Andrei Beliaev

Hey, I'm Andrei! I work on machine learning at Epivara, where I build vision models for tissue and cell analysis and develop systems for animal tracking, re-ID, pose estimation, and behavioral classification. I've also built agents for analyzing, structuring, and ingesting research data. Previously, at Tinkoff, I worked on anti-money-laundering systems and tools for measuring and improving product quality. I recently finished my master's in computer science at UIUC.

My interests in machine learning include representation learning and continual learning, particularly how models adapt to new tasks and environments, as well as energy-based models, iterative inference, and ML for scientific discovery.

I'm currently looking for machine learning, research engineering, and software engineering roles. Feel free to reach out.

I'd be happy to connect for collaboration. Feel free to email me.

Selected work

HistoForge logo

HistoForge

I built and released an open-source platform that lets researchers train, run, and share pathology models directly from QuPath.

More details about HistoForge

A DINOv3 + CNN-adapter model captures tissue-scale context while preserving cell-level detail. It processes 16× more tissue per pass, trains 3.7× faster, and halves inference cost compared with a state-of-the-art baseline. It improves segmentation and classification and is applicable to roughly 3/4 of surveyed tissue types.

Longer term, I'd like to build on this work toward a general-purpose, promptable model for segmenting tissue regions and individual structures, so researchers can quantify their number, morphology, and spatial relationships.

Skeleton-only 3D pose comparison of my pipeline, s-DANNCE, and LP3D, using matched frames, viewpoints, and scale. No animal footage.

Animal perception

Developing detection, re-identification, and multi-camera 3D pose tracking for animal behavior measurement, with a focus on maintaining identity through occlusion and re-entry. Work in progress.

The EpiBot analytics agent writing queries and generating charts.

Research data agents

Built a conversational analytics platform that queries research databases, runs statistical analyses, and generates charts and reports. A human-in-the-loop ingestion agent turns experimental data into structured records.

Experiments

  • Energy-Based Transformers

    Blog: Energy-Based Transformers

    Ran MCMC sampling ablations in Energy-Based Transformers, comparing perplexity across pretraining configurations.

  • Image immunization

    Code: Image immunization

    Implemented methods for protecting images from diffusion-based editing, using timestep-universal gradients and learned perturbation networks.

  • Entropy-aware sampling in vLLM

    Code: Entropy-aware sampling in vLLM

    Implemented GPU-batched lookahead to control sampling diversity by penalizing entropy-reducing tokens.