Projects

Research and engineering work in LLMs, NLP, and AI agents.

Open-Source Python Package

ragbenchpy — RAG Pipeline Benchmarking

  • Published to PyPI a benchmarking tool that runs every chunker × retriever combination for a RAG pipeline and reports accuracy (recall, precision, hit rate, MRR, nDCG), latency (p50/p95/p99), and cost per query side-by-side.
  • Supports fully local evaluation via Ollama and Hugging Face or bring-your-own OpenAI/Anthropic/Google keys, auto-generates benchmarks from unlabeled documents using an LLM, and isolates each run in its own subprocess so one bad combination logs a failure instead of crashing the matrix.
View on GitHub →   ·   View on PyPI →
Foundation Models & Reasoning Research

JEPA Implementation for Mathematical Reasoning & One-Shot RLVR

  • Improved exact-match accuracy on MATH-style problems by 6–8% over an SFT baseline by implementing a Joint Embedding Predictive Architecture (JEPA) pipeline for mathematical reasoning on Qwen2.5-Math.
  • Increased chain-of-thought correctness on GSM8K by 10% and reduced hallucinated reasoning steps by adapting RL pipelines (VERL/rllm) to fine-tune LLM policies with one-shot feedback (RLVR).
View on GitHub →
NLP Research

Hybrid DistilBERT–GAT for Relation Extraction

  • Improved macro F1 on SemEval-2010 Task 8 by 4–6 points over a DistilBERT-only baseline by designing a hybrid DistilBERT + Graph Attention Network (GAT) architecture.
  • Boosted classification precision on structurally complex sentences by 8–10% by integrating spaCy dependency parsing with custom graph attention layers.
View on GitHub →
End-to-End NLP Platform

Keyword News AI

  • Reduced manual data preparation effort by 70% by engineering an end-to-end AI news platform in Python, LangChain, and Hugging Face Transformers to ingest, clean, and normalize unstructured data.
  • Deployed an NLP summarization workflow with spaCy and Docker that generates concise, entity-preserving summaries, served via FastAPI to a React frontend.
View on GitHub →
LLM-Powered Agent

Stock Sense

  • Reduced manual research time from 15–20 minutes to under 1 minute by building a LangChain-based AI agent that orchestrates tools for real-time market data, feature computation, and stock recommendations.
  • Engineered Python pipelines for market ingestion, feature engineering, and LLM-assisted decision support, combining technical indicators with fundamentals.
View on GitHub →