Abhinay Shankar Belde

Computer Scientist & AI Researcher

M.S. in Computer Science at Purdue University | Deep Learning, Vision, Graphics, NLP

About Me

Hi, I’m Abhinay. I’m a researcher at Purdue interested in computer vision, graphics, and NLP.

Right now, I spend most of my time on generative models and representation learning. I care about outputs that are not just high-quality, but also structured, controllable, and consistent. I’m especially drawn to problems where models must respect objects, boundaries, and geometry instead of treating everything as one texture.

My master’s thesis was in medical imaging generation. I built a hybrid pipeline that combined self-supervised representations with diffusion and GAN-based synthesis to work in low-data settings. That experience made me care a lot about practical constraints like limited data, noisy signals, and the gap between something that looks good and something you can trust.

Currently, I’m exploring segmentation-aware conditioning for generation, using signals from models like SAM and Mask2Former to build per-instance conditioning that improves local structure and boundaries. In parallel, I work on adversarial robustness for NLP, focusing on query-efficient black-box attacks and evaluation tools for real deployed classifiers.

I enjoy collaborating with people who run careful experiments, move fast, and stay honest about what works and what doesn’t. If you’re working on generative modeling, robust NLP, or anything adjacent, feel free to reach out.

Technical Skills

🧠

Deep Learning Frameworks

  • PyTorch
  • TensorFlow
  • Keras
📚

Libraries

  • NumPy
  • pandas
  • Matplotlib
  • scikit-learn
  • cuDNN

High-Performance Computing

  • GPU Computing (CUDA)
  • Parallel Processing
  • HPC Clusters
  • Bash Scripting
  • SLRUM
💻

Programming Languages

  • Python
  • Java
  • JavaScript
🌐

Web Technologies

  • React
  • Node.js
  • HTML/CSS
☁️

Cloud and DevOps

  • AWS (EC2, ECS, EFS, S3, CloudWatch, VPC)
  • Git
  • Docker

Current Research

Computer Vision

Ongoing: Beyond Representation Sampling: Segmentation-Aware Conditioning for Generative Models — Combining SAM and Mask2Former to build per-instance conditioning vectors for controllable, fully unsupervised image synthesis.

Early experiments show improved FID and gains in precision and recall. Experiments are ongoing toward an ICML 2026 submission.

Natural Language Processing (NLP)

Submitted: CONQUEST: An Efficient Attack for the Constrained Hard-label Setting — designing query-efficient adversarial attacks when only top-1 labels are available, with strict perturbation and budget constraints.

The method combines adaptive search over discrete token edits with constraint-aware scoring, significantly reducing query counts compared to existing hard-label baselines on text classification benchmarks. The work is currently under review.

Publications & Research

Overcoming Black-box Attack Inefficiency with Hybrid and Dynamic N-nary Algorithms

EMNLP 2025

Published

Novel approach to improve model robustness using score-based feedback for adversarial perturbations in NLP models.

NLPAdversarial AttacksModel Robustness

Addressing Data Scarcity in Medical Imaging: A Hybrid IJEPA + Stable Diffusion + GAN Approach

Master's Thesis

Defended

Hybrid pipeline combining IJEPA, Stable Diffusion, and GANs for realistic synthetic medical image generation to address data scarcity.

Computer VisionMedical ImagingGANsDiffusion Models

Beyond Representation Sampling: Segmentation-Aware Conditioning for Generative Models

Research Project (Ongoing)

Ongoing

Exploring segmentation-aware conditioning techniques to achieve complete unsupervised image generation.

Computer VisionGenerative ModelsUnsupervised LearningSegmentation

CONQUEST: An Efficient Attack for the Constrained Hard-label Setting

Research Project (Submitted)

Submitted

Developing novel algorithms for adversarial attacks in scenarios where model confidence scores are unavailable.

NLPAdversarial MLText Classification

Projects

Personality Prediction - 2.0

Jan 2024 - Apr 2024

Developed a platform using Django and React.js to provide high school students with MBTI personality predictions, leveraging NLP and machine learning models.

Key Highlights:

  • Implemented state-of-the-art transformer-based models to analyze textual data
  • Integrated MongoDB for scalable data storage
  • Optimized backend performance using Python and Django
DjangoReact.jsMongoDBNLPMachine Learning

Abstractive Text Summarization

Jan 2024 - Apr 2024

Built an advanced text summarization system using deep learning techniques to generate concise and coherent summaries from long-form content.

Key Highlights:

  • Developed CNN-based model for text processing
  • Leveraged TensorFlow, Keras, and scikit-learn
  • Demonstrated expertise in deep learning for NLP
TensorFlowKerasPythonNLPTransformers

Image Coloring

Sept 2023 - Dec 2023

Developed a CNN-based model using TensorFlow and Keras to predict and generate color details for grayscale images, enhancing their visual quality.

Key Highlights:

  • Built CNN architecture for image colorization
  • Used deep learning for image processing
  • Enhanced visual quality of grayscale images
TensorFlowKerasPythonComputer VisionCNN

Purdue Marketplace (Telescope)

Sept 2023 - Dec 2023

Created Purdue e-commerce platform utilizing React, Node.js, HTML, CSS for seamless user experience with full-stack functionality.

Key Highlights:

  • Full-stack e-commerce platform development
  • Managed GitHub repository for collaboration
  • Implemented robust functionality and intuitive interface
ReactNode.jsHTMLCSSJavaScript

Experience

Graduate Research Assistant

Purdue University, Indiana, Fort Wayne

Aug 2024 – Aug 2025
  • Master's Thesis: Investigating a hybrid IJEPA + Stable Diffusion + GAN pipeline to address data scarcity in medical imaging. Leveraging GPU-based HPC clusters for high-volume synthetic image generation.
  • NLP Research: Transitioning from score-based adversarial attacks (EMNLP) to hard-labeled black-box scenarios, developing novel N-nary attack algorithms without model feedback.
  • Hands-on experience with High-Performance Computing (HPC), training deep learning models on multi-node clusters.

Software Engineer

Infosys Limited, Bangalore, India

Sept 2021 – Aug 2023
  • Built Spring Boot REST APIs integrated with Finacle Script for LMS–VAM (menu validation, redirects, SQL CRUD).
  • Refactored modules for modularity and reusability, separating orchestration from Finacle business rules.
  • Containerized microservices (fnhttp-va, lm) with Docker for reproducible, versioned releases.
  • Coordinated rollouts with infrastructure across staging and production to minimize downtime.
  • Monitored and optimized containers on AWS ECS, EC2, EFS, S3, and CloudWatch to improve scalability and reliability.
  • Streamlined release cycles and service reliability through standardized containers and modular API architecture.

Machine Learning Engineer Intern

The People's Corp, Bangalore, India

Jan 2024 – Aug 2024
  • Architected an enterprise RAG platform using OpenAI GPT-4, LangChain, and Pinecone to enable semantic search across 100k+ documents for cross-functional knowledge management.
  • Implemented Cohere ReRank integration to improve retrieval precision, reducing irrelevant responses by ~40% and enhancing answer quality for business-critical queries.
  • Deployed containerized RAG services on Kubernetes with a Redis caching layer, supporting high-concurrency workloads and reducing query latency by ~60% for 500+ daily users.
  • Collaborated with product and operations teams to build an AI-powered knowledge hub, streamlining access to internal documentation and enabling interactive querying of policies and procedures.

Student & Assistant Teacher Intern

JSpiders Institute, Bangalore, India

Jan 2021 – Aug 2021
  • Trained and certified in Java Full-Stack Development, Core and Advanced Java (J2EE), SQL, and PL/SQL.

Education

Master's in Computer Science

Purdue University Fort Wayne

Aug 2023 – Aug 2025

Fort Wayne, Indiana

Focus: Deep Learning, High-Performance Computing, AI Research

Bachelor of Science in Mechanical Engineering

Amrita Vishwa Vidyapeetham University

Jul 2017 – May 2021

Bangalore, India

Focus: Engineering Fundamentals, Problem Solving

Certifications

Build Better Generative Adversarial Networks (GANs)

DeepLearning.AI

June 2024

Natural Language Processing with Classification and Vector Spaces

DeepLearning.AI

July 2024

Supervised Machine Learning: Regression and Classification

Stanford Online

July 2024

Convolutional Neural Networks

DeepLearning.AI

Feb 2024

Neural Networks and Deep Learning

DeepLearning.AI

Jan 2024

Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization

DeepLearning.AI

Jan 2024

Get in Touch

Interested in collaborating on research, discussing AI innovations, or exploring exciting opportunities? Let's connect!

Contact Information

Connect With Me