Taming the Beast: Building Autonomous Agents to Solve German Tax Bureaucracy

German accounting is notoriously complex, running on a “deprecated backend” of physical paper, rigid bureaucratic rules, and a shrinking workforce of tax advisors—30% of whom are retiring soon. For developers, this presents the ultimate “legacy refactoring” challenge: How do you replace a manual, high-stakes human process with an autonomous agent without causing legal failures?

In this session, we dive into the architecture of Mika, an AI financial autopilot that has moved from 70% to 95% automation. We will move beyond the hype to examine the specific engineering patterns used to solve a domain where a 0.1% error rate is unacceptable.We will dissect three core components of our architecture: - The Multi-Agent Document Swarm: How we orchestrate specialized agents (Classifiers, Validators, and Knowledge Base RAGs) to navigate the nuances of German tax law better than a human. - The “Todo Firewall”: A self-learning feedback loop that utilizes structured memory to determine when to disturb the user and when to autonomously overrule strict validation errors. - Solving the O(N^2) Matching Problem: A deep dive into our 3-layer transaction matching engine. We will show how we combine classical ML ensembles with semantic LLM agents to match thousands of invoices to bank transactions with a 91.8% F1 score—all while adhering to strict Serverless memory constraints.

Join us to learn how to build “Agentic Workflows” that are not just chatty, but capable of handling the heavy lifting of a real-world complex use case.