PYX data product agentic engineering
The lifecycle

One journey, five stages, and a person at every gate

Not five products with a shared logo. One line a piece of work travels down, where each stage is refused entry until the previous one is honest, and where the AI drafts everything and decides nothing.

01 · RequirementThe stage in full →

The sentence somebody wrote, kept as they wrote it

A requirement arrives as prose, hedged and ambiguous. AI grills it into typed claims and blocking questions. A person answers them once, and the answers become part of the artifact.

AI drafts
Claims, blocking questions, data scenarios and the type of every term used.
You rule on
Every answer to every blocking question, and which claims were inventions.
It leaves behind
A requirement that can be argued with, dated and attributed, still readable in two years.

Argue with a draft, not with a blank page

Discovery reads your estate and proposes concepts, grain and join paths, each binding carrying a confidence and a stated reason. Modelling becomes a review, and the review is where the expertise goes.

AI drafts
Candidate concepts, the grain each measure could be at, join paths, and the physical bindings that would satisfy them.
You rule on
Grain, exclusions, the join decision, and every binding that gets accepted.
It leaves behind
A typed, diffable model where every inference states what it was inferred from.

Generated, not typed

Tables, layer pipelines, jobs and metric definitions all generate from the one model a person approved: Databricks-native, in source format, deterministic from what was agreed.

AI drafts
Bronze, silver and gold DDL, the layer runners, the job definitions and the metric views.
You rule on
What was approved in the model. Nothing downstream is authored by hand, so nothing downstream can disagree with it.
It leaves behind
Generated artifacts you can read, diff and run on your own platform, with the model they came from named in each one.

The gate the lifecycle passes through

Controls bind to the same model, at the moment the meaning is decided rather than months later. The verdict issues the release token, and what it leaves behind is reproducible eighteen months on.

AI drafts
The checks each claim implies, mapped to a vendor-neutral catalogue and parameterised to your assets.
You rule on
Thresholds, waivers, and which failures are allowed to hold a release.
It leaves behind
A verdict retained with its control version, parameters, query and observed value.
05 · VisualizeThe stage in full →

The number arrives with its reasons attached

Lakeview and Power BI outputs generated from the same model, so a figure on a page can still name the claim it answers, the control that held it and the requirement that asked for it.

AI drafts
The dashboard, its metric definitions, and the lineage that links every figure back up the chain.
You rule on
What deserves to be on the page at all, and which figures need their provenance shown rather than stored.
It leaves behind
A published product whose every number can be walked back to the sentence that asked for it.
Two ways in

Start from a requirement someone wrote, or a table you already have.

Design-first works with nothing connected. You can specify, argue about and generate a customer’s data product before anyone hands you credentials, which is when the design still matters most. Discover-first reads the estate you have and drafts the model from it. Both meet at the same stage two.