Skip to content
TechFabric

AI

AI consulting and development, built to survive production.

TechFabric builds AI systems that run in production: agents with governed data access, workflows that survive a restart, and the evaluation harness that tells you whether any of it is working. We run our own delivery on the same stack, so the patterns we recommend are ones we operate.

Most AI work stalls in the same place. The demo is convincing, and then somebody asks what happens when the model times out halfway through a refund, or who approved the thing it did at two in the morning, or whether last month's version was better. We build the system around the model.

Durable
A run survives a restart, a deploy and a provider outage, and picks up where it stopped rather than starting again
Governed
An agent runs under its own service principal and inherits Unity Catalog grants, so it cannot read what its caller could not
Measured
A fixed set of cases with known answers, scored on every change, so quality is a number rather than whoever spoke last

The demo worked. Everything after it did not.

Almost nobody gets stuck building the model. They get stuck at the point where somebody responsible has to be able to answer for what it did, and that shows up in three ways.

It reads things its user could not

The prototype connects with a service account that can see everything, because that was the fastest way to make it work. Now the answer it gives a salesperson quietly includes a row from payroll, and the fix is not a prompt.

It dies halfway and nobody notices

A run that takes forty minutes and waits on a person does not survive a deploy. It fails somewhere in the middle, the retry starts again from nothing, and the only evidence is a customer asking why their refund never arrived.

Nobody can say whether the change helped

Somebody swaps the model, or edits a prompt, and the team argues about whether it got better. Without an evaluation set the argument is settled by whoever is most confident, which is a poor way to run anything.

All three come from the same gap. The model was the project, and the system around the model was left for later, so permissions, durability and measurement each arrived as an afterthought or not at all.

The shift

Three ways this goes, and only one of them ships

Same ambition, same models available to everybody. What differs is the order the work is done in.

Promote the prototype

Fast, and it stays a prototype

Demo convinces
Pilot with real users
Security asks a question
Rebuild
Quietly parked

Buy a platform first

Governed, and nothing ships

Platform selected
Committee formed
Standards written
Pilots wait
Budget reviewed

Build the system around the model

Slower to the first demo, and it survives contact

One workload chosen
Permissions inherited
Durable run
Human approval where it matters
Measured, then extended

What actually gets built

Six parts, in your environment, on your cloud. Most teams need four of them and already have one, so the first job is working out which.

Retrieval that cites its source

Every answer comes back attached to the passage behind it, so checking it means reading the source instead of trusting the summary. This is also what makes an answer defensible to somebody who was not in the room.

Permissions the agent inherits

On Databricks the agent runs under Unity Catalog grants, so it cannot read what its caller could not. Row filters and column masks apply without anybody reimplementing them in application code, which is where they otherwise drift.

A runtime that survives a restart

Long runs and human waits are the normal shape of this work, and both outlast a deploy. Durable execution means the run resumes where it stopped, and the history of what happened is a property of the system rather than something somebody remembered to log.

A place a person says yes

Some decisions should never be automatic. The approval step is part of the workflow, with the context that decision needs on the same screen, and the record of who approved what kept alongside the run.

An evaluation harness

A held-out set, scored on every change, so the question of whether a new model helped has an answer rather than an advocate. Without this you cannot safely upgrade, which means you are stuck on whatever you launched with.

Model routing you can change your mind about

Providers are swapped through a registry rather than a rewrite, because the best model for a workload will not be the best one in six months. On Databricks that is the AI Gateway.

Two ways to start, both fixed fee

Fixed length, fixed price, and a written deliverable whether or not you do anything else with us. Which one you want depends on whether the question is what to build or what you are already paying for.

Two weeks

AI Readiness Assessment

Most AI programmes are not blocked by the model. They are blocked by data nobody trusts, permissions nobody can explain, and a platform where getting an agent to production means six approvals and a service account somebody created in 2023. Two weeks tells you which of those you have.

Book a readiness assessment

Two weeks

AI Tools Assessment

Three years of buying has left most companies with Copilot seats, a ChatGPT tenancy, a coding assistant per team, and four pilots nobody measured. Somebody in finance has started asking what it bought. Usually nobody can answer, because the seats and the outcomes were never on the same page.

Book a tools assessment

FAQ

AI development: common questions

What are AI development services?

Building the system around the model rather than the model itself. In practice that is retrieval and context, permissions so an agent cannot read what its caller could not, durable execution so a long run survives a restart, human approval where a decision needs one, and evaluation so you can tell whether a change made things better.

Most teams can get a prototype working. The engineering is in everything after.

Do you build on a specific model or provider?

No. Model choice changes every few months and a system tied to one provider is a system you rebuild. Our agent framework routes through a provider registry, and on Databricks that is Unity AI Gateway. We will have an opinion about which model fits your workload, and it is not an opinion worth hard-coding.

We have a pilot that works in a demo and nowhere else. Can you help?

That is the most common way this starts. Usually the demo is fine and the gap is durability, permissions, cost, or the fact that nobody can measure whether it is any good. We work out which of those it is before proposing anything, because the four have very different answers.

Do we have to be on Databricks?

No. Databricks is where we go deepest and where the governance story is strongest, because an agent can inherit Unity Catalog permissions instead of routing around them. We also build on Azure, AWS, Google Cloud and Cloudflare, and plenty of this work starts before a lakehouse exists.

How do you price AI work?

The assessments are fixed fee and fixed length, so you know the cost before the first meeting. Build work is scoped after one of those, or after a conversation if you already know what you need. We would rather write a scope we can defend than quote a number against a brief nobody has read properly.

Where we are

TechFabric is an AI development company in the Phoenix metropolitan area, founded in Gilbert, Arizona in 2017 and still run from here. Most of our work is national, and for companies across Phoenix, Scottsdale, Tempe, Mesa and Chandler we are a short drive rather than a flight.

Office

1530 E Williams Field Rd, Ste 201
Gilbert, AZ 85295

Phone

+1 (480) 681-6806

Serving

Phoenix, Scottsdale, Tempe, Mesa, Chandler