In DevelopmentPersonal Automation · AI Infrastructure
AI Command Center
The Problem
Property income and expenses across multiple units. WEG statements. Bills, receipts, and documents across dozens of categories. Eight years of self-filing complete German tax returns. All of it important. All of it repetitive. All of it eating my weekends.
The Approach
Hybrid architecture — local Ollama models running on-device for privacy-sensitive document processing, Claude API reserved for tasks requiring genuine reasoning. Seven specialist agents governed by a machine-readable Constitution that enforces a hard monthly budget cap and privacy rules.
Seven agents:
· Supervisor Orchestrator — routes tasks, enforces Constitution, tracks spend
· Tax Consultant — German tax law, Finanzamt documents, Elster filing
· Financial Planner — portfolio tracking, ETF rebalancing, P&L reports
· File Manager (The Librarian) — document classification, Google Drive organisation
· Health Agent — reimbursement claims, health document management
· Innovation Agent — AI news digest, research, idea capture
· Business Operations Agent — invoice management, scheduling, client documents
Routing logic: Sensitive documents processed locally via Apple Vision OCR — raw files never leave the device. Only anonymised values reach the cloud API. Cost governance enforced in real time.
The Result
60–70% of tasks handled locally at zero cost. Total monthly AI infrastructure targeting under €30. GDPR compliant by design. System runs autonomously — agents operate without user intervention.
Stack
Mac Mini M4 Pro (24GB)OllamaQwen3 14BGemma 3 12BClaude HaikuClaude SonnetPythonSupabaseApple Vision OCRHetzner VPS FrankfurtGoogle Drive API