Technical Skills: Python (numpy, pandas, scikit-learn, mlflow), Deep Learning Frameworks (PyTorch, TensorFlow), C/C++ , AWS, SQL , SnowFlake, Docker, Git, CI/CD(GitHub Actions),Shell programming.

Education

• Master in Data Science

  University of Colorado Boulder

• BE in Computer Science

  Rajalakshmi Engineering College

Work Experience

Machine Learning Engineer / Data Scientist @ WeDigiStudio

• Developed a Drowsiness detection model using the YOLOv8 model for Brakes India Private Limited Company
• Developed a two difference size model one with 25 epochs and 125 epochs and converted into TFLite format.

Computer Vision Intern @ TU Dortmund University

• Conducted data preprocessing and augmentation on diverse datasets
• Working on the data annotation process and writing comprehensive technical documentation

Projects

CleanMeet AI

GitHub Link

Clean Meet AI is a real-time content safety system for video conferencing. It continuously monitors video, audio, and text streams during online meetings to detect and prevent NSFW visuals, offensive speech, and toxic chat messages.
Built a real-time AI system to monitor video, audio, and text streams, reducing NSFW and toxic content exposure by 90% in online meetings.

Achieved 95%+ detection accuracy across multimodal content with <200ms latency, ensuring seamless and safe conferencing.

Integrated with Zoom, Google Meet, and Microsoft Teams, enabling scalability across 100% of major conferencing platforms.

Neural Machine Translation for Tamil

GitHub Link

Developed a model aiming to achieve a BLEU score of 41, surpassing current benchmarks for Tamil language translation models.
Optimized the model to achieve high performance with minimal processing power, reducing computational requirements by 20% compared to baseline models.
Processed and binarized the Bharath Parallel Corpus Collection (BPCC), containing 8GB of sentence pairs for Tamil language, for training.

Image Captioning:

GitHub Link

Built a model that annotates images with relevant captions, achieving a BLEU score of 0.5, indicating high relevance and accuracy.
Utilized a Convolutional Neural Network (CNN) to extract visual features such as color, shape, and texture from images.
Reduced training time by 30% through efficient preprocessing and model optimization techniques.

Satellite Image Segmentation:

GitHub Link

Designed a model to segment satellite images into classes like land, water, and roads, achieving a mean Intersection over Union (mIoU) score of 78%.
Combined U-Net with ViT (Vision Transformer) to enhance accuracy while requiring only 1,500~ labeled images, demonstrating effectiveness with low data.
Achieved a loss function value of 0.68, outperforming standard models on similar datasets by 15%.

WorkShop & Lectures

• Causality: The new science of an old question - GSP Seminar, Fall 2021
• Guest Lecture: Dimensionality Reduction - Big Data and Machine Learning for Scientific Discovery (PHYS 5336), Spring 2021
• Guest Lecture: Fourier and Wavelet Transforms - Scientific Computing (PHYS 5315), Fall 2020
• A Brief Introduction to Optimization - GSP Seminar, Fall 2019
• Weeks of Welcome Poster Competition - UTD, Fall 2019

• A Brief Introduction to Networks - GSP Seminar, Spring 2019

• Data Science YouTube

Publications

1. S Suresh Kumar, D Gayathri, R Shugavaneshwar, “Enhancing Deep Learning Models for Sentiment Analysis Integrating Texts and Emojis: A Comprehensive Survey,” IEEE, 2024.Link

• Machine Learning blog