Soham Roy

Soham Roy

Researcher & Engineer

I research how to make generative AI safer
and build the systems to prove it.

CS undergrad at KIIT University, working on multimodality, machine unlearning, and copyright protection in generative AI. Part of RespAI Lab, published at ICCVW 2025.

Research & Publications

ICCVW 2025

Dynamic Inference-Time Copyright Shielding for AI Image Generation

Soham Roy, et al.

We protect copyrighted content in diffusion model outputs using dynamic inference-time interventions, so models can generate responsibly without needing to be retrained.

arXiv

Current Research Areas

01

Multimodality

Working on ways to evaluate and improve how well agentic systems handle multiple modalities.

02

Copyright Protection

Making it possible to block copyrighted content in AI-generated images at inference time, no retraining needed.

03

Responsible AI

Figuring out how to make generative AI systems safer and more ethical as they scale up.

Experience

AI & Backend Developer Intern

Virtual Foundry

Jun 2025 - Present
  • Built EVA, an automated hiring pipeline using FastAPI, MongoDB Atlas, and AWS
  • Built a scalable system that processes thousands of resumes, with 80+ REST API endpoints
  • Integrated VAPI.ai, Twilio, and OpenAI APIs with full logging and monitoring

Research Member

RespAI Lab

Jun 2024 - Present
  • Researching machine unlearning and copyright protection for diffusion models
  • Working on new algorithms for AI safety in generative models

NLP Research Intern

IIT Bhubaneswar

Dec 2024 - Jan 2025
  • Built data pipelines for fine-tuning Llama 3 on travel planning tasks
  • Set up preprocessing workflows that improved model performance and output quality

Lead: AI/ML Team

Microsoft Learn Student Ambassadors, KIIT

Jan 2024 - Present
  • Leading a team of 6, handling project planning and development
  • Built an open-source Stable Diffusion fine-tuning framework, grew to 10+ contributors

Selected Projects

GeoAgent

A framework on top of state-of-the-art LLMs for geographic reasoning and navigation. Uses LLM-as-judge for iterative refinement, LoRA fine-tuning with HuggingFace, and Stable Diffusion integration.

LLMsLoRAHuggingFaceStable Diffusion

EVA: Automated Hiring Pipeline

Full-stack app with a FastAPI backend and MongoDB (6 collections). Includes Google OAuth, JWT auth, Docker, AWS deployment, and AI-powered resume analysis with Google Gemini.

FastAPIMongoDBAWSGeminiDocker

Automatic Interview Evaluation

Evaluates interview responses using Flask, LangChain, and FAISS. Has real-time speech recognition, NLP processing, and ROUGE similarity scoring.

FlaskLangChainFAISSNLP

Stable Diffusion Fine-tuning Framework

Open-source training pipeline with LoRA optimization and solid docs. Community-driven with 10+ contributors.

LoRAStable DiffusionOpen SourcePython

Achievements

Amazon ML Summer School

Selected for Amazon's prestigious ML Summer School, 2025

AIR 82, Amazon ML Challenge

All India Rank 82 in Amazon ML Challenge 2025

Department Rank 1

Achieved dept. rank 1 with an SGPA of 10.00 in 5th semester, Computer Science & Engineering

HackNITR 5.0 Finalist

Finalist at NIT Rourkela's premier national hackathon

MLSA Hackathon Winner

First Place at Microsoft Learn Student Ambassadors Hackathon (2024 & 2025)

Education

Kalinga Institute of Industrial Technology (KIIT DU)

B.Tech in Computer Science & Engineering

GPA: 9.24 / 10.00 2023 - 2027

Let's Connect

Always up for talking about research, cool projects, or AI/ML opportunities. Drop me a line.

sohamroy.dev@gmail.com
GitHub LinkedIn