Skills & Stack

What I actually use in production — across AI pipelines, backend integrations, infrastructure, and business operations.

🇩🇪 Deutsch 📅 Book Call
I build automation with code faster than others build it with drag-and-drop tools.
AI

Automation & AI Agents

Google ADK (Agent Dev Kit) Agentic IDE & MCP Orchestration Claude & Gemini Agent Workflows MCP Tooling (Google Ads, GA4, Postman) Two-Layer Agent Shields (Guardrails) Model Context Protocol (MCP & WebMCP) Multi-Agent Orchestration & A2A Hybrid RAG Architecture Agent Memory & State Management Code-first Automation IDE-as-Control-Plane Workflows LangGraph LangChain Vertex AI Search & RAG Deterministic + LLM Pipelines Business Process Automation
SEC

AI Security & Governance

Google Cloud Model Armor Two-Layer Agent Shields (Guardrails) Prompt Injection & Jailbreak Defense Cloud DLP & PII Masking (IBAN Anonymization) Agent Governance & IAM Least Privilege GDPR / DSGVO Zero-Trust Architecture Deterministic Safeguards in Code Zero-Leak Token & Secret Management Audit Logging & Cloud Trace Input & Output Sanitization Prompt Drift & Hallucination Mitigation Human-in-the-Loop Safeguards
LLM

Multi-LLM Orchestration

Hierarchical Model Delegation Tiered Routing: Architect → Coder → Evaluator Multi-Provider Routing & Fallbacks (OpenRouter) Cost-Conscious Inference (~90% API Cost Cut) Model-Agnostic System Design Claude + Gemini Multi-Agent Collaboration Local & Edge Inference (Ollama / vLLM) Multimodal Processing (Vision, Audio, Docs) Structured Output & Function Calling Latency & Token Budget Optimization
DEV

Backend, APIs & Data

Python 3.12+ (FastAPI, Asyncio) Google Cloud Firestore (NoSQL) PostgreSQL & Supabase (pgvector) Vector Databases & Embeddings Sentence-Transformers & HuggingFace Pydantic (Strict Validation) Pandas & Data Pipelines Postman (API Testing & Mocks) TypeScript & Node.js REST APIs & Webhooks SQL Migrations & Optimization System Integrations Git / GitHub Actions CI/CD Playwright (Headless Automation)
OPS

Cloud & Infrastructure

Google Cloud (Cloud Run & Build) GCP Secret Manager & IAM Supabase / pgvector Self-hosted Linux VPS (Frankfurt) Docker & Containerization GitHub Actions CI/CD Nginx & SSL / Reverse Proxy GCP MCP Servers Google GenAI SDK Railway / Vercel Redis (Caching & Limits) n8n & Make.com
BIZ

Business & Operations

B2B Operations Business × Tech Bridge Process Optimization German Compliance & GoBD Supply Chain Procurement Stakeholder Management P&L Ownership Client Negotiations SEO Meta Ads Google Analytics 4
WEB

Web, CMS & E-Commerce

WordPress & WooCommerce E-Commerce Online Stores Business Card & Landing Pages AI Chatbots & RAG for WordPress React 19 & Tailwind CSS Technical SEO & Core Web Vitals Custom Themes & Plugins Canvas API & Vite HTML5 / CSS3 / JavaScript Conversion Rate Optimization
LNG

Languages

English — Fluent (Professional) German — B1 (Certified, Integrationskurs) Russian — Native Ukrainian — Native
Tools I work in daily VS Code + Claude Code — code-first development. Not no-code platforms. The real stack matters more than the tools list.
See it in projects →
Engineering Methodology

Code-First Architecture Over Fragile Visual Builders

Deterministic reliability in enterprise production: While visual no-code platforms (Make, n8n, Zapier) excel for simple triggers, enterprise AI automation requires robust error boundaries, strict type safety, and reproducible version control. In production systems, I build with Python 3.12+, FastAPI, and Google Agent Development Kit (ADK). Every external API payload is strictly validated via Pydantic v2 schemas before execution, eliminating silent failures and unpredictable LLM outputs.

Two-Layer Agent Shields & Zero-Trust Governance: Autonomous LLM agents connected to company databases introduce injection and data leakage risks. My systems enforce a two-layer defense perimeter: an outer deterministic gateway (Cloud DLP tokenization, IBAN anonymization, regex sanitization) and an inner model governance layer (Google Cloud Model Armor, strict Tool Choice constraints, and IAM least-privilege service accounts on Google Cloud Run in Frankfurt). Sensitive data never enters external LLM contexts.

Observability & Orchestration

Multi-Agent Systems, FastMCP Tooling & Real-Time Tracing

Hierarchical model delegation: High-performance AI automation is never single-model. Complex tasks are broken down using tiered routing: high-reasoning models (Claude Code, Gemini 2.5 Pro) serve as architects and task delegators, while fast, lightweight endpoints (Gemini 2.5 Flash, local vLLM) handle real-time execution, data transformation, and vector search. This architecture slashes API inference costs by up to 90% while maintaining sub-second response times.

Production observability (4 Pillars): Every agent workflow is instrumented with structured JSON logging correlated to Cloud Trace via W3C traceparent headers. Distributed OpenTelemetry spans monitor tool call latencies, token consumption, and retrieval accuracy, allowing continuous evaluation and immediate bottleneck isolation.

Technical FAQ

Frequently Asked Questions: Technical Stack & Implementation

Why choose custom Python code over no-code drag-and-drop automation tools?
Visual drag-and-drop tools lack comprehensive regression testing, native distributed tracing, fine-grained secret management, and deterministic error handling. Code-first automation with FastAPI and Google ADK allows seamless Git CI/CD, automated integration testing, and complete independence from proprietary third-party platform lock-in.
How do you ensure GDPR / DSGVO compliance for European clients?
All server workloads run exclusively within the European Union (Google Cloud Run in Frankfurt am Main, europe-west3). Customer records and personal identifiers are scrubbed via automated Cloud DLP privacy proxies before reaching LLMs. We maintain contractual zero-data-retention guarantees with enterprise AI providers.
What role do Model Context Protocol (MCP) and WebMCP play in your architecture?
MCP provides standardized, type-safe communication bridges between AI agents and external tools (Google Ads, GA4, Postman, PostgreSQL). WebMCP extends this capability into the browser, enabling web applications to expose client-side tools and structured state directly to visiting AI agents via JSON-RPC 2.0 without screen scraping.
How do you prevent hallucinations in customer-facing RAG sales systems?
Hallucinations are prevented using hybrid vector retrieval with Vertex AI Search, strict Pydantic argument validation, and forced Tool Choice modes (ANY / AUTO with deterministic fallbacks). The agent answers exclusively from verified company knowledge bases and explicitly declines unverifiable assertions.
Can these AI pipelines integrate into existing ERP and CRM systems?
Yes. Every automation pipeline exposes clean REST webhooks and cloud event triggers, allowing effortless bi-directional synchronization with DATEV, SAP, Salesforce, HubSpot, Lexware, and custom PostgreSQL databases.