Inference API
OpenAI-compatible endpoint for open-weight models such as GLM and DeepSeek. Prompt caching for agentic workloads. Billed per token.
IT services from Vienna
AI Inference from Vienna, Sovereign LLMs, Self-Hosting.
I run open-weight language models on my own hardware in Austria and provide them to business customers. Companies that want to run AI themselves get my help building it.
01 / Compute
AI inference and GPU compute.
For companies that want to use open-weight models and must keep their data in the EU.
Dedicated GPU capacity
NVIDIA Blackwell servers (HGX B300) in Vienna, planned from early 2027. Reserved for 12 to 36 months.
EU data residency
Hardware, operations and data stay in Austria. Prompts and responses are not stored. Data processing agreement under the GDPR.
Running today
My own GPU server with two NVIDIA RTX PRO 6000 Blackwell cards serves GLM 5.3 Flash at about 200 tokens per second per request.
Export control and customer screening
- The GPU hardware is subject to US export controls (EAR, ECCN 3A090 and 4A090) and the EU Dual-Use Regulation. I comply with both.
- Every contract customer is screened before the contract starts: headquarters, ownership, intended use, and a check against EU and US sanctions lists.
- I provide no services to companies headquartered or ultimately owned in EAR Country Group D:5 or Macau, and none for military or weapons-of-mass-destruction purposes.
- The hardware is not resold. Usage is logged per contract (who, when, which capacity), without content.
02 / Services
Consulting: five focus areas.
Everything listed below, I run in production myself. No PowerPoint, no reseller markup.
AI Infrastructure
Inference on your hardware. Designed, deployed, monitored.
A single GPU server or a multi-node cluster. Model routing, auto-scaling, latency and token-throughput monitoring.
- vLLM
- SGLang
- LiteLLM
- Multi-GPU
Sovereign LLMs
Open-source models on-prem. Audit-ready, optionally air-gapped.
Llama, Mistral, Qwen, DeepSeek: hosted, routed, documented. GDPR- and NISG-2026-compliant.
- GDPR
- NISG 2026
- On-Prem
- Air-Gapped
RAG & Semantic Search
Knowledge retrieval over your own data. Provisioned and operated.
Embedding pipelines, vector stores, re-ranking, eval loops. For internal docs, codebases, support tickets, running on your infrastructure.
- pgvector
- Qdrant
- BGE / E5
- Re-Ranking
AI-Augmented Development
Claude Code & co.: introduced, guard-railed, measured.
Training, workflow integration, review pipelines that keep AI output honest. My standing rule: a claim doesn't count until the agent proves it with a command.
- Claude Code
- Codex
- Reviews
- Training
Self-Hosting
SaaS out, your own servers in. Reproducible and documented.
Mail, identity, monitoring, backups, CI: one stack, no vendor lock-in. Proxmox as the foundation under all of the above.
- Proxmox
- Vault
- Stalwart
- Wazuh
03 / About
Stefan Starflinger.
Independent, technical, hands-on.
Self-employed in IT services since 2018. I run the same stack I recommend: a 4-node Proxmox cluster, a GPU server with two RTX PRO 6000 cards, sovereign LLMs behind LiteLLM, self-hosted CI runners, monitoring, mail and vault servers.
Sole proprietorship based in Vienna. Consulting in German and English.
04 / Contact
Tell me briefly what you have in mind.
I respond within one business day.