Approach · AI infrastructure
Choose the model for the job, and keep it replaceable.
We build on established foundation models and cloud infrastructure, and engineer what surrounds them. That means deciding which model takes which task, what it can see, how its output is checked, and how to swap it when something better arrives. We don't train our own models.
How we approach it
- 01
Choose models per class of work.
From measured completions, not list prices. In one of our experiments the cheaper model cost $0.44 against $2.47 and never finished the task.
- 02
Use code where code can decide.
Rules are enforced by deterministic code. Small models can make narrow judgments; large models do the reasoning. A model's confidence never becomes permission.
- 03
Give each system only what its job requires.
An AI system sees the data its task needs. Sensitive records stay under the permissions they already have.
- 04
Keep providers replaceable.
Where we can, models are reached through infrastructure the stack already uses, so changing providers doesn't mean rebuilding the system.
Questions it answers
- Which model should handle which task, and what does a finished task cost?
- What data can each AI system see, and who decides?
- Where should a rule be enforced in code instead of by a model?
- What happens when a provider changes its price or its model?
In practice
Where this shows up in our own work, with each item's current status.
Tell us what you're trying to accomplish.
We'll talk through the problem, whether Lehvel is a good fit, and a sensible next step. The scope comes from the problem, not a package.