Stanford University, Management Science & Engineering
Jun 2026 → presentResearch Collaborator with Prof. Ross Shachter
Whether large language models can autonomously execute decision analysis, and what breaks when they cannot.
Software engineer at Barclays, researching how language and vision models behave when you measure them carefully
I build production AI and developer-platform systems at Barclays in Pune, and work as a research collaborator with Stanford, Harvard Medical School / MGH and Carnegie Mellon on evaluation, medical image generation and multi-adapter generative models.
A thread runs through most of what I do: the measurement is usually the problem. At Stanford I found that the order of fields in a model's response changes its answer to a decision-analysis problem, turning four different numbers into one stable one. At CMU, extending a benchmark from three seeds to ten cut our own reported gain from 6.8% to 4.3%. Frontier vision-language models score 39–48% on mirror reasoning where humans score 94%. None of those are results you get without building the evaluation first.
Alongside that I lead a team at Harvard Medical School and MGH generating patient-specific 3D brain MRI volumes without paired data, and work on low-latency systems: a C++ matching engine running around 3M orders per second, and a reproducible study of the order-book data structures underneath it.
Positions
Harvard Medical School / Massachusetts General Hospital
Apr 2026 → presentResearch Collaborator with Dr. Xiangrui (Taylor) Zeng
Leading a team generating patient-specific 3D brain MRI volumes from unpaired data with diffusion models.
Carnegie Mellon University, Prof. Min Xu's Lab
Jan 2026 → presentResearch Collaborator
Benchmarking vision-language models, multi-LoRA composition, and verifiable reinforcement learning environments for software agents.
Selected work
MIRA: mirror reasoning in vision-language models CVPR Workshop 2026
When a clock reads 2:50 in a mirror, humans instantly infer 9:10. Vision-language models report “2:50.” We name this reflection blindness and build the first benchmark to diagnose it: 713 real images, 3,529 QA pairs, 10 scenarios. Across 11 frontier models accuracy sits at 39–48% against 94% for humans, and multi-turn dialogue makes most of them worse.
Response-field order changes the answer Stanford
Asking a model to state its answer before its reasoning produced $60, $70, $82.35 and $67.50 on the same problem. Reasoning-first produced a stable $76.00. A measurement artifact, not a capability gap, and one that affects how LLM reasoning gets evaluated generally. Moving the arithmetic into generated code solved all 128 of 128 cases that in-context reasoning failed.
SR-LoRA / CLAC: multi-LoRA composition ICLR 2027 submission
A 480-composition-per-method GPU benchmark on Stable Diffusion v1.5. The ten-seed run showed our text-alignment gain was 4.3%, not the 6.8% three seeds suggested, and CLAC raises identity consistency from 0.839 to 0.915 over the strongest baseline.
MVEB: Massive Video Embedding Benchmark EMNLP 2026
A 23-task benchmark for video embeddings distilled from a 184-task pool. Across 33 models no single approach dominates, and audio's contribution turns out to depend on how the labels were annotated: a six-point gap consistent across model families.
News
- Oct 2026: Selected Student Delegate, Google DeepMind APAC Research Symposium 2026, Bangalore.
- Jun 2026: Began collaborating with Prof. Ross Shachter at Stanford MS&E on decision analysis and LLMs.
- Apr 2026: Began leading the unpaired brain-MRI generation project at Harvard Medical School and MGH.
- 2026: MIRA accepted at a CVPR 2026 workshop; MVEB accepted at EMNLP 2026.
- Jan 2026: Joined Prof. Min Xu's lab at Carnegie Mellon as a research collaborator.
- 2026: Finalist at the Meta PyTorch OpenEnv Hackathon, top 100 teams from 31,000+ registrations.
- Jul 2025: Joined Barclays, Pune as a Software Engineer.