Mikhail Vakulich

Napa Labs

• A durable LangGraph orchestrator for backbone-link incidents: ticket parsing, diagnostics on both ends of the link over NETCONF/PyEZ, an LLM-written conclusion from a JSON digest, port deactivation via commit confirmed with engineer approval, restore and re-check.
• Graph state in PostgreSQL checkpoints - a scenario survives service restarts and multi-day pauses while a field crew is on its way.
• A strict split of responsibility: the model only handles text and JSON; everything that changes network configuration is deterministic code.

ER-Telecom Holding (intern from Feb to Jul 2024)

• Backbone operations: configuring customer services (BGP, IX, L2/L3 VPN), network infrastructure and load balancing, running planned maintenance, resolving outages.
• Built and shipped a monitoring system for planned-work and outage conflicts - initiative lead and sole developer, roughly 200 hours. Projected impact: conflict detection in under 5 minutes instead of 15-30 minutes of manual cross-checking, MTTR down 30% on conflicts, 247 engineer-hours freed per year.
• Backbone traffic forecasting on a Temporal Fusion Transformer: P10/P50/P90 quantile forecast a week ahead, an N+2 capacity-expansion plan, an anomaly detector for the duty shift.
• An MCP server for Juniper devices: an audited toolset over NETCONF/PyEZ instead of shell access, port diagnostics in natural language.

Rostelecom, service centre

• Third-line subscriber support: fault clearance on subscriber lines, in-building distribution networks and district networks.
• Maintenance of district communication nodes, connecting new subscribers.

Training web agents on MiniWoB++ within a single 16 GB GPU: a frozen Qwen3-8B planner in 4-bit quantization plus a trainable PPO executor. A curriculum driven by the variance of the critic's estimates and a graded reward (ORM + DOM verification + waypoints). Over a 50,000 environment-step run: mean phase-3 SR of 42%, 61.7% across the last ten iterations, non-zero SR on 20 of 25 tasks.

An agent that monitors conflicts between planned maintenance and outages: events collected from the monitoring system, BPMS, corporate mail and manual input, deduplicated, then checked for overlaps by city and time. The LLM owns exactly one step - parsing unstructured letters from external operators; the rest is deterministic code. Prometheus, Loki, Flyway migrations, CI with integration tests.

A transformer over mel-spectrograms classifies a music track's genre, with inference wrapped in a Telegram bot. Full cycle: dataset, training, containerization, deployment.

Peter the Great St. Petersburg Polytechnic University, Institute of Intelligent Systems and Technologies

Thesis: “Training web agents with an adaptive curriculum and graded reward” - the WebAgent-RL+ framework.

Perm National Research Polytechnic University, faculty of electrical engineering

PNRPU