← Back

Independent GPU performance engineering

We work with companies that own their GPUs but have no dedicated performance team — typically 2–8× A100 / L40S / H100-class machines bought through a budget process, now running open-weight models in production.

That is a specific kind of company, and it is the only kind we take on. You bought the hardware deliberately. What you did not get was anyone whose job is to make it fast, and your salary structure does not have a slot for one.

How we work

• We measure first. No recommendation goes out before there is a baseline your own people can reproduce.
• We work inside your environment. Your data does not leave the building — for some of the companies we work with that is not a feature, it is the precondition.
• One variable at a time. Every change is verified on its own, and every gain is reported next to what it cost.
• You get a report you can sign off on — before/after numbers with the scripts to re-run them.

What we do not do

• We do not sell hours. A project is quoted as one number, and the risk sits with us.
• We do not take on companies that already have a dedicated performance team — you do not need us, and we would only be in the way.
• We do not ask you to take our word for anything. Reproduce the numbers yourself; that is the point.