A bit about myself and my research

I started my undergrad in mathematics and statistics, with an attempt at finance to see whether I would like it. People always say how much easier it would be to succeed there if you were good at math, so I tried. A couple of courses in business writing and management were enough. It ran on a set of social conventions I never quite picked up, and a few rounds of them would use up my day.

I have always enjoyed mathematics. Before college I was convinced I would be a math teacher someday, until I realized it would mean either a degree in education or, for academia, deep roots in pure math. Either way a decision made very early, which felt like a one-way road with no turns. So I stayed with mathematics without deciding what it was for, enjoying the theorems and the question of what came next. Certain subjects came to feel deeply intuitive to me, like analysis and abstract algebra, where the structure is the whole of the object. Statistics, interestingly, did not, even when I did understand a concept. Maybe because it begins with an observed world and asks what structure can be found in it, a bottom-up reversal of the way I was used to thinking. I ended up with a degree in statistics anyway, and it has been useful ever since. Intuition and competence turned out to be separable.

Later I added a CS major, nudged by the SDE hype around 2021, even though deep down I knew I was no geek material. I struggled with some systems classes, especially next to peers who had been writing code for years. I missed my “advantage” and my simple pleasure in mathematics, and kept going back to take more math courses and CS theory courses, because they felt so familiar and safe to me. Things changed after I learned about machine learning. It ran on the mathematics I liked, and it made that mathematics the main instrument rather than a prerequisite. I guess that was when I first found a place for myself in CS.

In the meantime I jumped around looking for research. Survival analysis for cancer. Statistical modeling for nanoparticle drug design and binding dynamics. Single-cell ML model training. Unsupervised learning algorithm development. At some point I also jumped into learning theory to see whether there was a spot for me. At the end of the day, not much. I came out with a long CV of experiences that had nothing to do with each other.

For my master’s I picked Columbia for the theory cohort and ended up in a genome and protein language model lab. I was exposed to more frontier AI4S research and genuinely found something I enjoy: understanding model behavior and explainability, especially in a biological or medical setting, where the ground truth is so much more complicated than in language or images.

And eventually, maybe life just took me here, I am starting a PhD in CS with applications to cancer research — biology and computer science, the two subjects I swore years ago I would never do. It’s funny how life turns out, because after the big circle I am back working on cancer, where my research journey started.

For years I would look at my CV and feel sad that I couldn’t find a way to explain my path. Now my scattered experience has finally found its unifying place in multimodal AI for cancer research: survival, genomics, single cell, spatial proteomics and transcriptomics, protein, explainability, even a bit of theory. It is also astonishing how much the field has changed. Back at the start I was pulling expression matrices out of public databases and running regressions to validate a single marker; now people are reading slides with foundation models and designing proteins to shorten the path to a therapy.

My research is now moving toward the thing that unsettles me most: that I cannot find much structure in biology, and I have never been good at remembering unconnected dots. And now we are trying to explain it with math, elegantly.

In practice that means asking a few concrete things. Whether there is real signal in the embeddings of a genome or a tumor slice, and what that signal means. Whether cancer can be deciphered from the feature space of a pathology model. How far the numbers we report can be trusted, and how inevitable leakage is in biology foundation models. And looking further out, how much AI can reshape treatment. It is a privilege to spend my days on this, knowing it sits close to something that matters and carries the potential to help people facing disease.