Cases

Three problems. Three production APIs.

A sample of what we've shipped: real client problems, turned into working AI systems on isolated infrastructure. Details anonymized, results are not.

Public sector analysis

A heat-transition intelligence workspace for municipal policy

40 Dutch municipal heat programmes were published as separate PDF documents, each with its own structure and language. Comparing solutions across municipalities, or simply answering a question about one, meant manually opening dozens of files. There was no single place to search or compare them.

What we built

  • Agents ingested all 40 programmes into a knowledge graph with a comparison matrix across municipalities
  • A living report on policy instruments and heat solutions that stays current as new documents arrive
  • Every claim in the report links back to the source document it came from
  • A public share link so anyone can ask questions in plain language and get sourced answers in Dutch

40

programmes made queryable

65+

linked knowledge files

seconds

to a sourced answer

Consumer app

An AI coaching backend for a golf app

A golf coaching app wanted to give players real insight from their own round and shot data, not generic tips that could apply to anyone. That meant grounding every AI response in a specific player's history, on infrastructure isolated from the rest of the platform.

What we built

  • A full AI backend on isolated infrastructure, separate from every other client
  • Knowledge-graph retrieval over each player's own rounds and shots
  • A two-tower score prediction model, trained per player and per hole
  • AI coach reports and social features, all shipped as authenticated APIs the mobile app calls directly

5+

production AI endpoints

per-player

grounded insights

live

score prediction in the app

Creative platform

An AI image generation platform for a photography company

The client needed studio-quality, on-brand product and campaign imagery generated on demand, at a scale no photo shoot schedule could match. Anything shipped had to look like it came from their brand, not a generic model output.

What we built

  • A dedicated image-generation platform running the latest diffusion and multimodal models
  • Brand-identity distillation straight from a store URL, no manual brand brief required
  • Parallel generation with automatic retry and safety handling on rejected prompts
  • Storage with signed URLs, with every call metered individually

production

API serving a real customer base

multiple

image models, one endpoint

per-call

cost control

Your case could be next

Tell us the problem. We design the system with you and ship it as a production API, the same way we did for these three.

Schedule an intro