Tathagat

MiniLLM

Built a GPT-style language model from scratch with 57.5M parameters, trained on 150M tokens. Implemented pre-training, QLoRA fine-tuning, DPO alignment, and deployed it as a FastAPI service with streaming inference.

View details

DevOps AI Agent

Built a RAG-based AI assistant that helps troubleshoot DevOps issues using retrieval, tool calling, and LLM reasoning. Reduced response latency from 28s to 10s through retrieval optimization and memory improvements.

View details

Proposition RAG

Designed a proposition-based RAG indexing approach inspired by Dense X Retrieval. Instead of embedding raw chunks, the system indexes atomic facts and retrieves full context through metadata, improving retrieval hit rate from 80% to 90% while reducing latency.

View details