Selected work

Projects

Towards Interpretable Models of Rumor Spread in LLM-Driven Agent Societies

  • Joint work with Pallavi Desai (UT Austin CS); course project for ECE 381K (Networks), December 2025.
  • Combined graph-theoretic analysis with Llama-3.1-8B agent simulations to model how rumors propagate, mutate, and are contained across synthetic networks (Erdős–Rényi, Barabási–Albert, Watts–Strogatz, SBM) and real Facebook/Twitter ego networks.
  • Extended SIR/SEIR diffusion with persona-driven decision-making (skeptical, neutral, conformist, sensationalist) and showed that strategically placed fact-checkers on hub nodes meaningfully suppress false-rumor cascades — true content reached ~10% higher saturation across all topologies.
  • Quantified information distortion via SBERT cosine similarity, GPT-2 perplexity, and Flesch–Kincaid scores; discovered a "Semantic Anchor" effect where clustered networks preserve rumor fidelity 2× better than random networks, and that mutations follow a directed evolution into distinct "generational dialects".

Reinforcement Learning for multi-node pricing and inventory management

  • Implemented and compared the performance of RL methods (A2C, DQN, PPO) against conventional Mixed-IP optimized pricing and (S, s) re-order policies for single/multi-echelon environments with stochastic demand.
  • Achieved a 25% increase in profit with dynamic pricing, and a 1.09× increase in profit compared to a traditional (S, s) order policy.

Multi-Modal Content Generation and Alignment with Efficient Optimization

  • Seamlessly merged a Stable Diffusion-based text-to-image model with AlignProp (a backpropagation-based approach to align diffusion models with downstream reward functions) and DreamGaussian (a 3D Gaussian Splatting model with mesh extraction and texture refinement).
  • Demonstrated DreamGaussian’s superiority in generating high-quality textured meshes — a 10× acceleration over existing methods, producing content from a single-view image in just 2 minutes.

A graph-based big-data optimization approach using HMM and CSP

  • Undertaken as a requirement for the PhD-level course Statistical Modeling 1 at UT Austin, supervised by Dr. Abhra Sarkar (Department of Statistics and Data Sciences, UT Austin).
  • Developed a framework that incorporates a Constraint Satisfaction Problem to reduce the state space of a Hidden Markov model, applied to financial time series. Used the model to predict closing prices of the Dow Jones Industrial Average.
  • Induced constraints by analyzing Twitter sentiment from Jan 20, 2017 — Jan 20, 2020. The approach reduced MAPE by 0.59%, increased accuracy to 90%, and reduced computational time by 0.02s versus a conventional model.

Deep Learning integrated with modeling a medical waste gasification-power production plant

  • Part of my undergraduate thesis, supervised by Dr. Dinesh Shankar Reddy (Associate Professor, NIT Andhra Pradesh).
  • Conducted literature surveys on the sources, types, and effects of medical waste, and on properties associated with characterizing medical waste.
  • Developed a Deep Learning model in PyTorch to simulate thermodynamic equilibrium modeling of a downdraft gasifier for the conversion of medical biomass to energy.