Legacy systems onto Kubernetes.
AI prototypes into production.

Freelance software architect for cloud platforms and AI systems

I build the architecture and stay with it until your team can run it alone.

Book a call Fixed-price packages

16+ years · Government, healthcare, public transit · Passau and remote

Why LLM prototypes never reach production

The demo works and leadership wants it live. Then legal stops the project. Usually one of seven things is missing.

No evaluation

A prompt or model change makes quality worse. Nobody notices.

No source grounding

The system invents citations. You cannot prove where a claim came from, and that is what compliance stops.

No observability

A bad answer cannot be reconstructed. Cost per request is unknown.

Naive retrieval

Chunking by character count, vector search only, no reranking. Hallucinations are then expected.

No context management

Long runs die mid-way. The token budget is unchecked.

No operating model

Unversioned prompts, no pipeline, secrets in the repo.

No tenant isolation

It works for one pilot. Not for twenty customers.

What you hire me for

Usually it is one of three jobs.

Legacy onto Kubernetes, no downtime

Target architecture, cutover, handover. The migration is done when your team deploys on its own.

Nationwide case-file system, WebLogic to Quarkus on Kubernetes. Zero downtime. Rollout from one day to 20 minutes.

See the migration assessment

LLM feature from prototype to production

Multi-agent systems, hybrid RAG with source grounding, evaluation and observability. Without that, the prototype stays a demo.

Audit assistant for certification (beta) and a RAG chatbot on a healthcare platform (in production).

See the LLM readiness check

A platform your team can run

CI/CD, infrastructure as code, secrets and monitoring. After that you should not need an external architect.

Public-transit platform with 20+ tenants and fully automated deploys. SBOM, Vault, DefectDojo in a government setting.

Cloud and Kubernetes

Zero downtime

Migration of a nationwide government application

1 day → 20 minutes

Rollout time after the migration

20+ tenants

on the migrated public-transit platform

4 years

my share of the government project

What clients say

Our J2EE application now runs on Kubernetes. On time, no downtime. In a government shop that is the part that fails when someone does not know operations.

Project Lead, Federal Agency

We had no CTO. Daniel built the architecture and got the team far enough that we run the stack ourselves.

CEO, AI Startup

The chatbot runs on 500+ video and PDF items. RAG pipeline and MCP server. The answers hold up.

Product Manager, Healthcare Platform

From the blog

Short pieces from real work. Often about the spots where an LLM prototype gets stuck.

September 4, 2026

Hybrid RAG and reranking instead of naive vector search

Character-count chunking and vector-only search make hallucinations expected. What hybrid search and reranking change, and what the older RAG article taught wrongly.

Read article