1. Multi Agent Financial Assistant

   GitHub

   The Financial Assistant is a multi-agent AI system designed to assist users with comprehensive financial inquiries. From generic conceptual discussions and continued conversations to highly specific price-action analysis (e.g., "What happened to NVDA in the last couple of days?"), the platform provides data-grounded and concise insights. Built on a modular architecture powered by LangGraph and FastAPI, the platform intelligently combines specialized autonomous agents to process technical, fundamental, and sentiment-based signals. The system is designed for multi-turn interactions, allowing users to follow up on complex findings with further questions.

2. Domain-Specific Code Generation System (Ongoing)

   Developing an LLM harness for domain-specific code generation in clinical network studies by integrating Model Context Protocol (MCP)-based tool interfaces with retrieval pipelines over technical documentation, sample code, and domain-specific resources. The system aims to improve the relevance and correctness of generated analytical code by grounding LLM outputs in framework-specific guidance and study constraints.

   Key Contributions:

   • Designed interfaces to provide domain-specific context, including structured prompts, technical documentation, study templates, and sample implementations to support LLM-assisted code generation.
   • Enabled natural language-to-R code generation over standardized clinical datasets, improving accessibility of analytical workflow creation for technical and non-technical users.
   • Exploring validation layers (upcoming) to improve structural consistency and correctness of generated study specifications within predefined analytical constraints.

   Key Technologies: LLMs, Model Context Protocol(MCP), Python, LangChain/LangGraph

3. SmrtFridge: IoT-based, User Interaction-driven Food Item & Quantity Sensing

   GitHub Presentation Video Full Paper

   SmrtFridge is a smart fridge prototype that uses thermal and RGB sensing, combined with deep learning, to identify food items and estimate remaining quantity—without the need for RFID tags or user labeling. It leverages natural user interactions (e.g., door openings) to trigger video capture and processing. The system, triggered by natural interactions, achieves ~85% accuracy in classification and ~75% accuracy in quantity estimation. Its low-intrusion design makes it promising for real-world adoption in smart kitchens.

4. CollabCam: Deep Learning based System for Energy-Efficient Pervasive Vision

   GitHub Full Paper

   CollabCam is an edge-based vision system enabling collaborative inference in multi‑camera deployments. Each camera sends mixed-resolution frames, downsampled in shared FoVs, and collaboratively uses peer camera bounding boxes to boost detection. It cuts down network and energy usage by 25‑35%, while keeping accuracy losses under 2‑5% when compared to non‑collaborative approaches.

5. Video Analytics System for Natural Language-Driven Surveillance

   This project designs a modular, AI-powered video analytics system capable of responding to natural language queries. It combines object detection (YOLO, GroundingDINO), tracking (DeepSORT), action recognition (SlowFast, TimeSformer), and fine-tuned VideoLLMs. A custom query agent translates user queries into subtasks and invokes models for real-time insights. The system targets industrial scenes (e.g., ports), addressing challenges like occlusion, poor lighting, and ambiguous motion patterns.