cv

Senior Applied Scientist working on foundation models for security, LLM alignment, and agentic AI.

📄 Download CV (PDF)

The PDF is the source of truth; it’s regenerated from LaTeX and replaces this file each release. The web sections below are an at-a-glance summary; see the PDF for full detail.

Now

Senior Applied Scientist, Amazon AWS (AI Security and Observability). (2022 – present.)

  • Leading agentic AI security: detection models on agent runtime traces (Amazon Bedrock AgentCore) that surface intent deviation and policy violations, extended with OS-level signals (eBPF, system logs) and auto-detection of agentic workloads in dynamic environments.
  • Co-authored the first paired-evidence study of kernel-level vs. application-layer signal for agent security (arXiv:2609.28915): the ACE corpus of 4,047 agent sessions pairing syscall traces with tool manifests and transcripts across 17 threat models. Kernel evidence is discriminative on its own, and cross-layer composition generally beats either single layer.
  • Co-developed PurpleAudit (NeurIPS 2026, Evaluations & Datasets), a co-evolutionary red/blue-team auditing framework for multi-agent systems, in which attacks and defenses evolve against each other over execution traces. Targets task hijacking, the case with no malicious verb to refuse: 31.26% attack success on a frontier backend where all 22 baseline attacks fail, and a prompt-level defense that holds at negligible utility cost.
  • Previously tech lead of the audit-log foundation model, shipped to production Sep 2025, cutting customer-facing false positives by 20–30%. Designed the 10B → 10M diversity-preserving sampling, contrastive fine-tuning for tenant-specific entities (asnOrg, api), multi-signal evaluation, and interpretability tooling.
  • Co-designed a memory-efficient hierarchical-log architecture for long-context, high-throughput security workloads (NeurIPS 2026 Workshop on Long-Context Foundation Models).

Earlier

Yahoo! Research, New York. (2017 – 2021, Research Scientist.) Contributed to ExtremeText (no-regret hierarchical softmax for extreme multi-label classification), VisualTextRank (unsupervised graph-based content extraction for ad text→image search), and unified multi-task CTR/CVR models that lifted advertising ROI by >10% in online A/B tests.

Education

  • PhD, Moscow Institute of Physics and Technology: Computer Science, Mathematical Modeling, 2016.
  • MS with Honors, MIPT: Applied Mathematics & Physics, 2013.
  • BS with Honors, MIPT: Applied Mathematics & Physics, 2011.
  • Yandex School of Data Analysis: Diploma in Data Science, 2012.

Interests

Agentic AI security · Automated red-teaming & co-evolutionary auditing · Foundation models for logs · Robust / self-supervised learning · Anomaly detection · Representation learning for structured / tabular data · Evaluation & interpretability · LLM alignment.