Retail format harmonization

ROSETTA

Infer a mapping once, compile it against a narrow transform language, then replay every row deterministically.

Role: Data engineering · LLM safety · Evaluation
Year: 2026
Repository: Private

01 / Interactive system slice

Inspect the workflow

Public sandbox · real engine with clearly labeled cached and preview states

ROSETTA·

RETAIL DATA HARMONIZATION

REAL ENGINE · CURATED SAMPLE

SERVICE WAKING

Starting the free-tier service…

Harmonization run

North Market

Inspect source structure, compile a mapping, and reconcile the output.

RetailerNorth Market

Rows processed—

Held for review—

Checks passed—

Ready when you are

Run the sample to inspect its data contract.

02 / Engineering evidence

Claims with boundaries

Derived from the local implementation and offline artifacts.

Adversarial suite

18 casesFormats include ambiguity and prompt-injection fixtures.

First-day model cost

$1.2843Fresh local estimate for 18 formats and one million rows; later unchanged runs use cached specs.

Row execution

0 LLM callsApproved transforms execute as deterministic code.

Replay

HashableOutputs and mapping artifacts can be reproduced byte-for-byte.

Problem

Why this exists

Retailers encode the same business facts under different headers, units, delimiters, and conventions. Calling a model for each row is costly and non-reproducible.

System

How it is bounded

A profiler gathers evidence, the model proposes one mapping, and a compile gate rejects unknown columns, transforms, and parameters before row execution begins.

Outcome

What the demo proves

The workflow separates probabilistic interpretation from deterministic transformation and records the origin of every mapped field.

Implementation surface

PythonPydanticPandaspytest

Building reliable data products?

Let's talk through the system.

Start a conversation