AI Implementation Engineer

What the work actually is

The assessment arrives already scoped and already priced. Your job is to make that number real. You prepare the client's own history so a model can learn from it, choose and train models for that specific work, wire the result into the tools the operation already touches, and test it against real cases from the client's last twelve months before anyone depends on it.

Then you write the runbook and sit with the person on their team who will own it until they can run it without you. We build in the client's accounts, on their billing, in their repository. Nothing here is measured on how dependent a client stays.

If you want to spend your days in a model playground, this is the wrong job. Most weeks include an afternoon of working out why one CSV export drops the timezone.

What you would be doing

• Prepare the client's history for training. Quotes, tickets, invoices, call transcripts, whatever the operation runs on.
• Choose the models for the work and train them. You make that call, not a client and not a vendor.
• Wire the system into the tools the operation already uses, and leave the parts that already work alone.
• Test against real historical cases and report the accuracy honestly, including where it is not good enough.
• Write the runbook and train the named owner on the client's team.
• Hand over accounts, billing, and code, then answer questions for 90 days.

What we are looking for

• You have shipped something that ran unsupervised against real customer traffic and you remember what broke.
• You are fluent enough with models in practice to tell a data problem from a model problem in an afternoon.
• You can read someone else's Zapier setup and leave it running when it works.
• You write. The runbook is part of the build, not paperwork after it.
• Python and whatever the client's stack forces on you.

What we are not looking for

• Research engineers. We do not train foundation models and we do not publish.
• Someone who needs a product manager between them and the client.
• Anyone who wants to rebuild the client's stack. We touch one operation and we leave.
• Anyone who will not get on a call with the operations manager whose job is changing.

Who should not apply

Do not apply if you have only worked at companies with a platform team. There is no platform team, and the client's data lives in four places, one of which is a shared drive.

Do not apply if demos are your best work. Nothing counts here until it has run against a real week.

Do not apply if you are outside the United States. We would like to say otherwise, but we can only hire and contract there at the moment, so an application from anywhere else is one we have to turn down on paperwork rather than on merit. We would rather you knew that before you wrote it than after.

How to apply

Email us with your CV attached and a short description of one system you built that ran without you watching it. What it did, what it was trained or built on, and what happened the first time it was wrong.

We would rather read 200 words about one system than a list of every stack you have touched.