Mikhail Vakulich
AIOps engineer
Saint Petersburg · open to relocation · remote
Profile
AIOps engineer building AI systems for carrier-grade network operations, currently also a PhD candidate in AI and machine learning. 4.5 years in telecom, starting in third-line support and moving on to IP/MPLS backbone operations. For the past two years, development of agentic LLM systems based on that network's telemetry and tickets, along with maintenance-conflict monitoring and capacity forecasting. Every system listed was delivered independently, from requirements to production.
Requires an EU Blue Card. Both universities are listed in anabin with status H+, both degrees with the assessment "entspricht" (equivalent to the German degree).
Experience
AI/ML for network infrastructure: 02/2024 – present · Telecom: since 03/2022
AI/AIOps Engineer
05/2026 – presentNapa 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. State is stored in PostgreSQL checkpoints, so a scenario survives restarts and pauses of several days. The model handles text only, and all network changes are made by deterministic code.
- Agent platform: a scenario runtime on Temporal and LangGraph (deterministic scenario compilation, human-in-the-loop approval, retryable and fatal error classes) and a security module that expresses the access model as executable code: zones, roles, agent autonomy levels, an append-only decision journal.
- PII Firewall for de-identifying work documents before they reach a language model: four detection layers (checksums of Russian identifiers, an NER model, a gazetteer, secret blocking), zone isolation at the network topology level, an encrypted substitution map.
Backbone Network Engineer
02/2024 – 04/2026ER-Telecom Holding (02/2024 – 07/2024 as an intern)
- IP/MPLS 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 as initiative lead and sole developer, roughly 200 hours of work. 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 with an audited toolset over NETCONF/PyEZ (no shell access) and port diagnostics in natural language.
Team Lead, Third-Level Support
03/2022 – 02/2024Rostelecom
- Operation of district network nodes and distribution network: fault diagnosis and resolution, third-line technical support, subscriber provisioning.
Selected projects
PPR/AVR Monitor
Monitoring conflicts between planned maintenance and outages: events from the monitoring system, BPMS, mail and manual input, deduplicated and checked for overlaps by city and time. The LLM is used only to parse letters from external operators, and the remaining logic is deterministic code. Measured in production, 09/2026: 93.8% of fields in operator letters parsed correctly, 15 of 16 duty-confirmed critical conflicts retained, 48% fewer critical alerts.
PII Firewall
De-identifies documents before they are sent to a language model. It takes PDF, DOCX or XLSX files and produces de-identified text with an encrypted substitution map, and the model's answer comes back with the original values restored. Identifiers are validated by checksum, and the de-identification zone has no network route outside. On a synthetic gold set (40 documents, 322 mentions) recall and precision are 100%.
WebAgent-RL+ (master's thesis)
github.com/1vmc1/webagent-rl-plusTraining 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. Over 50,000 environment steps the mean phase-3 SR was 42%, reaching 61.7% across the last ten iterations, with non-zero SR on 20 of 25 tasks.
NeuroML: decoding and controlling neuronal cultures
PhD research on features and decoders for the activity of living neural networks and on closed-loop stimulation control within a 50 ms budget. Spike inference from calcium imaging on open data reached a correlation of 0.549, compared with 0.067 for the classical method. More than 700 tests, with negative results documented alongside positive ones.
Full project and research timeline → vakulich.work/en/roadmap
Education
PhD candidate, 1.2.1 Artificial intelligence and machine learning
09/2026 – presentPeter the Great St. Petersburg Polytechnic University
Topic: ML decoding of neuronal culture activity and adaptive closed-loop stimulation control (the NeuroML project).
MSc, machine learning and artificial intelligence
09/2024 – 06/2026Peter 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.
BSc, infocommunication technologies and communication systems (11.03.02)
09/2020 – 06/2024Perm National Research Polytechnic University, faculty of electrical engineering
Professional course “Machine learning and artificial neural networks”
2024PNRPU
Skills
- AI and machine learning
- LLM agents · MCP · LangGraph · RAG · LoRA · PyTorch · Reinforcement learning (PPO) · Transformers · NER (GLiNER) · Computer vision · scikit-learn
- Networks
- IP/MPLS · BGP · OSPF · L2/L3 VPN · NETCONF / PyEZ · Juniper JunOS · Cisco · Huawei · Extreme
- Development and infrastructure
- Python · FastAPI · Temporal · gRPC · PostgreSQL · Docker · Kubernetes · GitLab CI · Ansible · Terraform · Nginx · Linux
- Observability and security
- Prometheus · Grafana · Loki · Zabbix · PII de-identification · Policy as code · mTLS
Languages
- Russian - native
- German - TestDaF TDN 4 (B2/C1)
- English - B2+