AI Agent Developer · Germany & International
Your team does the same work every day. I build AI agents that handle exactly that.
Reviewing documents, entering data, chasing approvals — reliably automated, without supervision. For companies in Germany and internationally.
About
I'm Ali Parnan. I build AI agent systems that take over real work — not demos, not chatbots.
For the past 8 years I've been building software products and leading engineering teams. The last two years I've focused almost entirely on AI agent systems — autonomous pipelines, MCP infrastructure, and production-grade orchestration with LangGraph.
My approach: agents should work independently and only involve humans when genuinely uncertain. I design for observability, correctness, and trust — because a system that runs unsupervised needs to earn that right.
Most of my projects are open source. You can read the code, see how I think, and judge for yourself whether the approach makes sense.
How I Build Agent Systems
Principles from 2 years of building agents that actually run in production.
Agents have behaviour, not features
You can't fully specify agent behaviour upfront. You observe it, shape it, and improve it over time. I design systems with this in mind from day one.
Human-in-the-loop only when uncertain
The goal is not to involve humans often — it's to involve them at exactly the right moment. When the agent is unsure, it asks. Otherwise it acts.
Observability is not optional
Every agent action is logged. Every decision is traceable. A system running without visibility is a system you can't trust — and can't improve.
Domain knowledge as editable skills
Rules, tax codes, account charts — these shouldn't be hardcoded. I model domain knowledge as Markdown-based skills the agent reads at runtime. Editable without redeployment.
Supervisor loops, not hope
Every production agent system needs a background supervisor: retry failed steps, detect stale documents, validate outputs. Reliability is designed, not assumed.
MCP for integrations, not custom glue
Model Context Protocol is how agents should connect to external systems — standardised, auditable, policy-controlled. I've built MCP infrastructure across multiple projects.
Read: Securing MCP Infrastructure →What I Build For You
Concrete AI agent systems for real business problems — DACH and international.
Autonomous document processing pipelines
OCR → classification → booking → validation — fully automated. The agent processes your documents, creates structured output, and only asks when genuinely uncertain. Like Workless does for accounting.
MCP infrastructure & agent management
Connect your existing tools to AI agents via MCP — with policy control, audit logs, and access management. No vendor lock-in, no black boxes. Based on AgentxSuite architecture.
German market AI agents (DATEV, GoBD, VAT)
Agents built for the German regulatory landscape — DATEV EXTF exports, GoBD-compliant audit trails, VAT ID validation, SKR03/SKR04 chart of accounts. Domain knowledge built in, not bolted on.
AI agent architecture consulting
Already building with AI but hitting walls? I review your architecture, identify where agent behaviour breaks down, and show you how to make it reliable. LangGraph, LangChain, or custom orchestration.
What I've Built
Real systems. Judge the work yourself.
Willow
A personal AI agent that runs locally on your machine. Browser and desktop automation, local vector memory, an Obsidian-compatible vault, and your own models — Anthropic, OpenAI, Gemini, or fully offline with Ollama. Your data never leaves the device.
Visit Willow →Workless
An AI-powered accounting agent for the German market. Upload receipts and invoices — OCR extracts the data, AI classifies and books them, a supervisor loop retries failures, and DATEV EXTF exports land at month end. Human review only when the agent is unsure.
View on GitHub →AgentxSuite
A control plane for AI agents. Connect agents to multiple MCP servers, define access policies, monitor every action in real time, and maintain full audit trails. Built for teams that need AI agents in production without losing control.
Visit AgentxSuite →Vispra
A modern HTML5 client for Xpra built with SolidJS and TypeScript. Renders remote application windows directly in the browser over WebSocket — no plugins, no installs.
View on GitHub →Handwerkfix.de
All-in-one SaaS for German tradespeople. Invoicing, quotes, scheduling — replacing paper-based workflows in a traditional industry.
Visit Handwerkfix →Experience
AI Agent Developer – Independent
Building AI agent systems with LangGraph, FastAPI, and MCP. Projects include Willow (local-first desktop AI agent), Workless (DATEV accounting agent), and AgentxSuite (MCP management platform).
CTO – amitego
Leading engineering for secure remote access and privileged access management software.
Head of Product – Orangery Holding GmbH
Built OrangeryOS, an enterprise office management platform. Led a cross-functional team of 20.
Owner – International IT Outsourcing
Custom software development and consulting for clients across Europe.
Writing on agents
Longer notes on MCP, team structure, and local AI.
Have a project that needs an AI agent?
Tell me what the process looks like today and what you'd want it to do on its own. I'll tell you honestly whether an agent makes sense — and how I'd build it.
info@aliparnan.com