Darrin Bright

Integrated Master of Technology in Software Engineering
Vellore Institute of Technology, Vellore

Biography: My research interests lie at the intersection of vision and representation learning. I'm particularly drawn to problems where neural networks need to extract structured knowledge from visual data, bridge different data modalities or learn from limited supervision.

Currently, I'm exploring self-supervised learning for medical imaging, specifically vector quantization and non-contrastive methods that can learn meaningful representations from unlabeled data. I've worked on cross-modal learning, Explainable AI for Interpretability, and semantic segmentation with transformer architectures. What connects these projects is a focus on making deep learning more data-efficient and interpretable.

Core Interests:

3D Vision
Medical Imaging
Representation Learning

Education

Vellore Institute of Technology - Vellore, India

Integrated Master of Technology in Software Engineering

School of Computer Science Engineering and Information Systems (SCORE)

Courses: Artificial Intelligence, Machine Learning, Natural Language Processing, Data Mining, Computer Networks, Data Structures and Algorithms, Database Management Systems, Operating Systems, Python Programming, Object Oriented Programming.
CGPA: 8.83/10

September 2022 - Present

Publications

PortionNet: Distilling 3D Geometric Knowledge for Food Nutrition Estimation

Darrin Bright, Rakshith Raj, Kanchan Keisham

Under review, Conference on Vision and Intelligent Systems 2025
Explainable Transfer Learning Ensemble with Dynamic Weighting for Kidney Disease Classification in CT Images

Darrin Bright, Praveen Kumar

Under review, Biomedical Signal Processing and Control (Elsevier)

Research

PortionNet: Distilling 3D Geometric Knowledge for Food Nutrition Estimation

Advisor: Dr. Kanchan Keisham | Vellore Institute of Technology, Vellore | July 2025 - November 2025

Developed PortionNet, a novel cross-modal knowledge distillation framework that allows RGB models to learn 3D geometric features from point clouds during training, eliminating depth sensor requirements at inference.

Designed a dual-mode training strategy with a lightweight RGB-to-Geometry Adapter that learns to generate pseudo-3D features for accurate geometric reasoning from a standard RGB image.

Achieved state-of-the-art performance on MetaFood3D dataset with 17.43% volume MAPE and 15.36% energy MAPE. Cross-dataset evaluation on SimpleFood45 achieves state-of-the-art energy estimation (12.17% MAPE), demonstrating effective generalization.

Explainable Transfer Learning Ensemble with Dynamic Weighting for Kidney Disease Classification in CT Images

Advisor: Dr. Praveen Kumar | Vellore Institute of Technology, Vellore | July 2025 - November 2025

Designed a dynamic weighted ensemble combining MobileNetV2, ResNet-50, and EfficientNet-B1 for multi-class kidney disease classification on 12,446 CT images across four diagnostic categories.

Implemented an explainability framework using Grad-CAM and LIME with domain-specific augmentations to provide interpretable predictions for clinical decision support.

Achieved 99.14% test accuracy using focal loss with label smoothing, effectively addressing class imbalance.

Self-Supervised Medical Image Segmentation with Vector Quantization

Advisor: Dr. Deepak Mishra | Indian Institute of Technology, Jodhpur | May 2025 - July 2025

Developed a hierarchical vector quantization framework that addresses limited labeled data in medical imaging by learning discrete representations from unlabeled ISIC 2018 images and using them as structured regularizers during segmentation fine-tuning on the PH2 dataset.

Designed three-stage training pipeline integrating VQ-VAE, SimCLR contrastive learning and segmenta- tion fine-tuning, with multi-layer quantization capturing features across semantic scales.

Achieved 79.76% Dice score with +1.38% consistent improvement across multiple runs, demonstrating the effectiveness of discrete representations as regularizers for medical imaging with limited labeled data

Experience

Indian Institute of Technology, Jodhpur - Jodhpur, India

Research Intern

Developed multi-stage self-supervised learning framework with hierarchical Vector Quantization for medical image segmentation, achieving 79.76% Dice score on PH2 dataset.

Conducted 13-experiment ablation study on Pascal VOC 2012 dataset investigating vector quantization integration with Swin Transformer encoders, testing codebook configurations, fusion mechanisms and initialization strategies.

Achieved 74.61% Dice score with optimized vector quantization configuration on Pascal VOC 2012, demonstrating +7.34% improvement over baseline CNN and +3.91% over vanilla Swin Transformer, validating VQ effectiveness for semantic segmentation tasks

May 2025 - July 2025

Astute - Vellore, India

Project Manager & Technical Lead - Intern

Developed an AI-powered SEO automation platform using Gemini-Flash that integrates keyword research, content optimization and technical auditing using SERP API and DataforSEO for data extraction and analysis.

Implemented website structure analysis with BeautifulSoup for HTML parsing and extraction of elements, creating automated solutions for identifying on-page optimization opportunities and generating technical SEO assets including sitemaps and robots.txt files.

Generative AI & ML Engineer - Intern

Built an SEO-optimized blog generator using Gemini-Flash that analyzes market trends by extracting high-potential Keywords from Google Ads Keyword Planner and Google Trends data, evaluating factors including search volume, interest over time, growth rate, CPC, and demographics.

Developed Social Spark model, an automated system for video advertisements and poster generation, combining Stable Diffusion for poster design and Gemini for tagline creation with optimized tagline placement.

July 2024 - April 2025

Vicuna Kouture - Vellore, India

AI Specialist - Intern

Experimented with various text-to-image diffusion models to generate unique t-shirt designs, achieving optimal results with Stable Diffusion.

Performed extensive prompt engineering to refine and identify the most effective prompts for generating visually appealing outputs.

Collaborated with the back-end team to integrate the AI model using FASTAPI, ensuring smooth deployment.

July 2024 - August 2024

IEEE Robotics and Automation Society - Vellore, India

Chairperson

Managing a chapter of over 500 students, organizing hackathons, workshops, and tech talks to inspire learning in robotics and automation.

Mentored a team of juniors by conducting sessions on ML and Generative AI frameworks like LangChain to enhance their skills and knowledge in the field of AI and automation.

December 2024 - Present

Skillset

Notable Achievements

Winners of Project 2039 Hackathon Vellore Institute of Technology, 2024
Winners of Alphaforge Ideathon Vellore Institute of Technology, 2024
Runners up of Biomimicry Innovation Challenge Vellore Institute of Technology, 2024
Runners up of Startup Demo Day Vellore Institute of Technology, 2025
Second Runners up of HackWar Hackathon Vellore Institute of Technology, 2025

Contact me

Email ID: darriebright@gmail.com

My inbox is always open; glad to connect, discuss and collaborate. Could relate to work, volunteering or a casual talk!
Feel free to send me an email ; will get back ASAP!