The practice
Named for the grain, not the postcard.
Production AI looks smooth from a distance and is all ridgeline up close. Dunes works at the grain — serving, retrieval, evals, cost, ownership.
The principal
In the work.
- AI services and infrastructure
- Practice of one, specialists as needed
- Remote worldwide · on-site by arrangement
A practice of one
Dunes is a practice of one principal who has spent a career in AI services and infrastructure — the unglamorous layers that make models useful: serving, data planes, evaluation, cost, and the teams that have to live with the system on Monday.
Specialists join when the work needs another pair of hands. You get principal-level judgment without a nine-month hiring cycle, and without a 40-person firm’s overhead.
We write the documents. We also write the code. The engagement is finished when someone on your side can walk the path without us.
Principles
Systems over slides
If it does not change the architecture, the eval, or the on-call path, it is not the work.
Named owners
Every recommendation has a person who will live with it after we leave. Otherwise it is decoration.
Cost is a feature
Inference, GPU, and retrieval spend are designed, not discovered in month three.
Honesty over fluency
A system that refuses cleanly beats a system that sounds sure. We design for the former.
What we don’t do
Saying no is part of the practice. If the work is a slide deck, a staffing request, or an unscoped agent, we will tell you quickly.
- Strategy decks with no system underneath
- Staffing or body-shop augmentation
- Unscoped 'add an agent' requests
- Foundation-model training unless that is the actual problem