Built an autonomous research agent using LangGraph + DeepSeek LLM: decomposes complex queries into sub-queries, runs parallel Tavily searches, summarizes findings, self-reflects with 0–10 confidence scoring, and generates fully-cited markdown reports. Supports configurable search depth/iterations, interactive Streamlit UI, CLI testing, and programmatic Python API for multi-hop research automation (e.g. academic surveys, etc.).

Developed "Headnote LLM", a domain-specific 8B-parameter model by fine-tuning Meta Llama-3.1 on 10k+ legal judgment datasets. Achieved state-of-the-art performance on legal reasoning tasks while adding only ∼168 MB of trainable parameters via LoRA adapters.

Engineered a modular algorithmic trading platform focusing on cointegration analysis (Johansen test) and Z-score signal detection. Implemented dynamic hedging using Kalman Filters and XGBoost for signal generation, integrating risk management (VaR) and real-time visualization.