LLM based Web Agents • Algorithmic LLM Reasoning • Prompt Engineering[Python, C++, PyClif, Protobuf, Jax, Apache Beam (equivalent), Abseil, Googletest]

Developed a deep learning framework in C++, with CPU & CUDA backend. Contributed to core modules(CUDATensor, Optimisers, Computation graphs) and upcoming methods and layers. This framework enables developers to make CV, NLP, RL based models. Achieved 1.3x faster training speed than Tensorflow (CPU backends).

• Was in a team of 4 people and garnered a letter of recommendation from the Head of Center of Excellence for Artificial intelligence, MCTE.• Developed an ML ensemble that classifies the news articles. Accuracy: 99.6 % on K10 cross-validation on huge dataset.• Used NLP techniques to monitor, extract patterns & provide insights from online news data used for surveillance.

• Made deepfake text generator which inputs a person's past tweets and generates and posts a tweet on twitter using the Twitter API.• Worked with multiple language models of variety of architecture: GPT2 (1.8 B parameter), RoBERTa transformer, etc.• Benchmarked fooling rate and compared finetuned GPT2 and custom trained model using Giant Language model Test Room(http://gltr.io/dist).• Fooling rate: - GPT2 (1554 M parameters) : 61 % - Custom trained model(124 M): 53 %