# Papayya — Ready-made infrastructure for AI pipelines

> Papayya is ready-made infrastructure for batch and periodic AI pipelines — a task queue whose unit is the item. It grades every item a run produces, keeps the input that produced it, and lets you re-drive the single item that broke instead of the whole run.

## Your run says 1,000 succeeded. Thirty of them didn't.

They returned a value. Nothing raised, nothing retried, the exit code was 0 —
and the answer was wrong. You find out from a customer, two weeks later, and
the smallest thing you can re-run is the whole batch.

Papayya grades every item, keeps the input that produced it, and lets you
re-drive the one that broke. Queue, retries, per-item state, per-tenant cost —
the scaffolding you were going to write around your pipeline anyway, already
built.

## Here are the thirty

The home page carries a working ledger over one run of `tag_tickets`: a
thousand tickets across four tenants, filterable, with a re-drive button that
changes the row in front of you. The sample run is fixed, and its shape is the
argument:

| Fact | Value |
| --- | --- |
| Items in the run | 1,000 |
| Items that did not work | 30 |
| Of those, items that **completed** and still returned a wrong answer | 21 |
| Items that failed loudly | 9 |
| Estimated run cost | ~$2.20, from your rate card — an estimate, not a bill |

Twenty-one of the thirty finished cleanly. A status check sees `completed` and
moves on. Papayya records `ran` and `worked` as two independent columns, and
the disagreement between them is the thing you are buying.

## A 150-page report. Page 73 came back blank.

From a real call with a claims company. The numbers are measured, not imagined.

| Question | The pipeline they have | On Papayya |
| --- | --- | --- |
| What can you name? | The document. 150 pages arrive as one row, so there is no page 73 to point at. | Page 73. Every page is an item with its own id, input, verdict and cost. |
| How do you fix it? | Re-run the document. All 150 pages execute again — 300 page executions to change one. | Re-drive page 73. The 149 steps that already worked are reused, not recomputed. |
| What does that cost? | Two days in the queue, and the other 149 pages written downstream a second time. | 40 seconds. $0.002. Nothing downstream moves that did not need to. |
| How did you know? | A customer told you. The run reported success. | The item is graded on the way through, so page 73 was in the work list before the call. |

What changes is not the speed. It is whether there is anything smaller than the
document to ask about.

## Status and access

The `papayya` SDK and CLI are published on PyPI and installable today. The
server they talk to — control plane, worker pool and dashboard — is installed
into your own infrastructure, so there is no shared endpoint and no access to be
granted. We do the first install with you: [book a demo](https://getpapayya.com/demo),
or write to [kingsley@getpapayya.com](mailto:kingsley@getpapayya.com).

## Next

- [Developer resources](https://getpapayya.com/developers): API, OpenAPI spec, CLI, SDK
- [OpenAPI 3.1 specification](https://getpapayya.com/openapi.json)
- [Documentation](https://docs.getpapayya.com)
- [About Papayya](https://getpapayya.com/about)
- [llms.txt](https://getpapayya.com/llms.txt)