th3chris
AI Engineering

AI Engineering & AI Solutions

RAG, MCP, AI agents — productively integrated, not just prototyped.

Coming from 25+ years of distributed systems, I build AI-native architectures that survive real production environments. Not 'the demo works' — but latency, cost, maintainability, and compliance handled simultaneously.

My approach

Most failed AI projects didn't have bad models — they had no architecture. Three principles I run projects by:

Modular over monolithic — LLM, retrieval, and tooling as swappable building blocks. That way the system survives the next 18 months of model generations.

Observable, not a black box — tracing over every inference, token cost per request, eval pipeline for regressions.

Conservative before magic — RAG before fine-tuning, tool use before multi-agent, deterministic before LLM whenever possible.

Capabilities

Production

RAG Systems

Document ingestion, chunking strategies, hybrid search (vector + keyword). Running productively with Dify + Weaviate on my own Kubernetes cluster — this website itself. Hands-on with pgvector and Qdrant.

Production

MCP Servers & AI Tooling

Multiple MCP servers in production — including enterprise DataHub queries at the Hoffmann Group. Connecting Claude, GPT-4, and local models to real backend tools.

Production

AI Agents & Workflow Automation

Custom AI agents that autonomously take on development tasks and communicate with the team through the ticketing system — real human-machine integration. Automated code reviews across GitLab, GitHub, and Azure DevOps.

Core

LLM Integration in Enterprise

OpenAI, Anthropic, Azure OpenAI plus open-source models. Model selection, cost optimization, fallback chains. Fine-tuning evaluated (LoRA, OpenAI) and applied selectively where RAG isn't enough.

DevOps

AI-Native Infrastructure

Kubernetes-native AI workloads, GitOps for ML pipelines, model versioning. Self-hosted on my own Kubernetes cluster — this website plus Dify plus Weaviate is the running example.

AI in production — real projects

th3chris.com — RAG-powered Portfolio (Live Demo)

This website itself: Dify-based RAG system with Weaviate vector DB, in-house knowledge base, GPT-4, and streaming responses. Self-hosted on a personal Kubernetes cluster with GitOps.

RAGDifyWeaviateKubernetesGitOps

Hoffmann Group — AI Integration in Enterprise DataHubs

The MCP server made the DataHub backend directly accessible to AI tools for the first time — developers can now formulate queries in natural language instead of memorizing GraphQL schemas. The AI assistance in the Playground noticeably accelerates onboarding of new use cases and consumers.

.NETAzureMCPAI Integration

Human-Machine Integration — AI Agents in the Dev Workflow

Custom AI agents that handle development tasks fully autonomously and communicate with human teammates through the ticketing system (GitLab). Combined with automated code reviews across GitLab, GitHub, and Azure DevOps. Bridge technology between classical microservice landscapes and LLM-driven tooling.

AI AgentsGitLabGitHubAzure DevOpsCode ReviewAutomation

When AI isn't the answer

Mastery means knowing when to leave AI out. Four typical cases where I recommend classical solutions or a human in the loop:

When determinism matters — accounting, compliance, billing. Rule-based systems beat LLMs in any audit.

When the corpus fits into the context window — passing all documents into the prompt directly is simpler, more deterministic, and more provable than RAG. Vector DB + embeddings only pay off once a full dump becomes too expensive in tokens.

When the output isn't verifiable — LLM responses without an eval pipeline and guardrails are a compliance risk, not a feature.

When the consequence is irreversible — medical diagnoses, compliance decisions, financial approvals, safety-critical control. AI belongs in these workflows as a proposer and accelerator, not as the final decision-maker. A human in the loop here isn't a bottleneck, it's a prerequisite: accountability has to remain delegable — and LLMs answer confidently even when they're wrong.

Sounds like your project?

A few targeted questions instead of a rigid form — I'll understand what it's about in 2–3 minutes and get back to you personally.