In the Prostate-Net project, I had the opportunity to work closely with two expert radiologists to enhance the localization of prostate cancer using cutting-edge deep learning techniques. Our goal was to leverage SwinUNETR’s self-attention mechanism to improve the accuracy of cancer detection.

One of the key challenges we faced was working with our in-house prostate dataset. This involved meticulous data analysis, cleaning, and restructuring to ensure that the model could effectively interpret and utilize the data. We transformed the dataset into a format compatible with our deep learning model, laying the groundwork for a more robust training process.

By integrating SwinUNETR’s self-attention mechanism, we achieved a remarkable Dice score coefficient of 0.84. This represented a 15% improvement over previous methods and enabled an 18% enhancement in early diagnosis. The meticulous process of data preparation, combined with the advanced capabilities of our chosen model, led to a significant leap forward in the accuracy and efficiency of prostate cancer localization.

• Python
• S3 & EC2
• AWS
• Swin-UNETR Transformer
• Django REST Framework

For my final year capstone project, I embarked on an ambitious journey to revolutionize the fashion retail experience through augmented reality. With the support of a sponsorship from RhythmFlows Solutions Pvt. Ltd., I developed Mirr(AR), a cutting-edge real-time AR platform designed to render photorealistic clothing and jewelry on customers, enhancing the virtual shopping experience.

The project leveraged a robust tech stack, including C++, Python, React, Flask, MediaPipe, and OpenCV, to create an intuitive and immersive platform. My key challenge was addressing latency issues, which I overcame by implementing performance optimizations using C++ and CUDA, ensuring that the real-time experience was smooth and responsive.

Another significant hurdle was achieving accurate fittings for diverse body types. At the time, there were no existing algorithms that provided precise measurements for individual body parts. To tackle this, I developed custom algorithms combining mathematical and geometric techniques with Python coding. These algorithms enabled precise fitting of clothing and jewelry, accommodating various body shapes and sizes, and thus enhancing the overall user experience.

The result was a highly successful product, which not only met but exceeded expectations, leading to a sale of Rs. 1.2M. Mirr(AR) significantly reduced return rates by 35%, thanks to its accurate virtual try-on capabilities and seamless user interface. This project stands as a testament to the transformative power of augmented reality in retail and highlights my ability to deliver innovative, high-impact solutions.

• Python
• React
• C++ & CUDA
• OpenCV
• Flask REST Framework
• FIgma
• Firebase Auth

In the Real-Time Surveillance project, I tackled a significant challenge faced by a gear box manufacturing industry: the lengthy and labor-intensive quality control process for components on the assembly line. The goal was to streamline the inspection of numerous small components before final assembly, which traditionally took about 2.5 hours and required four quality control personnel.

To address this, I developed a scalable, real-time video analysis system using C++ and CUDA, paired with AWS Kinesis, Lambda, EC2 GPU instances, and cuDNN optimization. This system was designed to detect and track object anomalies with high precision and speed, achieving over 30 frames per second.

I trained multiple Convolutional Neural Networks (CNNs) to identify irregularities on the surfaces of 18 different components. Components that failed the automated inspection were discarded, while those passing the check were subject to final human evaluation. This approach not only reduced the quality control time from 2.5 hours to 1 hour and 50 minutes but also decreased the required human labor from four to two personnel.

I successfully deployed the system, integrating parallel processing and setting up a centralized server to manage and configure all cameras. The real-time output and efficient processing led to a 10% boost in overall profits by enhancing productivity and accuracy. This project exemplifies how advanced video analysis and real-time processing can significantly improve manufacturing efficiency and quality control.

• Python
• C++ & CUDA
• AWS Kinesis
• AWS Lambda
• AWS EC2 GPU Instances
• cuDNN
• Convolutional Neural Networks (CNNs)
• OpenCV
• Roblox

In the Translate.me project, I aimed to create a user-friendly multilingual application that could bridge language gaps by converting input audio into text and audio across over 200 languages. The challenge was to develop a seamless and efficient platform that would cater to a global audience, enabling real-time communication and translation.

To achieve this, I integrated Googletrans for language translation, gTTS (Google Text-to-Speech) for generating audio output, and Flask to build the backend framework. The application harnessed the power of the SpeechRecognition library for accurately converting spoken language into text. Additionally, I employed RecorderJS to facilitate smooth audio processing and export, ensuring that users could easily interact with the application in their preferred language.

Translate.me stands out for its versatility and ease of use, making it accessible to a wide range of users regardless of their linguistic background. The application not only converts spoken language into text but also provides audio output in the selected language, making it a powerful tool for real-time communication in a multilingual world.

• Python
• Googletrans
• gTTS (Google Text-to-Speech)
• Flask
• SpeechRecognition library
• RecorderJS

In today’s fast-paced business environment, adaptability is key, and that’s the challenge I tackled with E-commerce.in. I designed and developed a fully customizable e-commerce platform that empowers administrators with unprecedented control over their online stores. With the ability to modify the entire layout, manage product listings, and define payment gateways, this platform enables businesses to pivot and change their entire business model instantly.

The platform's architecture leverages Angular for the web application, enabling the creation of a highly responsive and modular front-end. This structure facilitates component-level customization and expansion without compromising overall system integrity. For mobile applications, I chose Flutter to ensure cross-platform compatibility and consistent performance across Android and iOS devices.

To handle the backend, I integrated Firebase as the primary database. Firebase's real-time capabilities and scalable infrastructure were pivotal in ensuring that the platform could manage high volumes of transactions and user interactions with minimal latency. Additionally, Firebase's authentication services provided a secure and seamless user management system, while its cloud functions enabled serverless operations, further enhancing the platform's scalability and reducing maintenance overhead.

The development methodology employed a microservices architecture, allowing for isolated feature development and deployment. This approach optimized code maintainability and facilitated continuous integration and delivery (CI/CD) processes, ensuring efficient update rollouts with minimal service disruptions.

E-commerce.in is a testament to my ability to design and implement scalable, high-performance systems using a modern tech stack, delivering a solution that meets the complex and evolving demands of the e-commerce industry.

• AngularJS (Web)
• Flutter (Android/iOS)
• Firebase

During the height of the COVID-19 pandemic, ensuring the safety of educational institutions became a top priority. To address this challenge, I developed the Covid-Mask Tracker, a system designed to detect face masks and monitor social distancing, providing automated access control to university buildings. The system was built around a Deep Convolutional Neural Network (CNN), specifically trained to identify whether individuals were wearing masks, with the goal of enhancing campus safety and compliance with health guidelines.

The system's core was a Deep Convolutional Neural Network (CNN), which I meticulously trained using TensorFlow and Keras. The CNN was designed to accurately distinguish between masked and unmasked faces in various conditions. OpenCV was integrated to handle video feeds, enabling real-time processing and preprocessing tasks like resizing, normalization, and face detection. The model was deployed on local servers to ensure low-latency operations, directly interfacing with the university’s automatic door systems to control access based on mask detection results. Additionally, the system included a social distancing monitor, utilizing OpenCV's object detection to measure the distance between individuals and ensure compliance with safety protocols.

Developing the Covid-Mask Tracker presented several challenges, particularly in ensuring the system's reliability in diverse environmental conditions, such as varying lighting and crowd density. To address these issues, I implemented adaptive thresholding and dynamic background subtraction techniques within OpenCV, enhancing the system’s robustness. Fine-tuning the CNN was another crucial step, involving extensive experimentation to reduce false positives and negatives, ensuring that the system could operate effectively in real-world scenarios.

The Covid-Mask Tracker ultimately became a vital tool in maintaining a safe campus environment during the pandemic. Monitoring over 400 individuals daily, it enforced mask-wearing and social distancing protocols, significantly reducing the risk of COVID-19 transmission. The project not only highlighted the power of AI and computer vision in crisis management but also underscored my ability to design and deploy critical safety systems effectively.

• Python
• OpenCV
• TensorFlow
• Keras
• Deep Convolutional Neural Networks (CNNs)