Top

Automation that
runs itself.

I’m Meryem Türkmen, an automation engineer for workflows, AI agents, AI enablement and the infrastructure underneath them.

  • Berlin
  • B.Sc. Computer Science, TU Berlin
  • German, Turkish, English
See the work

order.createdready

  1. Press the button: a real model on Cloudflare Workers AI classifies a sample note while the packet travels.

Automation Engineer · Workflows, AI Agents, AI Infrastructure

The system graph: triggers, an orchestrator, an agent with tools, an LLM gateway, a database and the hosting behind them.

Four pillars, one system.

Workflows 1 of 4

Every repeatable task becomes a workflow with a trigger, a retry and an owner.

n8n and Windmill for orchestration, Python and TypeScript where a node is not enough.

  • Idempotent steps, so a retry never books twice
  • Retries with backoff and a dead-letter queue for what still fails
  • One log per run, readable by the person who owns the process

AI agents 2 of 4

Agents get a narrow job, real tools and a hard stop.

Classification, extraction and triage against internal APIs, in production since 2025.

  • Structured output against a JSON schema, validated before anything moves
  • A tool allowlist per agent and an evaluation set before rollout
  • Human approval wherever money or a customer is involved

AI enablement 3 of 4

A team is enabled when the model is one call away and the keys are under control.

Marketing, supply chain, support and B2B sales run their own automations on it.

  • One gateway, per-app keys with rate limits and budgets
  • MCP servers over operations data instead of CSV exports
  • Skills, templates and decision papers written for non-technical readers

AI infrastructure 4 of 4

Self-hosted where it pays, managed where it doesn’t.

vLLM with Qwen behind LiteLLM instead of a vendor bill; PocketBase and Supabase as the system of record.

  • Docker on small servers, Cloudflare in front, no single hand-made box
  • Logs and alerts from the first deploy, capacity watched per key
  • Documented so that someone else can run it on day one

What it changed

Measured in production at my last company. Names stay with the owner; the numbers are mine to show.

Selected work

Seven pieces of work. Pick one and see the path it took through the stack.

  • Case study
  • Case study
WebhookShop, ticketing, mailboxScheduleCron, pg_cronOrchestratorn8n, WindmillQueueJobs table, batchesAgentStructured output, toolsLookupERP, catalogueExtractJSON schemaWriteRecords, emailsLLM gatewayLiteLLM, keys and limitsModelQwen on vLLMMCP serverFastAPI over PocketBaseDatabasePocketBase, Supabase, PostgresHostingDocker, small serversEdgeCloudflareObservabilityRun logs, alerts

Toolbox

Workflows
n8n, Windmill, Python, TypeScript, REST APIs, Power Automate, GitHub Actions
AI agents
Claude Code, LLM integration, RAG, Structured output, Botpress, Retell AI, ElevenLabs
AI enablement
MCP connectors, LiteLLM gateway, Skills and templates, Decision papers, Workshops
AI infrastructure
vLLM, Qwen, Docker, PocketBase, Supabase, PostgreSQL, FastAPI, Cloudflare, Vercel
Data
Power BI, SQL, Data modelling, Anomaly detection, Odoo, plentymarkets