Researcher & Engineer
Fullstack AI Engineer
Computer science undergraduate at KIIT University, currently in my pre-final year. At present, I'm a SWE Intern at Cisco. I work on agentic AI systems, the benchmarks used to evaluate them, and applied machine learning more generally. I'm a member of RespAI Lab, where my work on copyright protection for diffusion models was accepted at ICCVW 2025. My interests may extend to anything else that contributes to the domain of Human Knowledge.
A method for keeping copyrighted material out of diffusion-model outputs at the point of generation, rather than by retraining the model. The work intervenes during sampling through adaptive guidance, so an already-trained model can be steered away from protected content without the cost of fine-tuning or a fresh training run.
A benchmark built around a single question: when an autonomous web agent is carrying out a task and lands on a website designed to manipulate it, will the agent give up the user's personal information? It reports two quantities together: how much private data actually leaks, and how reliably the agent recognises that it is under attack. The evaluation covers 91 adversarial environments, each paired with a benign control, organised under an eight-part taxonomy of manipulation tactics and tested on four current frontier models. The protocol was pre-registered before any data was collected.
GeoAgent introduces a benchmark for geolocalization in vision-language models based on embodied navigation rather than static image analysis. Instead of predicting a location from a fixed set of images, an agent actively explores Google Street View, moving and rotating before committing to a guess. The benchmark spans 1,200 navigable locations across 100 cities worldwide, covering both developed and developing regions and a range of recognisability. Across frontier VLMs, active exploration generally improves accuracy over static multi-view baselines, with Gemini 3.0 Flash and GPT-5 Mini performing best, though models rarely recover from a wrong initial hypothesis and tend to reinforce early mistakes. The results also reveal a marked geographic bias, with most models doing considerably better in developed regions than developing ones. The paper argues that embodied navigation exposes weaknesses in spatial reasoning that static geolocation benchmarks miss, positioning GeoAgent as a more realistic test of agentic geospatial intelligence.
AI systems that take actions in an environment, such as browsing, navigating, or calling tools, rather than only returning text. My interest is in how they behave when conditions are adversarial or simply unlike their training data.
Designing evaluations that say something honest about a model: controlled comparisons against a baseline, a clear taxonomy of how a system fails, and human reference points, instead of a single headline number.
Grounding a model's answers in retrieved source material, so that what it says can be traced back to real documents rather than produced from its parameters alone.
Cisco · Bangalore
Virtual Foundry
RespAI Lab
IIT Bhubaneswar
Microsoft Learn Student Ambassadors, KIIT
Selected for Amazon ML Summer School, 2025.
All India Rank 82 in the Amazon ML Challenge, 2025.
Department rank 1 with an SGPA of 10.00 in the fifth semester of Computer Science & Engineering.
Finalist at HackNITR 5.0, the national hackathon hosted by NIT Rourkela.
First place at the Microsoft Learn Student Ambassadors hackathon, in both 2024 and 2025.
B.Tech in Computer Science & Engineering
I'm glad to talk about research, the evaluation of AI systems, or relevant roles and collaborations. Email is the most reliable way to reach me.
sohamroy.dev@gmail.com