Optimization First: Cornell ORIE Welcomes Nikita Doikov to Advance Efficient AI
Nikita Doikov has joined Cornell Engineering's School of Operations Research and Information Engineering as an assistant professor. He brings a clear thesis: better optimization means leaner, faster, more reliable AI.
His work centers on designing and analyzing the algorithms at the core of machine learning systems and large-scale models. "Training an AI system is, at its heart, an optimization problem," Doikov said. "If we understand these algorithms better and if we can make them more efficient, we can significantly reduce the resources required to train modern models."
Why this matters if you run operations
For operations leaders, algorithmic efficiency shows up on the P&L. It cuts GPU hours, lowers energy spend, improves job predictability, and shortens time-to-value.
- Lower cost per experiment: faster convergence reduces retrains and dead-end runs.
- Higher throughput: better utilization of compute clusters and fewer stalled jobs.
- Operational stability: algorithms with guarantees behave more predictably under scale.
- Resource planning: clearer training curves make capacity and budget planning easier.
Research focus: continuous optimization for modern AI
Doikov's group will push on continuous optimization, targeting algorithms with provable convergence and practical speed. Expect work that studies the geometry of loss functions in large models and turns those insights into methods you can trust in production.
If you want the foundation behind this direction, the classic reference is Convex Optimization by Boyd and Vandenberghe (free online book). It's the playbook many modern training methods grew from.
A path built on curiosity and applied math
Doikov grew up in Vladivostok, a city near the Sea of Japan known for dramatic nature and wildlife. Early exposure to physics from his engineer father and music from his mother, a piano teacher, gave him range-computers stayed the constant.
At Moscow State University's Faculty of Computational Mathematics and Cybernetics, he blended computer science with core math and sharpened his algorithmic thinking through programming competitions. Internships at Google in Zurich gave him industry context, but research proved the stronger pull.
He completed his Ph.D. at Université Catholique de Louvain under Yurii Nesterov, a leading voice in optimization, then moved to EPFL in Lausanne for postdoctoral work in a machine learning and optimization lab. There he broadened his perspective by mentoring students and collaborating with applied teams.
Practical ideas you can apply this quarter
- Instrument every training run: track gradient norms, step sizes, and per-epoch wall time to spot inefficiencies early.
- Pilot algorithms with guarantees: benchmark methods with known convergence rates before scaling to full datasets.
- Budget by convergence, not epochs: set stop criteria tied to validation metrics and compute ceilings.
- Schedule for energy windows: align long runs with off-peak energy rates to cut cost without changing code.
- Maintain a small-batch baseline: a fast, repeatable baseline exposes regressions in data, code, or hyperparameters.
Teaching and what's next at Cornell
At Cornell, Doikov will build a group focused on the theory that makes large-scale training efficient and dependable. In Spring 2026, he's teaching a graduate course on continuous optimization-directly linked to the work his lab advances.
For teams upskilling on AI and operations
If your roadmap includes bringing AI into core workflows, a structured path saves time and budget. Explore role-based learning tracks here and automation-focused resources here.
Outside the lab
Recently, Doikov started learning to sail and is looking forward to exploring the Finger Lakes after settling in Ithaca. A good balance: precision in research, open water on weekends.
Your membership also unlocks:
