Application / AI Engineer

Tell us how you build.

Short, specific answers beat long ones. Your draft is kept in this browser until you submit.

01 About you

Full name

Location and timezone

Hours per week you can give

GitHub or portfolio

LinkedIn optional

This is a paid role, but the budget is still being negotiated, so we cannot give you a number yet. Does that work for you?

Yes, I can start before the budget is set I need the number before I start Contact me once the budget is set

02 Your experience

What is your stack?

Languages, model providers, LLM frameworks, evaluation tools, ML training, infrastructure. Mark what you have used in production.

How well do you know Terraform or OpenTofu on AWS?

Describe one LLM system you shipped.

What did it do, how did you measure quality, and what broke first?

03 How you would work here

Our model writes: "This instance is exposed to the internet." The facts only prove a security group rule allows TCP/22 from 0.0.0.0/0. An entity check passes, because every value in the sentence exists in the input. How do you catch this, and what gets published when validation fails?

The model provider ships a new version. You have recorded plan fixtures, golden-file tests, a response cache and an evaluation set. How do you decide whether we may switch?

Name the metrics and the pass or fail gate.

In terraform show -json output, a security group has no cidr_blocks in after, and after_unknown.cidr_blocks is true. What should the analyzer do?

Report no public exposure and pass Raise a high-severity finding Record the analysis as incomplete, and do not call it safe Ask the model to infer the likely value

04 Design problem

Jev, from TypeSafe AI, takes unstructured state and returns typed, calibrated decisions instead of free text. We want a unique model of our own in that spirit, for Terraform and OpenTofu pull requests, with one difference: it must be deterministic. Same input and same model version, same output.

The AI still may not decide what a finding is or how severe it is. Assume no spending on paid training or GPUs for now, and a small fixture library to start from.

How would you build it?

Cover these points. If you think the idea is wrong in part, say so and explain why.

Which decisions the model makes, given what it is not allowed to decide
Input representation and output types
How you get determinism, given that Jev's decisions are probabilistic
Training data and method, with few fixtures and no training spend
Evaluation and calibration
Hostile input, such as a resource tag written as an instruction to the model
What you would try in the first two weeks

0 words. Most good answers run 300 to 800.

Your answers are used only to review this application.

Tell us how you build.

Thank you.