2026
When the Defense Writes the Refusal: Auditing Keyword-Scored Evaluation of Inference-Time Defenses for Multimodal Large Language Models
Bulat Nutfullin, Vladimir Evgrafov, Dmitry Namiot · DAMDID/RCDL 2026
Hi, I am
Vladimir Evgrafov
Leading applied AI/ML systems and an organization-wide AI track. Researching LLM-agent security as a PhD student at MSU CMC.
About
I lead the AI function and the development of applied AI/ML systems in a large organization. I shape the technical architecture and research direction, and build the process that turns algorithms into reliable production software for biomedical applications.
Previously, I led technology at the viju streaming service for almost five years. I built a 35-person department across seven teams, migrated a legacy backend to Go microservices, and owned architecture, infrastructure, AI/ML, and recommender systems.
My work combines AI/ML and scientific algorithms, production engineering, and engineering leadership. In parallel, I am writing a PhD thesis at Lomonosov MSU CMC and publishing research on LLM-agent security, attacks, and defenses for AI systems.
Education
Lomonosov Moscow State University, CMC
Thesis on the security of LLM-based agentic systems.
A method for converting conventional artificial neural networks into spiking neural networks (ANN→SNN): the converted models preserved the original model's accuracy while enabling more energy-efficient event-driven inference.
Automatic generation of inputs for testing web page template engines: grammar-based fuzzing that identifies template-processing flaws leading to XSS, server-side code execution, and denial of service.
Experience
NDA — Head of AI
Lead an organization-wide AI track and the development of applied AI/ML systems. Shape the technical architecture and research direction, and establish the path from algorithms to reliable production software.
Viasat Tech — Chief Technology Officer (CTO)
Led AI/ML and recommender systems alongside architecture, engineering, and infrastructure. Built a 35-person department across seven teams. Designed the migration from a legacy backend to Go microservices, enabling 100× the load at one quarter of the resource consumption. Delivered 100+ releases across seven platforms, raised quarterly plan completion from 40% to 83%, cut Android crash rate 14×, halved video start time, and reached 99.88% iOS crash-free sessions.
Atlas Biomed — Backend Team Lead
Built systems for processing and interpreting medical, genomic, and microbiome data. Implemented ML mechanisms and scientific algorithms, designed computational pipelines, and translated research logic into reliable services.
TRMPLN Media — Golang Developer
Developed a high-load, fault-tolerant ad-serving system and decision algorithms based on user data, with safe handling of personal information in production.
Multy — Golang Developer
Built microservices for communicating with various financial systems.
Okmeter — Golang Developer
Worked on a monitoring service: agents collecting metrics from physical devices and individual processes on Linux.
SolidWall — Python Developer
Wrote and maintained the core Python codebase, built the software build and deployment system, and administered CI and monitoring.
Publications
Each paper has its own page with a short summary of what the work is about and why it matters. Full list on Google Scholar.
2026
Attack Methods and Defenses in LLM-Based Agentic Systems
VA Evgrafov, BM Nutfullin, DE Namiot · International Journal of Open Information Technologies 14 (5), 1-8
2023
A Review of Potential Risks of Systems with Multimodal Language Models
B Nutfullin, V Evgrafov · Tikhonov Readings, 121
2021
Spiking Neural Networks
VA Evgrafov, EA Ilyushin · International Journal of Open Information Technologies 9 (7), 21-31
Contact
Open to professional conversations about AI/ML leadership, applied AI systems, architecture, biomedical applications, and research collaboration.