Ship the Process, Not the Code.A production backend in days.

A product on Kipory is a project: its flows, tables, endpoints, triggers, schedules, vocabularies and events are configuration rows, checked on save and authored through the design API or the operator UI. A save that would break a step is refused, and anything still unwired is reported as an open issue. Nothing is compiled or deployed, and a change can be dry-run against the live project before it is made.

Tables hold typed, schema-validated records. Relations link them into a graph you can assert and walk. An embedding profile makes a table searchable by meaning - so one query filters on fields, follows links and ranks by similarity, with no pipeline between three stores. Vocabularies classify records across all three.

Every model call - generation, extraction, classification, embedding, transcription - is a step with a named model and a price in credits. Spend is metered per step, run and project; spending stops at a budget ceiling, with a warning on the way; what a run or preview cost is readable the moment it finishes. Each step chooses its own model, so routine decisions run on small, fast models and only hard ones pay for a large one.

A flow is a typed graph of steps. Deterministic handlers fetch, parse, store, branch, fan out and merge; model steps decide, extract, classify and write inside that structure. The model never controls the shape of a run - the flow does - so intelligent behaviour stays bounded and repeatable. Sub-flows and loops compose larger processes from proven parts.

Every run keeps a complete step log - never sampled - and a record of every write it made; sampled runs also keep a trace of what each output actually held. Compare a good run with a bad one, find the step that moved, reprocess the records it touched. Spend, events and configuration history are readable the same way, so the system can always account for what it did and why.

An eval suite pins what “working” means: cases with pass/fail assertions, and scorer flows where quality is a matter of degree. A suite can re-run on its own when its flow changes, and a run-to-run comparison says whether a change made things worse. Saves are checked against dependents, so a flow cannot quietly break the endpoints and tables that rely on it.