Oleksandr Turevskiy

Backend & Platform Architect — Go, Kubernetes & Applied AI
Viersen · Düsseldorf region, Germany · German citizen

I have 25 years in software, as a developer, team lead, CTO and architect, and I still write code. I stay responsible for what I build, including when it goes wrong. My work today is platform and AI systems where mistakes are expensive: edge fleets, supervised autonomy and lending platforms. Outside work I ship complete products, three of them on both app stores.

Pro bono. I keep a few hours a month for non-profits, volunteer organisations and open-source projects — architecture review, Kubernetes and deployment, or a second opinion on where AI does and doesn't belong in your product. If that is you, write to me.

Oleksandr Turevskiy

Main projects

Mira GmbH (Rheinmetall Group)

Senior Backend Developer. I own the Ansible-driven edge fleet-management platform for a supervised-autonomy programme — a 50-device production fleet, validated against 1,000+ simulated devices — plus the Helm delivery layer that ships it to development, production and four customer-specific environments. The simulation platform alongside it was my own initiative: it proved the messaging, monitoring and control architecture holds at twenty times the production fleet, and reproduces scenarios that are impractical to stage with real vehicles.

Twelvegrow

Senior Backend Engineer, freelance. Core Go services and the gRPC contracts between them, cash-flow modelling across both React front ends, BigQuery and dbt over banking data, and the Anthropic-backed document intake pipeline.

Before
Galeria Karstadt Kaufhof · Nylas · JoinClap · ControlExpert · VMRay · MySupermarket · SeeYourTravel (CTO) · Itera · Infopulse (Bosch Security Systems). Led teams and mentored engineers at Itera and Bosch. Full history in the CV.

Working with AI — in production, and in the workshop

In production
Grounded assistants. An in-app assistant for a fleet platform in a supervised-autonomy programme: it routes free text to navigation, a device action or a question. Actions always surface a confirmation step rather than executing directly, answers are grounded strictly in the product's own help content, and every figure is computed server-side from role-scoped endpoints — never generated by the model. In that setting the boundary is the design.
In production
Structured extraction. Document and email intake for a lending platform — PDFs, images and attachments into structured records. JSON-schema-constrained output with a prompt-based fallback for when constrained decoding fails, plus response caching and streaming. The fallback is the part that matters: pipelines degrade instead of breaking.
Team scale
AI-assisted engineering. Container and supply-chain security wired into the GitHub and GitLab pipelines — Trivy and Grype against CycloneDX SBOMs, Dockerised integration tests — so security, quality and compliance issues surface at build time rather than as customer-facing incidents and recovery cost. Plus agentic coding tools in a large codebase, and bringing the team with me.
Self-hosted
Local inference, agents and retrieval. A self-hosted RAG stack on my own hardware: a vector database indexing my Google Drive corpus for semantic search, plus retrieval over scraped sites and PDFs — all running on local models through Ollama, so nothing leaves the house. Hermes for local agents, MCP servers exposing my own tools to an assistant, custom agent skills, and n8n for the workflow automation that needs no model at all. At home, Frigate has run my video surveillance on local AI detection for a year now, and a Raspberry Pi with an AI accelerator serves as a small LLM server.
Speech
Low-resource TTS. Yiddish speech synthesis over FastSpeech2 checkpoints, and generated neural-voice pronunciation across 80+ languages for a vocabulary app. Most of the work is in the languages nobody built for.

Own projects — where I get to build the whole thing

SeeYourTravel

Live since 2015

The travel-community product I built as CTO in 2015 and never let go of. Brought back in 2026: containerised behind a Cloudflare Tunnel with no inbound ports open, the whole deployment driven from a Makefile. A ten-year-old codebase kept alive without a rewrite.

Docker · Cloudflare Tunnel · OAuth
seeyourtravel.app →

An OpenVPN-protocol server written from the wire format up — TLS control channel, AES-256-GCM data channel, rekeying, NAT, IPv6-in-tunnel. A stock openvpn client connects to it unmodified. I started it after one disappointment too many with VPN providers, and I use it myself.

Rust · verified against OpenVPN 2.6.14
Source on GitHub →

5001 Words

iOS · Android

Vocabulary trainer for iPhone, iPad and Apple Watch, with a native Android build. 290+ decks across 84 languages, neural-voice pronunciation, and downloadable packs with a range-fetch so the word list is usable seconds into a download. Yiddish is in there too: no provider offered a voice, so I built one.

SwiftUI · WidgetKit · Kotlin

5001 Tomatoes

iOS · Android

A focus timer that stays out of the way. Live Activities, home-screen, lock-screen and StandBy widgets, 76+ locales, and data that never leaves the device. It began as a five-minute timer for learning Spanish, and the folder is still called FiveMinutes.

SwiftUI · ActivityKit · WidgetKit

5001 Axolotls

Play in browser

An endless procedurally generated underwater maze. Server and browser client import the same rendering-free simulation module, so multiplayer is server-authoritative with no second implementation to keep in sync. Identity is a SHA-256 proof-of-work — no email, no password, no free text anywhere. My children found out that the roundabout can throw you off. I had not planned that, and I left it in.

Three.js · WebGL · Node · Capacitor · 76 languages

A trilingual site for a Ukrainian scouting organisation in Germany — Ukrainian, German and English, content managed by the people who run it rather than by me.

Next.js 15 · React 19 · Sanity · Vercel
Visit site →

Core skills — what I actually use

Languages
Go · Rust · TypeScript / JavaScript · Python · C# · SQL · Swift
AI / LLM
Grounded assistants · structured extraction against JSON schemas · prompt fallback for constrained-decoding failures · response caching · streaming · document and email intake · RAG · vector search · local inference (Ollama) · MCP servers · agent skills
Infrastructure as Code
Ansible — playbooks, roles and inventories driving fleet orchestration at scale · Terraform · Helm / Helmwave · GitOps · Argo
Platform
Kubernetes · Docker · AWS (EKS) · Google Cloud · Microsoft Azure · gRPC · RabbitMQ / AMQP · OpenTelemetry · Prometheus / Grafana / Loki · Trivy · Grype · CycloneDX SBOM · CircleCI
Web & Mobile
React · Next.js · Vue / Nuxt · Chakra UI · Tailwind · Sanity · SwiftUI · WidgetKit · ActivityKit · Capacitor · Three.js / WebGL
Data
PostgreSQL · MS SQL Server · MongoDB · Couchbase · Elasticsearch · BigQuery · dbt

Writing — selected notes

An engineering notebook I have kept since 2006. Recent entries first; nothing deleted for looking dated.

Certifications

Five since 2023, roughly two a year. I use certification as a forcing function to learn a thing properly — not as decoration.

iSAQB CPSA-F — Software Architecture, 2025 CKAD — Kubernetes Application Developer, 2024 AWS Developer Associate, 2023 AWS Solutions Architect Associate, 2023 HashiCorp Terraform Associate, 2023 Earlier, not maintained: PMP 2013 · ISTQB 2012 · CSM 2008 · MCSD 2001