Projects - Shounak Desai

Built a chatbot leveraging Retrieval-Augmented Generation (RAG) with GPT-3.5-turbo for providing soccer player insights, utilizing Sentence Transformers for player embeddings and FAISS for efficient similarity search
Designed and implemented a robust data pipeline using MongoDB to store player embeddings and attributes, and containerized the application using Docker for streamlined deployment.
Processed the EAFC 2025 dataset locally to create an interactive tool for exploring player statistics, enabling real-time responses through a Flask-based web interface.

Python Flask (RAG)Retrieval-Augmented Generation GPT-3.5-turbo Sentence Transformers FAISS MongoDB Docker

Pioneered a novel offside detection system utilizing a single camera, reducing the need for 10-30 traditional cameras by 90-97%, thereby significantly lowering hardware.
Leveraged pretrained model for precise Keypoint Estimation, employing vanishing point techniques to construct accurate 3D offside lines using Hough Lines in 3D space.
Segmented the play area to focus solely on the soccer pitch and players, excluding extraneous details like the audience, advertisement boards, and scoreboards from the input frames provided by the broadcaster.
Executed Line Detection on the soccer pitch for detecting parallel lines in 3D space using Hough Lines from Real-time input frames.

Computer Vision OpenCV Python PyTorch RCNN Hough Lines Image Processing Augmented Reality

Developed an end-to-end binary classification ML pipeline to automate manual compliance work in VoIS’s legacy tools for ticket management.
Implemented SMOTE analysis to tackle imbalanced data and crafted an NLP model utilizing various classic ML algorithms. Deployed the XGBoost NLP model on GCP Cloud Composer utilizing Apache Airflow and Google Cloud Storage delivering an end-to-end automation pipeline with an accuracy of 88%
Automated the work of 25-30 full-time employees at Vodafone Group, saving $350k-$800k in revenue.

Naturla Language Processing Scikit-Learn Python Google Cloud Platform Apache Airflow Machine Learning Shell scripts

Developed a Flask API that read input Business .pdf files and generated output .csv files with relevant information for Vodafone Clients.
Worked on Optical Character Recognition using Tesseract library, flask for the web app and used Docker to deploy on the SAP Cloud Platform, using Postgres database.
Clients were benefited from not reading multiple .pdf contracts since the required specific data was easily available in the output .csv files at their end. Successfully delivered this project for Spain, Germany, South Africa, Italy markets.

Optical Character Recognition Tesseract PostgreSQL Spark Docker SAP Cloud Platform

Crafted multiple event-based notification APIs for the Vodafone Business Clients during the shifting of client business files from SAP Cloud Platform to GCP.
Facilitated a seamless transition from SAP to GCP, ensuring uninterrupted client operations and minimizing downtime. Used GCP Pubsub, Cloud Run and Cloud Functions services for this automation project.

Google Cloud Platform Pubsub Cloud Run Cloud Functions Python APIs

Designed a real-time interactive game of Rock-Paper-Scissors to be played between a human and a machine.
Developed Python scripts to create custom labelled dataset with 4 labels at real-time using OpenCV and Python.
Worked at the intersection of Computer Vision and Deep Learning with highly efficient real-time results using a custom made Convolutional Neural Network (CNN) with an accuracy of 91%

Python Computer Vision TensorFlow Deep Learning OpenCV CNN

Created a Flask Web application that detects Human faces in user input Video files, recognizes and displays the Facial Expression type at any given instant in any video.
Leveraged OpenCV, custom trained CNN, Flask web framework, and Docker to containerize and make it able to deploy the web-app virtually on all Cloud platforms with an accuracy of 75%

Python TensorFlow CNN Flask OpenCV Docker

Developed custom trained Decision Trees and Adaboost models based on the Math and algorithm required.
Created a Flask website where user can input a sentences or paragraphs to determine if the input language is English or Dutch.
The models results is efficient with 98% accuracy

Python Decision Trees Boosting models Flask Machine Learning

Developed an advanced AutoEncoder Transformer for the purpose of Image Super-resolution.
Created custom model using TensorFlow.
Constructed for 256x256 input image sizes with an inference of 4.5 RMSE values

Image Processing Python TensorFlow Transformer Deep Learning

Designed a real-time interactive Handgesture Recognition model which recognizes and displays the number of open fingers on Hand.
Used Convexity Defects math for this Recognition with OpenCV implementation.

OpenCV Python Computer Vision Image Processing