Platform & Infrastructure Engineer · LLM agents, evals, agent safety
Head of Platform & Infrastructure @ Algous Studio · MSc AI @ ITMO (AI Talent Hub)

I run the platform and infrastructure track at a VFX studio in Saint Petersburg — 60+ Linux hosts, ~1 PB of storage, all on-prem. Lately the work has been moving up the stack: self-hosted inference, LLM agents, and deciding what an agent is actually allowed to do.

I tend to care less about the stack than about being able to prove a system works — frozen evaluation protocols, benchmarks with real execution, property tests instead of claims.

Auto mode for Kilo Code — an auto-approval layer for an open-source coding agent that decides on what a command does, not how it was written. Upstream PR, ~9.7k lines. Benchmarked on 498 episodes with real command execution: attack success rate 100% → 2.6% at 80% agent utility. 3rd place in the Alfa-Bank case at AI Product Hack 2026.

AI Helpdesk Automation — a local LLM agent that closes routine IT requests in the studio, in production since 2026. Safe tool calling by design: the model picks an action, never builds a command. 96% routing accuracy on a frozen holdout against 63.5% for a keyword baseline, κ 0.97 on blind annotation. The candidate model was rejected on safety, not accuracy.