AI infrastructure is the layer every AI feature in a company stands on: one access point for cloud and local models, cost control, a log of every call and anonymisation of data before it reaches a model. I build it so that swapping a model is a setting, not a rebuild.
Choosing a model should be a setting, not an architectural commitment for years.
Who it is for
- Companies with several AI features, each calling models differently
- Companies that don't know what their AI costs or who uses it
- Companies that want to combine cloud and local models
If you have one AI feature and one provider, you don't need infrastructure yet.
How it works
- I map where and how you call models today, what it costs and what data goes there.
- I design one layer for model access, logging and limits.
- I build it and move existing features onto it step by step.
- I set up a cost and quality overview and a procedure for swapping models.
What you get
One access to models
OpenAI, Anthropic and local models via Ollama behind one interface.
Cost control
Limits and a cost overview per feature and team.
Call log
What went into the model, what came back and what it cost.
Anonymisation
Names, national ID numbers and IBANs are replaced with tokens before the model.
Queues and background processing
Long tasks run in the background and resume after an outage.
In practice
AGL Assistant runs as a NestJS background worker with PostgreSQL and Redis and processes mail from every inbox. ClaimPilot replaces personal data with tokens before every model call and runs entirely on local models, where measurement decided the model choice.
What changes in the company
You know what AI costs and who uses it. You can try and deploy a new model without rewriting applications.
Technology and data security
Data can stay on your server, personal data is anonymised before the model and every call is logged.
Price
€1,800–4,500, 2–6 weeks. Pricing
FAQ
What is AI infrastructure?
A shared layer for all AI features: model access, costs, logs, anonymisation and queues. Features built on it don't each have to handle which model they call and how.
Do we need it if we have one chatbot?
Usually not. It pays off when you have several AI features or want to combine cloud and local models.
Can you use local and cloud models at the same time?
Yes. Sensitive data can go to a local model and the rest through an EU-region API.
How much does it cost?
€1,800–4,500 depending on how many features are connected, 2–6 weeks.
Related
- Services
- Local LLM
AI on your own server. Data stays in the company.
- LLM evaluation and monitoring
A test set and numbers that show whether your AI really works.
- AI agents
The agent drafts, a person approves.