Research

I work on evaluation and generation: building the measurements that show what models actually do, and the generative systems that have to survive those measurements. Most of it happens through collaborations at Stanford, Harvard Medical School and Carnegie Mellon.

Stanford University, Management Science & Engineering

Research Collaborator with Prof. Ross Shachter · Jun 2026 → present

Decision analysis is an unusually clean domain in which to test whether language models can act autonomously: the influence-diagram representation is formal, the pedagogy is well established, and the correct answer to every question is known exactly, in closed form, before any model is queried.

The measurement artifact. I designed experiments testing whether LLMs can execute decision analysis, and found that the order of fields in the response format alone changes the answer. Asking for the numeric answer before the free-text reasoning produced $60, $70, $82.35 and $67.50 across conditions; asking for reasoning first produced a stable $76.00. This is not a capability gap but a property of how the output is elicited, and it affects how LLM reasoning is evaluated well beyond this domain.

Where the arithmetic should live. I built a code-generation setup in which models solved decision problems exactly, 128 of 128, in cases where in-context reasoning failed outright. Error feedback lifts GPT-4o from 10.2 to 23.0 out of 29 tests, a real improvement that never reaches perfection. The architectural conclusion: models should supply genuinely uncertain judgments, and code should perform any numerically sensitive computation.

Harvard Medical School / Massachusetts General Hospital

Research Collaborator with Dr. Xiangrui (Taylor) Zeng · Apr 2026 → present

I lead a team generating patient-specific 3D brain MRI volumes without paired training data, using diffusion models in PyTorch.

Setting the direction. I hypothesised that gradient guidance would fail precisely where adapter conditioning succeeds, and designed an oracle experiment to test it rather than argue it.

Chasing the blur. Residual blur in the generated volumes traced back to conditional-mean averaging over plausible anatomies. I localised it to high-frequency wavelet subbands and designed a wavelet-domain loss that measurably improved sharpness. I also built the ADNI data pipeline and the nnUNet-based segmentation evaluation pipeline, and generated all second-phase results.

Carnegie Mellon University, Prof. Min Xu's Lab

Research Collaborator · Jan 2026 → present

MIRA (advisor: Yunbei Zhang). Vision-language models systematically hallucinate the canonical orientation of mirrored content instead of reasoning about the reflection. The benchmark covers 713 real images and 3,529 question-answer pairs across 10 scenarios, from licence plates in rear-view mirrors to analog clocks, chemical formulas and ambulance mirror-text. I evaluated 11 frontier models: all land at 39–48% against 94% for humans, and multi-turn dialogue degrades most of them rather than helping. Accepted at a CVPR 2026 workshop.

SR-LoRA / CLAC (with Xingjian Li). I built and ran the 480-compositions-per-method GPU benchmark on Stable Diffusion v1.5 (rank-4 LoRAs, DPM-Solver++, CLIP ViT-L/14) along with the τ/α/δ and N=3 ablations. Extending from three seeds to ten showed our own text-alignment gain was 4.3%, not the 6.8% the smaller run reported. CLAC raises identity consistency from 0.839 to 0.915, 7.6 points over the strongest baseline. Submitted to ICLR 2027.

TRACE and APISHIFT. Verifiable reinforcement learning environments for network troubleshooting agents and API contract migration. I contributed to environment design, verification logic, and reward mechanisms that score an agent's actions against ground truth rather than relying on a judge model: the distinction that makes the reward signal trustworthy enough to train against.

KL University

Undergraduate Research Assistant with Dr. K. Ashesh · Aug 2022 → Mar 2025

First-authored an IEEE CONIT 2024 paper applying outlier analysis to credit card fraud detection, using statistical and machine learning methods over transaction amount, location, time and user behaviour to raise detection accuracy while holding down false positives.

Low-latency systems

Outside the collaborations, I work on execution systems: a C++ limit order book and matching engine running at roughly 3M orders per second, and a reproducible, cross-platform study of the order-book data structures underneath it.