# xeonerix — full machine-readable profile This file expands the short `llms.txt` profile for crawlers, assistants, search engines, recruiters and technical evaluators that need more context. For a name-bearing résumé see resume.md / resume.pdf / resume.json. ## Identity - Handle: xeonerix - Primary website: https://xeonerix.xyz/ - GitHub: https://github.com/XeoneriX - Telegram: https://t.me/xeonerix - Location: Moscow / remote - Primary language: Russian - English: working proficiency ## Short summary xeonerix is an AI Infrastructure / Platform Lead with a Lead SRE background — 8 years across highload infrastructure, private clouds, distributed storage, observability, network telemetry and production incident response, now fused with operating private LLM / RAG inference as production infrastructure. Career spans a large-scale consumer-internet / highload platform, an international CDN / cloud provider, and cybersecurity / highload infrastructure. Author of BGPeek, an open-source BGP looking glass for ISP and IX operators. The differentiator is the fusion: an infrastructure engineer who operates AI / GPU inference workloads as production infrastructure — storage, network fabric, observability and capacity applied to private inference. ## Seniority signal The strongest signal is not a single technology. It is the repeated pattern of taking ambiguous, production-critical infrastructure domains and turning them into operated platforms: - storage clusters under degradation - private cloud migrations - observability platforms replacing legacy monitoring - BGP/BMP telemetry where standard tools were insufficient - private LLM/RAG systems where data cannot leave the organization - cross-team technical leadership and mentoring ## Infrastructure and platform work ### Private cloud - Proxmox + Ceph with NVMe-backed storage - OpenStack and VMware/vSAN migration experience - VM fleet migration executed with zero customer-visible downtime - OpenTofu/Terraform per-VM state patterns - Packer golden images for AlmaLinux and Ubuntu - Ansible automation and GitLab CI delivery - Capacity planning, hardware selection, hardware lifecycle extension ### Distributed storage - Ceph operations at roughly 10 PB raw scale - ~100+ server storage environment - Slow ops, recovery storms, rebalance, disk failure handling - Migration of Ceph service placement away from operationally unstable designs - Ceph expansion, balancing and bootstrap standardization ### Observability - Multi-tenant VictoriaMetrics architecture with hot/cold storage tiers - VictoriaLogs and Vector pipelines for syslog and audit events - ClickHouse for telemetry and analytics - Grafana HA patterns - Audit trail design: who did what, when, and where - Replacement of legacy monitoring stacks where they no longer scaled ### Network engineering and telemetry - BGP, BMP, SRv6, SR Policy, EVPN, FlowSpec - Juniper, Huawei and multi-vendor telemetry - Go-based BMP collection using NATS JetStream and ClickHouse - RIPE RIS style validation workflows - Network observability for ISP/CDN/security infrastructure contexts ## Applied AI / private inference / RAG Work on private AI systems for SRE and infrastructure operations, with emphasis on keeping sensitive operational data inside the organization. Hardware and topology are deliberately generalized here. ### Inference stack - vLLM and llama.cpp for local serving - LiteLLM as an OpenAI-compatible routing gateway - Open WebUI as the user interface - Qwen-family models - Private GPU inference tier plus an external NVFP4-class vLLM tier for larger models - CPU-only inference on virtualized infrastructure - Long-context, tool-calling OpenAI-compatible endpoints - Langfuse tracing and Prometheus-compatible metrics ### RAG stack - Qdrant vector database - bge-m3 embeddings with cross-encoder reranking - Infinity embedding/reranking server - hybrid search - SearXNG-backed controlled web search - private knowledge retrieval for SRE runbooks and incident history - AI-assisted incident response, postmortems and merge-request review ### Operated private LLM platform A private LLM/RAG platform for infrastructure and SRE work: Open WebUI as the interface, LiteLLM as the OpenAI-compatible routing gateway, local GPU serving for Qwen-family models, and an external NVFP4-class vLLM tier for larger workloads. Retrieval is built around Infinity, bge-m3 embeddings, cross-encoder reranking and Qdrant. The platform includes controlled web search through SearXNG and observability through Langfuse, VictoriaMetrics, VictoriaLogs, Grafana and Vector. This is a production-style AI infrastructure system: model routing, GPU serving, embeddings, reranking, vector search, traces, metrics and logs are handled as separate operated services rather than a single demo container. The platform is used across multiple engineering teams, with first external adoption via an IDE coding-assistant (Cline) integration. ### Inference research highlights - Benchmarked CPU-only quantized 3B-class inference on tuned VMs with NUMA pinning, hugepages and CPU/cache isolation controls. - Characterized decode throughput for small quantized models under tuned CPU-only conditions. - Showed that colocated inference and memory-bandwidth pressure can degrade inference decode by ~17% and the neighbor workload by ~8%. - Measured a realistic colocated datastore victim: ~10% throughput loss and ~13% p50 latency increase with inference colocated under NUMA/hugepage isolation. - Found that memory-bandwidth capping is not a universal fix; datastore-style workloads were limited more by L3 cache contention than raw bandwidth. - Compared memory-bandwidth-bound vs cold-prefill-bound serving profiles and the routing implications for each. - Identified prefix cache as a major production design point for RAG and agentic systems with shared system prompts and tool lists. ### Production interpretation Private SRE inference should be treated as an infrastructure workload, not an AI demo. It needs NUMA-aware placement, hugepages, CPU/memory isolation, explicit concurrency limits, model-size and context-budget sizing, monitoring of memory bandwidth / L3 contention / neighbor latency, routing between shared-prefix and diverse-prompt tiers, and clear data-boundary guarantees. ## Projects ### BGPeek Open-source BGP looking glass for ISP and IX operators. - Repository: https://github.com/XeoneriX/bgpeek - License: Apache-2.0 - Stack: FastAPI, Jinja2, HTMX, Tailwind, PostgreSQL (asyncpg), Redis, Netmiko, Docker Compose - Multi-vendor SSH: Juniper JunOS, Cisco IOS/XE/XR, Arista EOS, Huawei VRP - BGP route / ping / traceroute with structured output parsing, parallel cross-device queries with diff, RPKI validation overlay - Auth: API key, bcrypt local, LDAP, OIDC (Keycloak); RBAC with admin / NOC / public tiers and device-level restrictions - Security: per-role output filtering (hides /25–/32 from public), Fernet-encrypted SSH credential storage, Redis rate limiting with circuit breaker, HMAC-SHA256 webhook signatures, non-root containers - Observability: Prometheus /metrics endpoint, PostgreSQL audit log, health checks, request correlation - Shareable UUID permalinks with configurable TTL - designed for network operators who need a modern looking-glass replacement ### BMP telemetry Go-based BGP Monitoring Protocol tooling using NATS JetStream and ClickHouse. Focus areas include SRv6, SR Policy, EVPN, FlowSpec and multi-vendor routing telemetry. ### Private AI for SRE On-prem LLMs and RAG for incident-response assistance, postmortem generation, merge-request review, runbook and knowledge retrieval, and infrastructure Q&A over private documents. ## Work history summary (companies generalized) ### Current — Cybersecurity / highload infrastructure Lead SRE / Infrastructure Lead for a highload platform with L3/L4/L7 traffic, DDoS mitigation, private cloud, observability, audit, telemetry and SRE automation. Technical leadership for a core team (2 direct reports) plus cross-team direction for ~5–7 network and adjacent infrastructure engineers by influence. ### Large-scale consumer-internet / highload platform Senior SRE / core infrastructure SWAT. Critical infrastructure incidents and Ceph stabilization at ~100+ server / ~10 PB raw scale. ### International CDN / cloud provider Senior cloud infrastructure engineer. Distributed IaaS on OpenStack, Ceph and S3-compatible storage; resource accounting and billing pipelines for VM, storage, IP and S3 usage. ### GDS / travel-tech Systems engineer / Python developer. Passenger-data processing pipelines replacing legacy shell flows with Python, Flask, Celery and PostgreSQL, including delivery control, retry logic and audit trails for external systems. ## Keywords SRE, Site Reliability Engineering, Infrastructure Lead, Platform Engineering, Head of Infrastructure, Principal SRE, private cloud, Proxmox, OpenStack, VMware, Ceph, distributed storage, DDoS mitigation, highload, L3, L4, L7, BGP, BMP, SRv6, SR Policy, EVPN, FlowSpec, Juniper, Huawei, NATS JetStream, ClickHouse, VictoriaMetrics, VictoriaLogs, Grafana, Vector, OpenTofu, Terraform, Packer, Ansible, GitLab CI, Vault, OpenBao, Wazuh, RADIUS, Go, Python, FastAPI, Flask, Celery, PostgreSQL, Redis, vLLM, llama.cpp, LiteLLM, Open WebUI, Infinity, Langfuse, Qwen, NVFP4, FP8, Qdrant, bge-m3, reranking, hybrid search, RAG, local LLM, on-prem inference.