Graduate Projects

EduNotes - Multi Agent Study Assistant

Aug 2025 - Present | UMass Amherst | GitHub | Live App Demo

  • Multi-agent RAG study assistant centralizing study materials into a semantic knowledge base users can search and build over time
  • Content classification, adaptive summarization and notes generation based on input type with dynamic query routing
  • Multi-stage web search via DuckDuckGo and Google fallback with AI-driven query refinement at each step
  • Orchestrated via LangChain & ChromaDB, to produce AI-generated flashcards, quizzes, notes, summaries, and progress analytics
  • Research Mode with vision model analysis of PDF figures, tables, equations along with references from OpenAlex, arXiv & Semantic Scholar
  • AI Chat agent with 7 modes (Socratic Tutor, Compare Topics, Research Writer, etc.) powered by Gemini 2.5 Flash and searchable conversation history
  • Distributed workload across Llama 3.3 70B for generation, Llama 3.1 8B for lightweight tasks, and Llama 4 Scout for vision analysis
  • Built with FastAPI & Streamlit, supporting 6+ input formats (PDF, URL, text, topic, web search) and 3+ output formats

PerFine, an Iterative Critique-Refine Framework for Enhancing LLM Personalization

Jan 2025 - Jul 2025 | Cisco, UMass Amherst | arXiv

  • Training-free critique-refine Agentic RAG framework for long-text personalization using LangChain
  • Enhanced outputs using profile-grounded critic feedback retrieved via Pinecone, FAISS, MCP for style and content refinement
  • Evaluated using LLM-as-a-Judge (G-Eval), improving personalization by 13% and Meteor by 10% over RAG-based baselines

Text Generation in Continuous Space

Aug 2024 - Dec 2024 | UMass IESL Lab

  • Research under Prof. Andrew McCallum
  • Autoregressive model with lookahead decoding in superposition
  • Efficient token sequence search with just two forward passes
  • Applied to MT5 for improved machine translation quality

Automated Data Cleaning Pipeline

Jul 2025 - Aug 2025 | UMass Amherst | GitHub

  • Production-ready system that auto-learns cleaning rules
  • Intelligent field detection with full monitoring
  • Detailed reports for data quality tracking

LS-GAN: Human Motion Synthesis with Latent-space GANs

Jan 2024 - Dec 2024 | UMass Amherst | IEEE WACV'25

  • Text-to-motion generation using GANs, VAE, CLIP in latent space
  • Trained on HumanML3D and HumanAct12 with distributed PyTorch Lightning
  • Achieved FID of 0.48 with 91% FLOPs reduction vs diffusion models
  • Published at IEEE WACV 2025 Workshop

Realtime Stock Analysis

Oct 2024 - Dec 2024 | UMass Amherst | GitHub

  • Real-time sentiment analysis from news using Kafka, PySpark, Flan-T5
  • 74% sentiment accuracy at 3.5 throughput
  • Weighted scoring for daily stock sentiment trends

Safe to Serve: Aligning Instruction-Tuned Models for Safety and Helpfulness

Jan 2024 - May 2024 | UMass Amherst | arXiv

  • Aligned LLaMA-2 7B using LoRA, QLoRA on PKU-SafeRLHF
  • Implemented SFT, RAFT, RLHF, DPO in Unsloth and TRL
  • 93% safe on DPO vs 40% on SFT (Llama-Guard evaluation)
  • LLM-as-a-judge for safety and helpfulness metrics

Automated Model Selection for Tabular Data

Sep 2023 - Dec 2023 | UMass Amherst | arXiv

  • Feature interaction-aware model selection framework
  • Priority-based random grid search and greedy search methods
  • Forward selection and backward elimination approaches

Optimization in Reinforcement Learning

Sep 2023 - Dec 2023 | UMass Amherst | GitHub

  • Implemented 5 RL algorithms: Reinforce, Actor-Critic, PPO, SARSA variants
  • Tested on CartPole, Acrobat, 687-Grid-World
  • Achieved mean rewards: 470/500 (CartPole), -100/0 (Acrobat)

Undergraduate Projects

HOME Decor VR Application

IIITDM Kurnool

  • VR app for 3D house rendering on Oculus Quest 2
  • Interactive object spawning, wall color modification, furniture placement
  • Research paper written, awaiting publication

DIY-Home AR App

IIITDM Kurnool

  • Android AR app for home interior design
  • Virtual furniture placement with texture customization
  • Flooring and wall painting options

Sentiment Analysis - Smart India Hackathon 2020

IIITDM Kurnool

  • GRU-based neural network for Amazon review classification
  • 95% accuracy on positive/negative sentiment
  • Top 3 idea at institute level

16-Bit Processor

IIITDM Kurnool - VLSI Course

  • 8 functionalities: arithmetic, logical operations, shifters
  • Wallace tree multiplication, non-restoring division
  • Recursive doubling for add/subtract