For delivery platforms

We build operations agents for platforms that relay orders between delivery apps and restaurant tills. The Proof costs from 1,500 euro and takes 5 working days: a read-only reader on your order feed, graded by 3 to 5 tests you write with us on day 1. It runs in shadow mode, so it reads and reports and writes nothing.

What the agent does on your feed

• It finds orders an app accepted that never reached the kitchen.
• It finds menus out of sync: a dish the app still sells that the kitchen has switched off.
• It watches for expiring till tokens, silent printers and an app channel that goes dark.
• It opens one incident per problem, with the evidence.
• It logs every step in a ledger: the evidence, the rule applied, the exact call and how to undo it.

Four authority tiers

A model may propose an action. A separate controller decides whether it runs, against a rulebook you sign off.

A kill switch stops everything. You set the tiers: a platform that wants a person to approve every write starts with everything on Ask first.

The Proof on your data

• What we read. Your order feed, through read access you grant and can remove: an API, a database replica or a scrubbed export.
• What it does. It runs the detectors on your real orders and writes nothing to your feed, the apps or the tills.
• How it is graded. By the 3 to 5 tests written on day 1. For example: an order the app accepted with no confirmation from the till is flagged as stuck.
• What you get on day 5. The incidents it found, the action it would have proposed for each, and the test results. You own the code.

Half is paid up front, half when it passes the tests. The Proof is credited in full against a Pilot booked within 30 days.

The demo, in three worlds

The delivery operations agent runs one rulebook in three worlds: the Gulf, India and Europe, each with its own currencies, cities, apps and tills. Every order, store and failure in it is synthetic.

In the Gulf world, on 7 replayed days, 249 planted incidents, it scored precision 0.98 and recall 0.99. It handled 95% without a person and handed 5% to a person, with the evidence attached. These numbers are against failures we planted, not on your feed. That is what the Proof is for. The weakest detector is named on the accuracy page.

A worked example, on synthetic data

On a busy evening a till's token expires, and orders the app accepted stop reaching that kitchen. The stuck-order detector opens one incident for the store, with the waiting orders attached. The controller finds the fix on the Do it tier: refresh the token and re-send the orders. It runs, and the ledger records the call and how to undo it.

What a Build adds

A Pilot from 6,000 euro puts the workflow live in your systems for 30 days, against a baseline you agree first. Writes come through your own APIs, each one logged with an undo.

A Build from 20,000 euro, in 3 to 6 weeks, gives you the governed agent: detectors across every app and till you relay, a rulebook with the four tiers, a ledger, and its own accuracy page. Care from 500 euro a month per live system covers monitoring, fixes, small changes and a monthly accuracy report.

Send us the job in the form below and a plan with a fixed price comes back within 24 hours.