I use cookies and tools to provide this website and — with your consent — to analyse how it is used (e.g. Google Analytics, Google Ads). More about this in the privacy policy.
n8n · LLM integration · MCP · AI governance
Bot check failed, please try again.
Get in touchWord version available on request
agents that handle tasks on their own up to an approval step
Ten AI agents run my business operations with defined roles: research, content, sales, accounting, administration. They read mailboxes, create tickets, draft texts and report in Slack. Nothing is published or sent without approval.
88 active workflows in production: scheduled jobs, webhooks, IMAP inboxes, Slack agents and reusable sub-workflows. Every job has a runbook, failed runs go to monitoring.
connecting language models to existing applications
Claude API via the Anthropic SDK, a self-hosted LiteLLM gateway, Symfony AI. Every call is logged with model and token usage. The public chat assistant on jacomeit.de has been in production since July 2026.
Model Context Protocol
My own open-source MCP servers: lexware-php-mcp (PHP, official MCP SDK) for Lexware Office and mite-mcp-server (TypeScript) for the mite time tracker. Day to day, my agents reach Jira, Slack and the knowledge base through MCP.
Transparency obligations under Art. 50 of the EU AI Act implemented on the day they took effect (2 August 2026), on four layers. The chat assistant's guardrails are available as an MIT-licensed Symfony bundle: injection prefilter before every paid call, kill switch, budget and rate limits, retention periods.
Integrations with accounting (Lexware Office), time tracking (mite), ticketing (Jira), Slack and custom REST APIs. Asynchronous processing over message queues, results written back into the existing process.
AI implements tickets up to the pull request
A custom development pipeline has agents implement tickets in isolated containers, runs the quality gates and opens pull requests with green CI. 17 pull requests merged since 16 October 2026, each after review and approval.
| Project | What it shows | Evidence |
|---|---|---|
| AI team in my own business (since July 2026) | AI agents, n8n, MCP, approval workflow | 10 agents, 88 active workflows, 600+ tickets since July 2026 |
| Operations cockpit (since September 2026) | agent API with permissions, AI cost per ticket in euros, one-click updates | Node.js, TypeScript, Fastify, PostgreSQL; 35 releases |
| Autonomous development pipeline (since October 2026) | agentic coding with a review cascade | 17 merged pull requests |
| GuardedChatBundle (open source) | LLM guardrails for Symfony | github.com/Stefan39/GuardedChatBundle, MIT |
| lexware-php-mcp, mite-mcp-server (open source) | own MCP servers in PHP and TypeScript | GitHub Stefan39, MIT |
| Marketing digest with the Claude API (since June 2026) | LLM integration in Symfony, token logging, cost estimate | in production on jacomeit.de |
| Quality assurance | AI-written code passes the same gates as hand-written code | 4,147 tests, over 99% test coverage, PHPStan level 8 |
The AI work rests on 25+ years of backend development. My three most recent engagements:
Energy provider · February to June 2026
Event-driven architecture with Symfony Messenger, step-by-step replacement of PHP legacy code, team of 3–5.
Stack: Symfony, PHP, Symfony Messenger, CQRS, async command processing
Shipbuilding · February 2024 to December 2025
Step-by-step migration of a legacy PHP system to Symfony, partly in a classified environment, team of 11.
Stack: PHP, Symfony 5+, Oracle, RabbitMQ/AMQP, JavaScript/ExtJS
Public-sector platform · September 2021 to March 2023
Symfony API with a React frontend, backend focus, team of 8–9, test coverage consistently above 98%.
Stack: Symfony 5+, PHP 8, CQRS, Symfony Messenger, JWT, PHPUnit, Docker
My focus is integrating and operating AI in business software, not building models: no model training, no data science. If a role is mainly Python ML, I am not the right fit. On client projects I work AI-assisted, including with Microsoft Copilot.
Send me the project description. You will get a reply within one business day with my assessment and availability.
Bot check failed, please try again.