Statement of Purpose

Arman Tavakoli — PhD in Computer Science

At 1:40 a.m. on the busiest sale night of 2024, the ranking model I owned at Digikala began to fail quietly. Offline metrics had promised a strong release; online, conversion in two provinces dropped nine percent within an hour. The cause was mundane — a holiday-driven shift in query distribution the model had never seen — but the lesson was not: a learned system is only as trustworthy as its behavior on the data we failed to anticipate. I have spent the three years since working on that gap, first as an engineer patching it in production and then as a researcher trying to characterize it. I am applying to your PhD program to make distributional robustness the center of my work.

My preparation began at Sharif University of Technology, where I completed both my B.Sc. and M.Sc. in computer engineering, finishing my master's in the top five percent of my cohort. My thesis examined out-of-distribution generalization in session-based recommenders. The finding that shaped my research taste was negative: two widely cited invariance-based methods, tuned carefully, failed to beat empirical risk minimization under the covariate shifts we actually observed in e-commerce logs. Rather than bury that result, I built a public benchmark of realistic temporal shifts so the comparison could be reproduced, and presented the work at a national AI conference; an extended version is available as a preprint. The exercise taught me that robustness research fails most often at the evaluation stage, before any algorithm gets a fair hearing.

Industry sharpened the question. At Digikala I designed the drift-monitoring service that now guards our ranking stack, and I ran a group distributionally robust optimization experiment that lifted worst-region NDCG by four points at a cost of half a point on average. That trade-off — who pays for average-case performance, and how much — is exactly the kind of question I could pose but not fully answer with production constraints and quarterly deadlines. Answering it requires the time horizon and the intellectual community of a PhD.

Your department is my first choice for concrete reasons. The Trustworthy Machine Learning Lab's recent line of work on certified robustness under distribution shift addresses the failure mode I met in production, and its emphasis on conformal methods offers a language for the guarantees practitioners actually need. I am equally drawn to the Systems for Machine Learning group, because I believe robustness is partly a systems property: monitoring, retraining cadence, and rollback policy shape real-world reliability as much as the loss function does. A program where those two communities share a hallway — and where the qualifying coursework spans statistical learning theory and large-scale systems — fits the shape of the problems I want to work on. The department's pattern of publishing benchmark and evaluation papers, not only methods, tells me my research taste would be at home.

During the PhD, I want to pursue three connected threads: evaluation protocols that predict deployment-time degradation from offline data; adaptive methods that trade average accuracy for worst-group guarantees in a controllable way; and open tooling that lets industrial teams adopt both. My master's benchmark and my production experience give me a running start on the first and third; the second is where I most need to grow, and where the coursework and community I would find in your program matter most.

After the PhD, I intend to work as a research scientist at the boundary of machine learning and reliability engineering — ideally shipping robustness tooling that becomes as standard as unit tests are today — while keeping an open door to academia. Five years of watching models meet the real world has convinced me the field needs people who take deployment failures as seriously as leaderboard gains. I would like to become one of them, and I believe your program is the right place to do it.