“AI may not dream yet, but I’m definitely dreaming up some pretty exciting projects in the field!”
I’m a graduate student at UC San Diego, passionate about building intelligent systems that can learn, reason, and collaborate with humans. My current interests include LLM reasoning, AI agents, and LLM systems, which I explore through research, coursework, and hands-on projects.
In my recent internship at Kognitos, I developed an agentic AI pipeline for SOP generation from desktop recordings, built a web app supporting multimodal inputs, and deployed a scalable PII masking service using AWS Lambda.
Previously at Samsung R&D, I worked on privacy-preserving ML models, edge-based LLM optimization, and real-time mobile inference, with solutions deployed across a wide range of devices.
From on-device intelligence to cloud-native agent systems, I’m excited about building AI that is useful, trustworthy, and grounded in real-world applications.
Feel free to explore my Projects and Research to learn more.
- Large Language Models (LLMs)
- Agentic AI
- Embodied AI
- Human-Machine Interaction
- Computer Vision
- Natural Language Processing (NLP)
- Reinforcement Learning (RL)
-
MS in Computer Science, 2024 - Present
University of California San Diego
-
B.Tech in Computer Science & Engineering, 2015 - 2019
Indian Institute of Technology Kanpur (IITK)
Experience
- Researched and prototyped an agentic AI pipeline for SOP generation from desktop recordings, integrating key-frame extraction (CV-based), OCR, and multi-agent reasoning; benchmarked GPT-4o, GPT-4.1, GPT-5, and Gemini-2.5.
- Built and deployed a lightweight multimodal web app (video, images, documents) for automated SOP and flow-diagram generation, containerized with Docker and deployed on AWS EC2; enabled internal testing across diverse inputs.
- Developed and productionized a PII masking solution using GLiNER model (supports 40+ entity types), evolving from a REST API on EC2 (PoC) to a scalable AWS Lambda service with Docker packaging and role-based access control.
- Led the development of an automated Python-based Android profiling tool to benchmark rendering uniformly across devices, identifying bottlenecks (scroll janks) and enhancing rendering performance through improvements in the Android framework source.
- Led the development of four edge-based personalization LLM solutions, deriving actionable insights to enhance user experience.
- Fine-tuned the FLAN-T5 LLM model for our specific use-case, optimizing performance & achieving a ∼92% evaluation score.
- Collaborated with engineers and testers to improve the quality of ∼15k samples for one of the said personalization solution.
- Designed an edge ML solution to analyze phone usage data and detect boredom with ∼80% accuracy, enhancing user experience. Built an Android app for real-time inference with under 50ms latency, showcasing the model’s effectiveness.
- Pioneered a Federated Learning (FL)-based solution to predict gender and demographic age, enhancing privacy for ∼10k users. Explored innovative distributed learning techniques and tested diverse FL algorithms across 20+ input and model configurations, laying the groundwork for future privacy-preserving AI advancements.
- Developed the Robot Camera Visualization Android app for real-time visualization of depth maps and 3D point-clouds from a ToF camera at 30 FPS, with gesture-based UI features for enhanced user interaction, tailored to stakeholder requirements.
- Engineered a privacy-preserving edge ML solution for Next App Recommendation, published in IEEE ICSC 2022, by designing a memory-efficient model (99% size reduction) to minimize FL bandwidth costs. Trained and deployed the model in Java using DL4J, integrating it on Android edge devices with a User Trial (UT) app for training and inference across 500+ devices.
- Developed Neural Network (NN) model to predict the current location of a user based on their recent locations and time of the day
- Developed a simulation environment for replicating which cell tower in a given area a user would be connected to while in transit
- Developed an ML classification model to predict which cell tower a user is most likely connected to at any time of the day
- Top‐1 & Top‐3 prediction accuracies for the final model were 85‐90% and 90‐95% respectively on the evaluation dataset
- Implemented CNN models in Python using TensorFlow for image classification, leveraging Google ML Engine APIs to accelerate training on Google Cloud. Deployed the model with TensorFlow Serving to expose REST APIs for generating model predictions.
Research Work
Projects
