Leonid Gremyachikh

Several teams — a trading firm, a crypto-social product, on-chain market infrastructure. The work has the same shape in all of them: model how markets and people actually behave, then put that model into production where an adversary is adapting to it.

• Systematic trading research on tick and limit-order-book data — microstructure features, out-of-time validation, backtests with fees, slippage and fill models, and a positive out-of-sample result on live futures.
• Reinforcement learning agents bidding in on-chain auctions; part of this work is published.
• Scoring and detection of manipulation and abuse at scale — sequence models over behavioural time series that cut fraud losses by more than half and false positives by an order of magnitude.
• The production side throughout: real-time and batch pipelines, scheduled retraining, monitoring, and the tooling analysts work in.

Market microstructureReinforcement learningPyTorchXGBoostTime seriesProduction ML

Research projects taken on directly with external clients — my own pipeline, from the first conversation to the delivered system.

• Reinforcement learning for advertising budget allocation. An agent that reallocates spend across campaigns under a total budget constraint — environment and simulator, policy- and value-based agents, statistical comparison of approaches. Delivered as a working MVP, under NDA.
• Satellite collision avoidance. Years on one problem, from research to a running service: a reinforcement learning system that estimates collision probability between space objects and computes the cheapest avoidance maneuver, on top of an orbital simulator built for it — then refinement of orbital position predictions, and the backend and web service around the models. First place at the Roscosmos Academy project competition (presentation), an MVP, a VR presentation of the system, and a paper I led. Reported at the IAA SciTech Forum, the RAS space debris conference and the SMILES school; covered by Yandex Academy and Izvestia.
• Reinforcement learning for no-limit poker. A smaller piece, but one I kept coming back to — environment, simulator, and policy- and value-based algorithms for decisions under incomplete information, with the pipeline used to compare them.

Reinforcement learningSimulationPythonR&D management

Computer vision for orthodontics: segmentation of teeth on 2D images and reconstruction of their 3D geometry from depth maps and UV unwraps, used to plan tooth correction.

PyTorch3D reconstructionSegmentationMedical imaging

Deep and reinforcement learning across physics, space and infrastructure, alongside teaching and student supervision.

• Space Navigator — reinforcement learning for satellite collision avoidance; where the work above began.
• Satellite positioning — machine learning on simulated ephemerides; authored the IDAO 2020 olympiad task.
• Seismic inversion — recovering velocity models of complex media with deep convolutional networks.
• Storage failure prediction — NLP over log streams and online change-point detection.
• Storage digital twin — conditional generative models standing in for production hardware.
• Reinforcement learning on digital twins — policies for hybrid distributed storage, tests on chaotic systems, Gym wrappers for the simulators.
• Distributed deep learning — training pipelines across GPUs and machines, containerised and reproducible.
• Model Gym — hyperparameter optimisation for comparing predictive models.
• awesome-log-analysis — an open reading list on log analysis, anomaly detection and AIOps.
• Teaching, thesis supervision, and mentoring at the SMILES school.

Reinforcement learningGenerative modelsNLPResearchTeaching

On a blockchain the order of transactions inside a block is itself worth money, and searchers compete in an auction for the right to set it. We model that auction — Polygon's Atlas — as a learning problem: a simulation environment for the auction dynamics, and a PPO agent that bids inside real-time latency budgets. Deployed alongside existing searchers the agent reaches 49% Maximum-Profit Capture, and improves substantially on baseline strategies under counterfactual evaluation.

Read on arXiv

Recovering the velocity structure of a complex medium from boundary measurements — the seismic inverse problem, in an acoustic setting. Forward modelling uses the grid-characteristic numerical method; the inverse direction is learned by deep convolutional networks. We modified the baseline UNet and showed a statistically significant gain in how closely the recovered velocity profiles match ground truth.

Read on arXiv

Real failure data from storage hardware is scarce, so we generated it: a discrete-event simulator in Go that reproduces the structure and dynamics of high-level storage building blocks. On top of that, an adapted change-point detection algorithm — density-ratio estimation via binary classifiers — finds failures online in the resulting time series. The point is to catch a failing component early in a data centre holding thousands of devices.

Read in JPCS · arXiv

Satellites in low orbit have to dodge debris, and every maneuver costs propellant that cannot be replaced. The paper describes a system that estimates collision probability for a pair of space objects and searches for the maneuver that avoids the collision at the smallest fuel cost, using reinforcement learning over a simulated orbital environment.

Read on arXiv

A simulator of a hybrid distributed storage system, built so that reinforcement learning agents can be trained against it rather than against production hardware. It reproduces the behaviour of the real system closely enough to tune policies, and exposes the interfaces an RL loop needs.

Read in PeerJ CS

Supervising student research in machine learning — deep learning, reinforcement learning, anomaly detection and time series. At HSE University in 2021–2022, at MIPT since 2024.

• Machine Intelligence, an interdisciplinary perspective — an online machine learning course for scientists, developed and taught on Pelican. Lectures and assignments in Python and PyTorch Lightning: deep learning foundations, computer vision, autoencoders and normalizing flows, anomaly detection, time series.
• Introduction to Data Analysis — seminars, HSE University, bachelor's programme.
• Python for Data Science — online course developed for Skillbox.
• Volunteer and mentor, 6th Machine Learning in High Energy Physics summer school, SMILES, Skoltech.

• 2026 — mentored student projects in MIPT's online master's programmes.
• 2025 — certified there as a mentor for technology projects.
• 2022 — summer school on entrepreneurship, HSE Faculty of Computer Science.

A VR installation built in Unreal Engine. Two actors share a world — you, and an artificial intelligence standing in for everyone who was there before you. You can leave traces and images in that world, and those traces are the only channel of communication you have. What does cooperation look like when language is taken away?

I built the artificial intelligence side of it: the agent the visitor meets, and the way it reads and answers what people leave behind.

Shown at the Game Junction exhibition, Khodynka gallery, Moscow, 2021.

• 2022 — Interview on the HSE Faculty of Computer Science website
• 2021 — Artigitl, digital art course
• 2020 — Poster, AI and Satellite Problems, SMILES international summer school, Skoltech
• 2020 — Authored the task for IDAO, the International Data Analysis Olympiad — data preparation, baseline and review of participant solutions — and ran the webinar on data analysis for satellite tracking
• 2020 — Computer Vision course, OTUS
• 2019 — Report at Space debris: fundamental and practical aspects of the threat, Institute of Astronomy RAS
• 2019 — Autumn School on Generative Models, HSE × Yandex
• 2018 — Paper report, IAA SciTech Forum, RUDN
• 2018 — RL Hackathon, A!ONE (MTS)