We train our own AI models — and why that matters for your business

Most consultancies resell API keys. We train our own models.

There is a quiet difference between an AI partner who wires up someone else's API and one who trains, fine-tunes and operates models in-house. The first rents you intelligence and bills you forever. The second builds capability into your company — and hands you the keys. At Probex we deliberately chose the second path, and this page explains what that buys you.

What we actually run in-house

• Language models for tone and domain — a fine-tuned LLM adapter powers the property copy in ImmoStory.ai. Generic APIs wrote "spacious family home with lots of light". Our model writes copy that sounds like the best local agent.
• A self-hosted assistant — the chat assistant on this very site runs on our own Ollama infrastructure in Belgium, on a model we selected, prompted and tuned. No per-message cloud bill, no data leaving the country.
• Vision for photo selection — the model that ranks listing photos in ImmoStory.ai was trained on what actually sells a property, not on generic aesthetics.
• Extraction and enrichment pipelines — structured data out of messy sources: listings, catalogs, documents — feeding clean inputs to everything above.

Why it matters for your business

Rented AI (API resell)Per-call costs foreverData leaves your controlGeneric tone and domainVendor changes = your outageOwned AI (Probex way)Fixed infrastructure costData stays in BelgiumYour tone, your domainYou keep the weightsThe same reasoning we apply before every build: buy where it pays, own where it matters.

The honest tradeoff

We do not train everything. Where a hosted API is genuinely cheaper and better — low-volume embeddings, exotic languages — we buy, and we say so. The skill is knowing which side of the line each workload belongs on. That judgment is the product; we published our exact decision rules in the field guide linked above.

What this looks like for you

When Probex builds for you, the models that matter run on infrastructure you can point at, tuned to your voice and your data, at a cost you can predict. When the project ends, they are yours. That is what "AI you actually own" means — it is not a slogan, it is an architecture decision.

Have a workflow in mind? Tell us about it — within two weeks you will have a build-vs-buy recommendation with real numbers and a working prototype.