Research

Current Projects

• Interactive Robotics Lab
  • Hypernetwork based Optimizers
    

    Developing new class of semantically enhanced Hypernetwork-based Optimizers for faster convergence in 10+ numerical and RL policy optimization tasks; Built LLM fine-tuning pipelines with Quantization, LoRA, and lightweight reasoning distillation from the Gemini LLM with chain-of-thought prompts.

Past Projects

• Interactive Robotics Lab
  • Prompted Policy Search - Project Link
    

    Co-developed Prompted Policy Search, an LLM based optimizer that incorporates linguistic reasoning for numerical and RL policy optimization; achieved SOTA results by outperforming PPO and TRPO on 8 of 15 benchmarks; presented at NeurIPS 2025 (Main Track).

  • Feedback-Optimized Robotic Policy Generation with Language Models
    

    Co-developed an LLM-based program synthesis framework for compliance-aware robotic manipulation without gradient updates, achieving 1.6 mm peg-in-hole accuracy with a UR5 robot arm; presented at Southwest Robotics Symposium 2025.

• Arizona State University
  • Self Supervised Dense Point Tracking in Turbulent Videos - Poster
    

    Implemented a test-time trained dense point tracking model robust to atmospheric turbulence by leveraging DINOv2 semantic features and RAFT optical flow model, leveraging the DINO-Tracker model (EEE 598 Class Research Project, 2024).

• Samsung Research Institute
  • Quick Learning, Active Intelligence on Bixby

    Built low-latency microservices on Samsung Bixby to identify target devices from user utterances within 700 ms by developing Flask APIs for sentence embeddings (benchmarked across 4 models), caching vectors in Redis for a 25% latency reduction, and implementing a time-decayed similarity algorithm to identify 3+ target devices (2021).

• Manipal Institute of Technology
  • Cross-Geography Generalization for Classification of Flooded Regions in Aerial Images - Paper
    

    Developed and implemented various neural network architectures for estimation and segmentation of flooded regions in aerial images captured using UAVs, while maintaining cross-geography generalization of data (CoRR, 2021).

  • Music Emotion Recognition Using Hybrid Spectral - Temporal Feature Vector
    

    Built a music emotion classification system based on Russell’s two-dimensional valence-arousal model using hybrid CNN and LSTM architectures; Extracted a hybrid spectral-temporal feature vector involving audio fingerprints for training the model (Final Year Research Project, 2020).

  • Multiclass Classification and Verification of Online Signatures
    

    Developed online signature verification system using time-series data to detect forgeries and classify owners, while improving efficiency through feature reduction with the Ramer-Douglas-Peucker sampling (Industrial Research Project, 2019).