Evaluation before models
If we cannot measure it, we do not ship it. The test set is the first artefact we write, and it belongs to you.
profilo-studio
Made by ariya-tech.net
The Studio
A small studio that ships large systems.
Mirage is six research and platform engineers in Masdar City. We are deliberately small so that the person who designs the evaluation is the person who carries the pager.
We started in 2021, after years of watching promising models die in slide decks. The demo always worked; the production system never arrived. Mirage exists to close that gap — to treat an AI feature like any other piece of software that has to run at 3 a.m.
That means we start from the evaluation, not the model. We write the test set with your team first, agree on what "correct" means in your domain, and only then choose the smallest model that clears the bar. It is a less glamorous way to work, and a far more reliable one.
Applied AI, built to deploy
Masdar City, Incubator Building, Abu Dhabi
// The team
One team, six orbits
The disciplines that turn a model into a system, and the order they usually run in.
- 00 Evaluation
- 01 Data
- 02 Fine-tuning
- 03 Retrieval
- 04 Agents
- 05 Operations
How we work
Latency is a requirement
A correct answer that arrives in nine seconds is a wrong product. We treat the P95 budget as a spec, not a hope.
Your data stays yours
We deploy in your cloud or the UAE region, train on nothing without written scope, and hand back every weight and eval.
Boring in production
The best compliment a system earns is silence. We optimise for predictable, observable, and easy to roll back.
The people on call
LM
Layla Al-Mansoori
Founder · ML Lead
OH
Omar Haddad
Research Engineer
YN
Yara Nassar
Retrieval & Data
KF
Karim Farouk
Platform & Ops
HZ
Hind Al-Zaabi
Evaluation Lead
Where we stand
2021
Founded
12
Systems in production
99.95%
Aggregate uptime
6
Engineers, no layers
Have a task worth measuring?
Bring the problem and one example of a correct answer. We will bring the evaluation.
Start a project