Vocal Eyes - AI powered Smart Glasses for the visually Impaired

Vocal Eyes is an assistive vision system designed to give real-time environmental understanding to visually impaired users through smart glasses. The system integrates Vision Transformers (ViT) for high-accuracy scene recognition with Scene Graph Generation (SGG) to model spatial relationships between objects, enabling contextual, human-like reasoning about the environment rather than simple object detection.

A key contribution of the project is its hybrid edge cloud deployment strategy, which balances low latency inference on commodity hardware with scalable cloud processing making the system practical and cost-efficient without requiring expensive dedicated hardware.

Tech: Vision Transformers, Scene Graph Generation, Python, PyTorch, Edge-Cloud Architecture 

Football Match Analysis System ~Computer Vision Project 

An end-to-end football match analysis pipeline built to extract tactical and performance insights from broadcast footage. The system uses YOLOv8 for real-time player and ball detection, combined with multi-object tracking to maintain consistent identity across frames even through occlusions and fast movement.

Motion estimation techniques are applied to extract player trajectories, measure distances covered, and compute team-level possession metrics, all from a single broadcast camera feed without any specialized equipment.

Tech: YOLOv8, Multi-Object Tracking, Motion Estimation, Python, OpenCV 

Glaucoma Detection with Deep Learning ~Medical AI Project 

A deep learning model for early-stage glaucoma detection from retinal fundus images. The project fine-tuned EfficientNetB0 using transfer learning on a labeled retinal image dataset, achieving 86% classification accuracy, a clinically meaningful result for a low-cost screening tool.

The trained model was deployed publicly on Hugging Face Spaces, making it accessible as a web-based diagnostic aid without requiring any local setup.

Tech: EfficientNetB0, Transfer Learning, TensorFlow, Hugging Face Spaces, Python 

RISC-V RV32I CPU Implementation ~Digital Design Project

A ground-up hardware implementation of the RISC-V RV32I instruction set architecture in SystemVerilog. The project involved designing all core functional blocks of a Single-Cycle processor, including the ALU, register file, control unit, instruction memory, and data memory and verifying correctness through simulation.

The design was validated on Quartus Prime for synthesis and ModelSim for functional simulation, covering the full RV32I base integer instruction set.

Tech: SystemVerilog, RISC-V RV32I, Quartus Prime, ModelSim, Digital Logic Design  

Warehouse Automation Robot ROS System   ~Robotics Project 

A simulated warehouse automation system built in ROS and Gazebo, featuring two coordinated robots working in a continuous pick and place delivery loop.

A 3-DOF robotic manipulator handles object picking and placement across three joint configurations (PICK → TRANSFER → PLACE), controlled via ROS joint position controllers with tuned PID parameters. A differential-drive rover autonomously navigates a square delivery path, synchronized with the arm through a custom /object_placed ROS topic, ensuring collision free sequential operation across 27+ continuous delivery cycles.

Tech: ROS, Gazebo, Python, URDF/xacro, ROS Control, RViz