Menu
Describe a task

How we work

Five stages. Each one ends with something you can check.

If the numbers aren't there, we stop early, while it's still cheap.

  1. 01

    Discover

    You show me the task. I ask about your tools, your data and what a good result looks like.

    You receiveA written use case, or a plain “not a fit”.

  2. 02

    Prove

    Core: a free 3-day sample check on 20–50 of your examples. Bigger builds: a paid 5-day Blueprint, credited if you go ahead within 60 days.

    You receiveA fixed price and a result target, in writing.

  3. 03

    Build

    Core takes 2–3 weeks, Standard 4–6, Advanced 6–10. The client portal is built alongside.

    You receiveA working system and its portal, in a test environment.

  4. 04

    Pilot

    I run the scoring script on the test set we froze before signing. When the target is met, you pay the final part and it goes live.

    You receiveA test report against the target.

  5. 05

    Hand over

    Your code in your repository, a runbook, and 30 days of fixes. An optional care plan after that.

    You receiveA named owner on your side, and the keys.

In every build

How every build handles an important action.

The AI proposes; a person approves anything that sends, pays, changes or deletes.

The accuracy promise

A result target, agreed before you sign.

Before you sign, I run your own samples through a first version and we agree in writing the result the build must reach, for example ‘every field correct on at least 47 of 50 of your invoices’. I only commit to a number I’ve measured on your data. When the build reaches it, the work is done. If it hasn’t by the delivery date, I keep working at my own cost until the extension date in the contract. If it still hasn’t, the final payment isn’t due.

The first payment covers the work done and isn’t refunded, unless I cancel or don’t start the work (then I return it in full) or your local law requires a refund. Accuracy on live data after launch is measured and reported, not guaranteed. Nothing here limits your rights under consumer law, such as the Australian Consumer Law.

Acceptance test report · Invoice & PO Automation

CheckResultStatus
Every field correct48 of 50 (target: at least 47)Met
Sent to a person2 of 50 (limit: 3)Met
Entered without approval0 (limit: 0)Met
Approval step worksyesMet
Every action in the audit logyesMet

Test set: 50 invoices · frozen, checksum listed · model and settings fixed

Sample report. Demo: built by Hadavin to show how this works. Sample data. Not a client project.

Every build includes

  • Built in your cloud account (AWS by default; Azure or Google Cloud on request)
  • The client portal: approvals, audit log, cost and results
  • A person approves anything that sends, pays, changes or deletes
  • A test set from your real data, and a target agreed before you sign
  • A runbook, and 30 days of fixes after go-live
  • Your code, yours after the final payment

What I don't take on

Regulated work that needs a certification or licence I don't hold, such as HIPAA health data, card payment data (PCI), or financial, medical or legal advice. Also hiring, tenant, credit or insurance decisions, web scraping, and systems that act on important things without a person approving. I'm not SOC 2 audited.

Any project still goes through the same stages, so a poor fit shows up in the first call or the sample check, not after you've paid for a build.

Start with a free 20-minute call.

Bring one task. You leave with a written use case, or a plain answer that it is not a fit.

Describe a taskSee pricing