Developed a dual-agent QA system using fine-tuned Qwen2.5-70B and DeepSeek-R1-Distill-Qwen-32B models to generate synthetic datasets for blood relations and seating arrangements, achieving 80% questioning accuracy and 70% answer correctness. This project showcased expertise in leveraging advanced LLMs and synthetic data generation for complex reasoning tasks.

Developed a voice-powered personal assistant agent using LiveKit, integrating Google Calendar, task management, weather, and web search capabilities using MCP to intelligently manage schedules, tasks, and provide real-time information. This agent streamlines daily activities by automating meeting scheduling, task updates, email communication, and offering proactive recommendations based on user data and web insights.

A privacy-centric psychological counselor app using Liquid-AI LFM2.6B SLM  that runs entirely offline on mobile devices, ensuring data privacy and accessibility system.

MultiAgent system built using crewai that automatically monitor brand mentions, analyze sentiment, and generate market feedback reports from social media and web sources

This product assistant built scrapes product data from any online store, builds a vector database, and delivers users real-time, accurate product information.

A scalable document analysis chatbot built using LangChain, optimizing the RAG pipeline with DeepEval for enhanced contextual responses and multi-format file comparison.

An agentic system that generates Python code from natural language queries and executes it securely in a Docker sandbox.

A RAG-based agent that scores resumes against job descriptions, providing detailed section-wise feedback and actionable improvement suggestions.

Optimized and deployed a quantized MobileNet SSD model in INT8 TensorFlow Lite format for classifying stray animals on highways, achieving over 96% accuracy and real-time inference at 50-60 FPS on Raspberry Pi 3B+ with camera input.

FPNrNet, an attention-based model for reducing fixed pattern noise in CMOS sensors, was quantized and deployed on edge devices like Raspberry Pi and Jetson Nano, with performance and resource usage evaluated.

Deep learning image classification models were optimized using techniques such as quantization, pruning, layer fusion, and knowledge distillation. Standard toolkits including TensorRT, AIMET, and OpenVINO were leveraged for weight and activation-aware quantization from float32 to int8, with custom calibration caches generated to enhance post-training quantization and improve edge device inference performance.s

Object detection and tracking of Foreign Object Debris (FOD) on airport runways were achieved using a fine-tuned YOLOv10 model, with robustness ensured by fusing RGB and IR images from Intel RealSense thermal cameras for reliable day and night detection.A ground station image stitching pipeline was established to combine video streams from drone swarms, providing a unified panoramic visual interface.

A smart wearable for soldiers that monitors vital signs (heart rate, body temperature), detects falls, and sends SOS alerts to nearby soldiers and the base station via Zigbee.

The project developed a low-cost Braille printer using Arduino, stepper motors, and a servo motor to print Braille from English text onto paper. It provides an affordable, portable alternative to commercial printers, enhancing accessibility for the visually impaired.

Vehicular Ad-Hoc Network (VANET) concepts were experimentally demonstrated using remote-controlled cars equipped for Zigbee-based V2V and V2I communication. Ultrasonic sensor-enabled proximity sharing was employed for collision avoidance, and the environment was simulated with Ns3 and Sumo for validation.