Design partner program is openThree partners, real decision traffic →
razoo

Design partner program

Three partners. Real decision traffic. The number, measured together.

We are looking for teams that already run agents on recurring work: invoices, contracts, triage, renewals, inbox. Work where a person makes the same call every week and would rather not. You bring the traffic. We bring the kernel and sit with you while it learns.

What you get

The kernel, the team, and a number you can defend.

The kernel, as an appliance or the SDK

One container on your infrastructure with a persistent volume, or one package, delivered as a tarball, inside your product. Your choice, and you can change it. A commercial licence is included for the program.

A weekly working session

With the Razoo team, on your decisions, until the first learned rule fires on its own. Then as often as the queue needs.

The conversion report

Your share of governed decisions that stopped needing the model, with the evidence behind each learned rule and the envelope it fires inside. Every week.

The technical review, under NDA

Architecture, the invariants and the tests that pin them, threat model, and red-team history. You read it before you install anything.

Free for the program, preferred terms at GA

No charge while you are a design partner. Preferred commercial terms at general availability, in writing, before you start.

razoo · conversion report · week 7illustrative · not a measured result

Governed decisions

412

Now deterministic

27%

Unsafe actions

0

  • signature-missing
    authored · 0 tokens
    Rule
  • renewal-note
    replayed clean · envelope: action
    Learned
  • contract-expiry
    candidate · gathering evidence
    Learned
  • value-above-threshold
    money_legal · always a person
    Human
Every learned row links to its evidence.Revoke

What we ask

Real traffic and honest feedback. Especially when it is wrong.

Real traffic

At least one workflow with recurring human decisions, governed end to end. Synthetic data will not produce the number, and we will not publish one it did.

Honest feedback

Where it is slow, confusing, or wrong. Counterexamples are the point: a counterexample revokes a rule, and that is the system working.

A case study, when the number exists

Once the conversion figure is real and you are happy with it. Reviewed by you before anything is published, and never before.

An engineer, a few hours a week

Someone who can wire an inbox, a webhook, or an SDK call, and who will tell us when the docs are wrong.

Who it suits

Recurring human decisions with a side effect at the end.

The kernel earns its keep where a person settles the same call over and over and then confirms something: an approval, a reply, a chase, a delegation.

Invoices and payables

Approval within delegated limits from known suppliers. Exceptions to a named approver. The money line a rule may never cross.

Contracts

Completeness, signature status, expiry inside a window, value above a threshold. The reference domain the kernel ships with.

Triage

Sorting incoming work by the fields that decide it, routing the rest to the person who decides, and learning which of those routes repeat.

Renewals

Who to chase, when, and in what tone. The follow-up a person drafts the same way every week.

Inbox and operations

Replies, follow-ups, and delegation, everything staged and nothing sent unconfirmed.

Not a fit, yet

Open-ended chat with no side effect. Decisions nobody makes twice. Workflows with no human in the loop today, because there is nothing to learn from.

Eight weeks

Install, judge, and watch the first rule earn its place.

The timeline is paced by your traffic, not by us. A rule has to earn its place against real judgments, so the weeks below assume a workflow that produces a few decisions a day.

Week 1Model

Install and connect

Appliance or SDK on your side. One workflow connected: a mailbox, a webhook, or an SDK call. Validated by doctor, the CLI's checker. Every decision the agent makes starts on the model rung and is recorded.

Weeks 2 to 4Judgment

Judgment and evidence

Your people make the calls they already make, inside the queue. Each one is captured as evidence. The reasoning-debt view shows where the model is still being asked the same thing.

Weeks 5 to 8Learned

First learned rules and the report

Candidates replay against your history. The ones that pass are promoted by a person and fire inside their envelope at zero tokens. The first conversion report is the number we measured together.

Apply

Tell us about the decisions your agents keep asking about.

One email is enough. Which workflow, roughly how many decisions a week, which agent framework and model provider, and whether you would run the appliance or embed the SDK. We reply to every application within two business days.

Apply by emailhello@viabuild.app

Three partners. Real decision traffic.

If you run agents that make the same calls every week, we want to govern them and show you what stops needing the model.