dcjain [at] andrew [dot] cmu [dot] edu
dishankjain0112358 [at] gmail [dot] com
Hello! I am a first-year MIIS student at the Language Technologies Institute at Carnegie Mellon University. I am interested in furthering the state of the art in Large Language Models(LLMs) and making LLMs usable beyond chat. I want to approach this grand goal by applying Reinforcement Learning, Data Centric pretraining and AI Agents.
Previously, I worked as a data engineer at Texas Instruments working in the end-to-end data lifecycle from data collection to model development, deployment and monitoring. I completed my Bachelors in Artificial Intelligence from Indian Institute of Technology, Hyderabad.
Developed a model for multimodal reasoning, enabling it to solve grade school puzzles by analyzing both images and text. Finetuned InstructBlip2 model consisting of 13 billion parameters. Used 4-bit Quantization and Low rank adaptation of transformers to fit the model on a single 40 Gb GPU.
Built a model to generate a 3D point cloud of an object using only one image of the object. First generated point clouds from different viewpoints. Then fused them after rotation
Implemented various Deep Reinforcement Learning Algorithms like Q Learning, Reinforce, Monte Carlo Tree Search and Bandit Algorithms and applied them to games like Lunar Lander, Tic-Tac-Toe and Pong.
Fine tuned a BERT model for Question Answering and used methods like Saliency Maps, Integrated Gradients and T-SNE to explain outputs from the BERT model.
Developed a big data pipeline for moving 200GB of data daily using Pyspark resulting in cost savings of $8M annually. Deployed using Airflow and Kubernetes.
Designed a methodology to recognize Long Landing Events crucial for flight safety. Handled 250 Gb of flight data. Performed Data Mining and developed a Gradient Boosting model capable of predicting Landing Distance with a mean error of only 65 meters.
Built a multilingual RAG system capable of quering Documents and Videos in multiple languages. Used Llama-index, Prometheus, Graphana, Streamlit and FastApi. Led to $500k annual savings.
Built an agent to automate manual tasks in manufacturing processes. Able to work in ambiguous environments, interact with non-technical users and collect correct input information for the task at hand.