Skip to content
TechFabric

Ship something new

Take the AI demo to production

The model is not the hard part. We build the durability, the permissions and the evaluation around it, and put the thing in front of real users.

It works in the notebook and nowhere else

The prototype convinced everyone. Then somebody asked three questions and the project stopped.

What happens when it times out

A run that takes two hours and calls six systems has no answer for a deploy in the middle of it, so it starts again, or worse, half-finishes.

What is it allowed to see

It got a broad service account because that was quickest. Its access is now wider than any person's and nobody has reviewed it.

Is this version better

The prompt changed. Whether the thing improved is decided in a meeting by whoever is most confident.

None of these are model problems and a better model fixes none of them. They are all properties of the system around it, and that system is what does not exist yet.

The shift

Three ways a demo becomes a product

Two of them are the demo again, with more infrastructure.

Ship the prototype

Fast, and it breaks in public

Demo works
Wire it up
Ship
First timeout
Roll back

Rebuild it properly, later

Correct, and it never starts

Demo works
Plan a rebuild
Scope grows
Priorities move
Shelved

Build the system around it

In production, and explainable

Demo works
Durability added
Grants resolved
Scored against ground truth
In front of users

What we build around the model

The half nobody demos, which is the half that decides whether it survives.

Durable runs

The work is a workflow. A restart, a crash or a deploy mid-run resumes rather than starts over, and a retry does not repeat what already happened.

Grants, not prompts

The system reads through the same permissions a person would. An agent reaching a table it has no grant for fails the way any principal fails.

Retrieval that cites

Every answer carries the documents behind it, so checking it means reading the source rather than trusting the summary.

An evaluation suite

Ground truth you own, deterministic scorers, and model-based judges where correctness is a matter of degree.

Human approval where it belongs

The actions that warrant a person stop and wait, and the approval is recorded against the exact operation.

Cost you can see

Model spend, retrieval infrastructure and the review time it still needs, instrumented at build time rather than discovered on an invoice.

2

Of our own products running on it, in production

The same work we did on our own

Fabric and Canvass are both in production and we are on the pager for both. Every part of this list exists because one of them needed it before a client did.

FAQ

AI features to production: common questions

We have a prototype already. Do you start over?

Almost never. The prototype proved the idea, which is the expensive part to be wrong about. What is usually missing is everything around it, and that gets built underneath what you have.

Do we need to be on Databricks?

No. It is where we are deepest, and this work runs on Azure, AWS and Google Cloud too. What matters is whether the thing has to be governed and explainable.

How do you know it is working after we hand over?

The evaluation suite is yours and it runs on every change. That is the point of building it: the answer stops depending on who is in the room.