV, we were able to add more information to the model, we designed an end-to-end alert mechanism system for any abnormality noted using Transfer Learning in Tensorflow.

  3. For better interaction, we created a user interface in Streamlit.

13735_2022_227_Fig1_HTML_edited_edited.jpg

SafeDriveDriod

  1. Developed a road safety device app, with rich user experience and available in the regional language. The application blocks incoming calls and messages and informs the caller that the driver is currently driving. 

  2. Upon reaching a speed of over 20 km per hour, the app automatically activates and blocks all incoming phone calls and text messages. Upon determination of an accident, emergency contacts are immediately notified, enabling medical assistance to be rendered in a timely manner. 

  3. Designing a speed limit system based on road signs, traffic density, school zones, and adjusting voice over of the speed limit. Exploring its potential for next-generation AI cars.

bottom of page