cv

Graduate Research Assistant

Oregon State University

Machine Learning Research: Deep Learning, AI Planning, Agentic AI, Reinforcement Learning, LLM Agents

• Developed multi-agent LLM framework for PDDL planning with verifier-guided training, reducing LLM calls by 40%
• Created GNN-based POMDP framework (GammaZero) achieving 20x computational efficiency with zero-shot transfer
• Designed GNN action ranking approach outperforming LLMs (GPT-O3, Gemini-2.5-Pro) by 20-30x in planning coverage
• Led algorithm development and mentored team of 5 on AI2Thor embodied agent project (91% success rate)

Software Engineer

Cisco

Enhanced security for IOS XR routers and developed NLP-enhanced interfaces

• Enhanced security for IOS XR routers supporting global internet infrastructure (C++)
• Reduced user onboarding time by 15% with NLP-enhanced command line interface
• Leveraged reinforcement learning for continuous model improvement through user feedback

PhD

Oregon State University

Computer Science and Artificial Intelligence

Integrated M.Tech

IIIT Bangalore

Computer Science

Planning with Affordances: Integrating Learned Affordance Models and Symbolic Planning

arXiv 2025

Novel approach integrating PDDLStream with learned affordance models for task and motion planning

GammaZero: Learning to Guide Belief-Space Search for Long-Horizon POMDPs

arXiv 2025

GNN framework for POMDP planning with 20x efficiency gain and zero-shot transfer to 6x larger scenarios

Hierarchical Object-Oriented POMDP Planning for Object Rearrangement

NeurIPS SpaVLE 2025

Hierarchical framework achieving 71% task completion, scaling to 20 objects across 4 rooms

Graph Neural Network Based Action Ranking for Planning

NeurIPS 2025

GNN approach for learning relational policies achieving 89-100% success rates with 13x better coverage than baselines

Learning Agents with Prioritization and Parameter Noise in Continuous State and Action Space

ISNN 2019

Prioritized DDPG algorithm outperforming DDPG on 80% of MuJoCo benchmarks with 40% faster convergence

Ag Aid Hackathon Winner - Water Challenge

Oregon State University

ML solution for precipitation forecasting with 75% temporal and 84% spatial accuracy

Dean's Merit List

IIIT Bangalore

Second Runner-Up - Hack.Bangalore

Daimler

Idea presented at Mobile World Congress (MWC), Barcelona 2017

Second Runner-Up - Tech Maestro

NetApp

All India tech quiz

Winner - DreamWorks Hackathon

IIITB

Verifier-Guided Multi-Agent Orchestrator for LLM Planning

Multi-agent LLM framework for PDDL planning using compiler feedback as dense training signal

• Developed Contextual Action Ranking (supervised) and RLVR + PPO training approaches
• Reduced total LLM calls by 40% through improved routing
• Trained Llama-3-8B orchestrator matching prompted GPT-4 at significantly lower inference cost
• Applicable to any compiler-equipped domain (Python, C++, SQL, PDDL)

GlobXAI: LLM-Based Natural Language Interface for Explainable AI

Natural language querying system over global explanations of image classifiers

• Fine-tuned Gemma-2 9B using QLoRA achieving 98.7% parse rate and 96.2% intent classification
• LLM parses queries while Pandas executes deterministically - hallucination-free outputs
• Created synthetic dataset pipeline producing 50K+ training examples

GammaZero: GNN Framework for Scalable POMDP Planning

Graph neural network framework achieving dramatic efficiency gains in POMDP planning

• 20x computational efficiency gain with inference in seconds vs minutes
• Zero-shot transfer to scenarios 6x larger than training
• Uncertainty-aware graph representation for partial observability
• Integrated with MCTS for drastically reduced search

GNN-Based Action Ranking for Classical Planning

Learning-to-rank approach for guiding classical planners with GNNs

• 89-100% success rates with strong out-of-distribution generalization
• 75% less training data than value-function methods
• 13x better coverage on hard problems, generalizes to 8x larger test problems
• Outperforms LLMs (GPT-O3, Gemini-2.5-Pro) by 20-30x in coverage

Hierarchical Object-Oriented POMDP for Object Rearrangement

Hierarchical planning for multi-object rearrangement in partially observable environments

• 71% task completion in partially observable multi-object environments
• 2-3x improvement over baselines through hierarchical decomposition
• Scales to 20 objects across 4 rooms (baselines: 5 objects, 1 room)
• Handles blocked paths, object dependencies, goal conflicts

Intelligent Embodied Agent for Object Retrieval

Navigation system for object retrieval in interactive AI2Thor environments

• 91% success rate across 5000 test scenarios
• Handles container opening, tool usage, obstacle navigation
• Led algorithm development while mentoring team of 5

Denoising Diffusion for Motion Planning

Using diffusion models for collision-free path generation

• 80% generation accuracy in complex 2D environments
• 3x speedup over RRT for paths longer than 50 steps

XGBoost Precipitation Forecasting (Hackathon Winner)

ML solution for agricultural precipitation prediction

• 75% temporal and 84% spatial prediction accuracy
• Used Dynamic Time Warping, spatial aggregation, correlation analysis

Prioritized DDPG with Parameter Space Noise

Enhanced reinforcement learning for continuous control

• Outperformed DDPG on 80% of MuJoCo benchmarks
• 40% faster convergence using parameter space noise exploration