Boring where it counts
Inference is infrastructure. Infrastructure should be predictable, observable, and dull to operate. We save the interesting parts for the routing engine.
Company
We build the layer teams stop thinking about.
Vantafold started because shipping an AI product meant writing the same integration three times and rewriting it every quarter. That work is real, but it is not anyone's product.
Mission
Make the model layer something you configure, not something you rewrite.
A product team should be able to pick a model on Monday, discover on Thursday that a cheaper one clears the same bar, and switch on Friday without a migration. That is a configuration change in principle and a two-week project in practice, and closing that gap is the whole company.
We are deliberately narrow. Vantafold does not train models, does not sell an application, and does not compete with the products built on top of it. We run the request path — routing, reliability, and the accounting that comes with it — and we try to be unremarkable at it.
How it started.
2023
The third integration
Our founders were on a platform team maintaining three provider integrations for one product. Each had its own retry semantics, its own streaming format, and its own idea of what a tool call was. The fourth was already on the roadmap.
2024
An internal gateway
What began as a shared proxy became a router: if the call is normalised anyway, the model choice can be made per request rather than per service. Spend fell by roughly two thirds on the first workload it touched, and quality held.
2025
Vantafold
The gateway left the building. The first external users were platform teams with the same problem and less patience for building it twice. Observability came next, because nobody would trust automatic routing they could not audit.
2026
Routing engine v3
Nine regions, active-active, with a scorer that estimates request complexity instead of following static rules. Millions of requests served daily, and a public status page we would rather not have to update.
Six things we keep coming back to.
Show the work
Every routing decision returns its reasoning. Every cost is itemised. If a number in our dashboard disagrees with your invoice, that is a bug and we treat it as one.
Aligned incentives
We do not mark up model tokens. Routing you to a cheaper model that still passes your evaluation is good for you and neutral for us, which is the only way it stays honest.
Portability by default
Nothing we build should make leaving harder. The API is documented, the exports are complete, and the local gateway runs without us.
Write it down
Design notes, incident write-ups, and deprecation timelines are published. Institutional memory that lives only in someone's head is a reliability risk.
Small surface, deep floor
Five endpoints, not fifty. We would rather make a small API work under extreme conditions than ship a large one that works in the demo.
A small team of infrastructure people.
We are distributed across San Francisco and Berlin, with a strong bias toward written communication and a weak tolerance for meetings.
Names and photographs are placeholders. Replace with your own team.
Backed by operators from the teams who built modern AI infrastructure.
- Logo placeholder
- Logo placeholder
- Logo placeholder
- Logo placeholder
- Logo placeholder
Come build the boring part.
We are hiring across infrastructure, developer experience, and security.