Research
Teaching a model to read two brain-tumour MRI scans months apart and say whether the tumour grew — RECAP-Net, ranked 3rd worldwide in a 2025 international challenge and presented at MICCAI, the main conference for AI in medical imaging.
Machine-learning researcher · medical brain imaging · New Delhi
Building ML pipelines& enjoying life through backpropagation
I'm an undergraduate researcher (NSUT and IIT Madras) who builds machine-learning models that read brain MRI scans — my MICCAI 2025 paper placed 3rd worldwide in an international brain-tumour challenge — and I build and ship the software around the research.
loading 0%
The problem
One scan says almost nothing.
A brain-tumour patient is scanned every few months. The question that matters — did the tumour grow, shrink, or hold? — needs two scans compared, and that comparison is slow and subjective by eye.
The work
RECAP-Net reads the pair.
My model takes two MRI scans of the same patient, months apart, outlines the tumour in each, and classifies the change — progressing, stable, or responding — the way radiologists do under RANO, the standard rulebook for judging treatment response.
The result
World Rank 3.
3rd of all teams worldwide in the BraTS Lighthouse 2025 Tumor Progression Challenge — an international competition where every team's model is scored on the same hidden MRI data. Presented as an oral talk at MICCAI 2025 in South Korea (the main medical-imaging AI conference) and published in Springer LNCS.
A paper is a piece of maths that other people will build on without re-deriving it — every product that uses it borrows it on trust. That's why I take research seriously. And it's why I ship what I build: because an idea only becomes real when it survives being used.
Three threads
Publish it. Ship it. Open-source it.
Research that gets peer-reviewed, engineering that gets deployed, and open source that other labs actually run. The same person does all three, and each one makes the others better.
Engineering
An internal assistant that helped up to 35,000 employees at a healthcare company find the approved way to use AI for their role; and PyAMorph, a library at IIT Madras that turns images and CAD models into geometry a physics simulation can run on.
Open source
Fifteen-plus merged pull requests to BrainGlobe, the open-source toolkit neuroscience labs use to map whole-brain microscopy images; a leading contributor to its image-alignment tool.
Experience
Jun – Aug 2026
Optum (UnitedHealth Group)
AI Engineer Intern — AI-DLC Pilot Team
- Built an internal assistant, in TypeScript on the company's agent framework, that helps up to 35,000 employees find the approved, governed way to use AI for their specific role — instead of searching policy documents.
- Built an automated judge for an internal hackathon: a model that scored, consistently and by fixed rules, how well each team followed AI-DLC (the company's development process, where every change carries a written plan and record), so the human judges had a reliable second opinion.
- Interviewed seven kinds of specialist — up to six people each — and turned what they told me into role-specific workflows, prompts, and guardrails the assistant could hand out.
Jan 2026 – present
IIT Madras
Research Intern — Scientific Computing (remote)
- Developing PyAMorph, a Python library that turns images, CAD models, and equations into geometry a physics simulation can run on. It works by computing, for every point in space, how far it is from the nearest surface (a signed distance function).
- Added shape combination (union, subtraction, intersection), conversion from standard 3D mesh files (STL), and GPU-accelerated bindings, all plugged into AMReX — the US Department of Energy's framework for large simulations on supercomputers.
- The work was selected for an oral presentation at INCAM 2026, India's national applied-mathematics conference, at IIT Kanpur; a journal paper is in preparation.
Jul – Aug 2026
Amazon
Amazon ML Summer School
- Selected for Amazon's competitive machine-learning summer programme for students.
Selected projects
All projects
Open to what's next.
Looking for software-engineering and research internships in 2026–27. If you work on medical imaging, scientific computing, or AI systems that have to be right, I'd like to hear from you.