Machine Learning Engineer & Researcher — LLM & Agent Security
Google Scholar · Telegram · Email
I build infrastructure for LLM agents — runtimes, harnesses, and long-term memory — and research their security. My published work focuses on prompt-injection attacks and defenses, trojan/backdoor detection in neural networks, and uncertainty-aware evaluation of vision-language models. PhD candidate at MSU's Department of Information Security (expected 2027). I also collaborate on computer-vision and statistical methods for forest ecology.
Selected Publications
- V. Kostumov, B. Nutfullin, O. Pilipenko, E. Ilyushin. Uncertainty-Aware Evaluation for Vision-Language Models. arXiv:2402.14418, 2024. [arXiv]
- D. Khomsky, N. Maloyan, B. Nutfullin. Prompt Injection Attacks in Defended Systems. DCCN, 2024.
- N. Maloyan, E. Verma, B. Nutfullin, B. Ashinov. Trojan Detection in Large Language Models: Insights from the Trojan Detection Challenge. arXiv:2404.13660, 2024. [arXiv]
- V. A. Evgrafov, B. M. Nutfullin, D. E. Namiot. Attack Methods and Defenses in LLM-Based Agentic Systems. IJOIT 14(5), 2026.
- N. Maloyan, B. Nutfullin, E. Ilyushin. Dialog-22 RuATD: Generated Text Detection. arXiv:2206.08029, 2022. [arXiv]
- A. E. Bykanov, G. V. Danilov, V. V. Kostumov, O. G. Pilipenko, B. M. Nutfullin, et al. Artificial Intelligence Technologies in the Microsurgical Operating Room. Sovremennye Tehnologii v Medicine 15(2), 2023.
Experience
ML Engineer — AI startup (2024 – present)
Building the agentic runtime and harness that LLM agents run on. Built the document parser at the core of the pipeline from scratch — handling PDF, DWG, and other formats — and optimized it for accuracy on hard real-world inputs. Designed the agents' long-term memory system.
Viasat — ML Engineer (2021 – Jul 2024)
Computer vision and generative ML for a streaming platform: automatic video highlighting, face replacement for actor substitution, and personalized poster generation with Stable Diffusion. Built low-latency online and offline recommendation serving on ClickHouse.
Forest ecology — research collaboration (2022 – present)
Statistical testing of tree-species co-occurrence hypotheses and species-compatibility metrics.
Forest ecology — startup (2018 – 2022)
3D point-cloud and remote-sensing CV: tree classification and mistletoe detection from lidar point clouds, satellite-imagery analysis, and geometric registration of field-survey and lidar maps.
SAS — Intern (Summer 2019)
Sales-data analytics; benchmarked open-source vs. proprietary ML toolchains.
Education
PhD Candidate (Cand. of Technical Sciences), Computer Science (2021 – present, expected 2027)
MSU, Faculty of Computational Mathematics and Cybernetics — Department of Information Security. Dissertation: Adaptive defense methods for multimodal LLMs at inference — the MARS framework (adaptive routing of inference-time defenses).
MSc, Computer Science (2019 – 2021)
MSU, Faculty of Computational Mathematics and Cybernetics, Dept. of Open Information Technology. Thesis: automatic speaker diarization methods.
BSc, Computer Engineering Technology (2015 – 2019)
Skills
- Languages — Python, SQL
- LLM / Agents — agentic runtimes, long-term memory, prompt-injection & robustness evaluation
- ML / Data — PyTorch, XGBoost
- Backend / Infra — FastAPI, Temporal, Docker, AWS, git
- Databases — PostgreSQL, ClickHouse, SQLite