ABAli Backour
Ali Backour is a Syrian mathematician and ML researcher at MIT, studying computer science and mathematics (B.S. 2022-2026, MEng 2025-2027). Born and schooled in Syria, he represented his country at the 61st International Mathematical Olympiad (IMO 2020), scoring 28 points and earning a Special award, and returned for IMO 2021. At MIT he has done research at CSAIL (geometry of generative models, circuit lower bounds with Prof. Ryan Williams, the ATL tensor language), the MIT Operations Research Center, and spent time as a software engineer at Genesis Therapeutics; in 2026 he joined Prelude Labs as a Machine Learning Researcher. His published work spans diffusion-model geometry, turbulence super-resolution, and cryptography/ML security.
- By 2026 he holds two concurrent roles: ML Researcher at Prelude Labs (Brooklyn, NY) and Research Assistant at MIT CSAIL, on top of a full-time MEng - a research-plus-startup load typical of a strong MIT ML student.2026
- Represented Syria at the 61st IMO (2020), scoring 28/42 (rank 86, 86.2%) with a Special award 'S'; returned for IMO 2021. One of only six Syrians on the 2020 team.Sep 2020
- Matriculated at MIT from Syria (2022), an unusual path for a Syrian olympiad student; congratulated publicly by Syrian Youth Empowerment alongside fellow Syrian admittee Subhi.Mar 2022
- Spent 2025 as a software engineer at Genesis Therapeutics (AI drug discovery), building a Kubernetes job-retry decision engine that raised pod-failure observability from 25% to 100% - unusual production-infra depth for a mathematician.2025
- Research spans an unusually wide band - diffusion-model geometry, turbulence super-resolution, circuit lower bounds, and cryptography/ML security - suggesting a generalist theorist rather than a single-subfield specialist.2025
- Publishes under the Atomic Architects group (Tess Smidt's MIT lab); co-authors include Julia Balla, Jeremiah Bailey, Elyssa Hofgard, Tommi Jaakkola, Ryley McConkey.Sep 2025
Working on ML models creativity
• Studied the geometry of generative models, including how guidance affects diffusion sampling and how learned representations reflect the underlying structure of the data distribu
• Developed a decision engine for Kubernetes job retries, expanding pod failure observability coverage from 25% to 100% and reducing queued job times by ~10%. • Built a data analyt
• Built predictive models for inventory replacement and cross-item demand interactions for Zara. • Performed large-scale data transformation and feature engineering on hundreds of
• Provided 1-on-1 session for 600+ students to debug and optimize their weekly Python programming projects during office hours as a part of the ’Fundamental of Programming (6.1010)
Competition record
Syria · IMO