Deep Reinforcement Learning: Advantage Actor Critic
Implemented Advantage Actor-Critic (A2C) with generalized advantage estimation from scratch in PyTorch; trained on OpenAI Gymnasium environments (Lunar Lander, Cart Pole) to validate convergence and policy stability.
Agent landing a space ship after learning from experience
RaDeR: Retrieval-Augmented Dense Retrieval
Research paper on enhancing dense retrieval models through retrieval augmentation techniques for improved information retrieval performance.
AI Poker Agent with Deep Reinforcement Learning
Engineered a deep reinforcement learning agent for Texas Hold'em, leveraging a Monte Carlo Tree Search (MCTS) guided by a custom deep learning policy network to learn decision-making under uncertainty; secured 7th place in a 22-team tournament.
CLIP Image and Text Embedding Pipeline and Inference System
Built an end-to-end ML inference API for CLIP embeddings using FastAPI; implemented adaptive batching with a queue-and-timeout mechanism to maximize GPU utilization; indexed 100k images in Pinecone for similarity search.
Fine-Tuning LLMs for System Initiative Prediction
Quantized and fine-tuned Llama-7b on tens of thousands of MS Dialog conversations, presenting findings at a final poster session.
Face-to-Sketch Conditional GAN
Built a generative image model based on a research paper, improving sketch quality by pretraining on a larger face image dataset.
Reducing Echo Chambers in Collaborative Filtering Models
Preprocessed user ratings of books and used graph algorithms to create a prototype model for reducing echo chambers.
