AI in 2026: Build vs Buy — Two Products, One Honest Answer
From Genk, two AI products live in 2026: ImmoStory.ai (we built the model) and tamam.be (we bought the embedding API). Here's the honest build-vs-buy reasoning, with the architecture and the parts that broke.
Why this article exists
Probex is a small AI studio from Genk, Belgium. We build consumer products and we build software for other companies. In the first months of 2026 we started two products — ImmoStory.ai in late January, tamam.be in May. Both run on AI every day. They took different paths. This page is what we learned, written for the founder who needs to make the same call.
We are not a vendor. We are not selling a stack. We are telling you what we built, what we chose not to build, and when each side of the build-vs-buy call actually pays off.
Company at a glance
Founded in May 2024 in Genk (C-Mine). Two live products. One AI engineering team. We design, ship and operate the kind of AI we wish someone had built for us.
At a glance — what we ship and where it lives
ImmoStory.ai — we built it
Real estate agents paste a property URL. Within 90 seconds they get a 30-second vertical video, a description in five languages, and a clean listing. We self-host the vision model, fine-tune the LLM adapter, and run a render farm in the same datacenter. No per-call API fees after the first month.
What this means for you
If you are a real estate agency, brokerage or portal: paste a listing, get a video ready for TikTok, Instagram Reels and YouTube Shorts, in any of our five languages. Branding is yours. The output is yours. You are not renting a closed SaaS — you can walk away with your own model weights.
Step by step
- Scrape + clean — a single Playwright job pulls the page, strips noise, and writes a structured JSON payload (price, sqm, photos, energy label).
- Vision + LLM — the Vision model picks the 6–8 best photos, the LLM drafts copy in the target language and tone.
- Render — a Remotion headless worker composes the video, brand-aware, fluent in five languages.
- Publish — output is uploaded to S3-compatible storage with a signed URL, ready to share on social.
tamam.be — we bought it (and glued it)
Tamam.be is Belgium's home services marketplace: homeowners compare verified local professionals, quotes and reviews. We did not build a custom model for the matching engine. A few hundred matches a day, latency matters, and a general-purpose embedding API is already good enough. We used a hosted vector API, a managed Postgres for ranking, and a small NestJS service to glue them. The data is ours, the API is replaceable, the operating cost is predictable.
What this means for you
If you are a tradesperson or handyman in Belgium: tamam.be is a free channel to local customers. If you are a homeowner: it is a way to compare quotes without giving your details to five call centres.
Why we bought instead of built
We could have built a custom ranker. The math didn't work. Below 500 matches/day the hosted embedding API is cheaper than the GPU we'd need to fine-tune. Below 1,000/day the quality difference is invisible to the end user. We built where the build pays off, we bought where the buy does.
When to choose each path
- Build when you have more than 1,000 events/day, the domain is proprietary, or per-call cost exceeds 0.05 EUR.
- Buy when the workload is below 500/day, the model is general-purpose, and quality is already good enough.
- Glue the two: a small owned service that calls the API and keeps the data model under your control.
The 8 things that actually broke
- Vision model hallucinated room counts on wide-angle photos. Fix: constrain to architectural features.
- Translation tone felt robotic. Fix: a small style fine-tune on locally-gathered descriptions.
- Render queue got stuck when a worker died mid-job. Fix: lease + heartbeat, not TTL.
- First-pass signed URLs leaked across tenants. Fix: keyed prefixes per tenant.
- Scraper triggered CAPTCHAs at scale. Fix: residential proxy pool + backoff.
- Embeddings drifted between model versions. Fix: pinned version + reindex on upgrade.
- Latency budget exploded when we added re-rendering. Fix: cache by (input-hash, parameter-set).
- Cloud bills ballooned because nobody watched the queue. Fix: dashboard + alerting from day 1.
How to start — the 30-day plan
- Week 1 — pick one workflow that already costs you time and money. Don't start with a flagship.
- Week 2 — call a vendor API. Time it. Measure the cost. Decide build vs buy with data.
- Week 3 — glue it into your own service. Keep ownership of the data and the queue.
- Week 4 — add a small in-house model only where the API is too expensive or too generic.
Our products, in plain language
ImmoStory.ai turns a real estate listing URL into a 30-second vertical video, in five languages, ready for social. We built the model, we run the render farm, you keep the brand and the output.
tamam.be is Belgium's home services marketplace. Homeowners compare verified local professionals, quotes and reviews. Trades sign up free and get a steady flow of local leads.
Want a hand?
Probex designs, builds and operates AI you actually own. Tell us the workflow — we'll come back with a build-vs-buy recommendation and a working prototype within two weeks.
Press: Our work was featured in Het Belang van Limburg (Aug 2026) — "Schooldirecteur richt AI-bedrijf op dat promotievideo's in minuten maakt".