Contact
Get in Touch
Happy to talk about research, collaboration, or open roles in industry or academia.
I'm a machine-learning researcher working on vision-language models, LLM reasoning, and AI safety, and on the deeper question of how data and training shape what neural networks become.
I came to AI through an unusual range of fields: a PhD in computational biophysics, a stint designing algorithmic trading strategies, and years designing computer systems for accelerating Machine Learning and Quantum Computing before turning to the science of modern AI. Looking from all those angles yielded my working premise: AI models are the first systems we engineer top-down rather than from the ground up, so we must study them the way we study natural systems — investigating when and how they fail, then learning to control them to build reliable, purposeful AI.
What lies at the heart of the challenges for the wide-spread adoption of, and the safety concerns about AI models is that they are the first engineered systems built top-down rather than from the ground up. That makes it hard to understand their failure mechanisms and to establish their reliability.
I believe it is therefore essential to study these models just as we study natural systems, and to learn to control them. Neural networks learn to approximate target functions presented to them through a training dataset and a loss function, in effect reverse-engineering a target task from data. I am interested in understanding how, and which properties of the data and training procedure shape that approximation, so that we can guide and control it.
My research in recent years has focused on dataset-selection methods for fine-tuning multimodal models, improving the quality of task-specific synthetic images, the safety and risks of reward models for generative systems, and reinforcement-learning methods for reasoning models. Recent work includes in-context learning methods that help vision-language models resist misleading training priors, reward designs that strengthen multi-step reasoning, and data-selection recipes that reach strong performance with a fraction of the usual training data.
My earlier work in computer architecture, quantum-computing systems, and high-frequency trading gives me a systems-level instinct for these questions: how computation, data, and incentives interact to produce behavior, whether in silicon, in markets, or in a trained model.
My career has moved through fields that rarely share a hallway — biophysics, quantitative finance, computer architecture, quantum systems, and AI. The map on the home page is the interactive version of this, each field shaded by how recently I worked in it.
Mentor, AI4All Summer Program, introducing high-school students from underrepresented groups to AI. Preceptor for COS 126, Princeton's largest intro CS course. Graduate teaching in biochemistry and fixed-income finance.
PI on 4 NSF-funded projects · Roche Research Foundation Postdoctoral Fellowship · Boehringer Ingelheim PhD Scholarship · National Math Olympiad Training Team (top 20 of 1M+ students) · TÜBİTAK Undergraduate Fellowship
A selection across AI, systems, quantum, biophysics, and psychology. Full list on Google Scholar.
Happy to talk about research, collaboration, or open roles in industry or academia.
