AI data cleaning — fix messy files in plain language

Messy files don’t need hand-editing. On the review step, Mildport suggests fixes for what validation actually found — “Fix 3 emails”, “Format 2 phones” — and takes requests in plain language: “split the full name into first and last”, “map aktiv/inaktiv to active/inactive”.

What makes it different from other “AI cleaning” features: the model never touches your data. It only proposes a short list of cleanup steps; the importer executes them itself and shows you the exact before → after preview. Nothing applies until a person clicks Apply.

How it feels in practice

1. Upload a messy file. Phone numbers in five formats, names crammed into one column, German status words where your app expects English ones.
2. Click “Clean up”. Suggested fixes appear ranked by how many cells they repair — each one already verified against your actual rows, so a chip never over-promises. Or type what you want changed, in any language.
3. Review the diff. A compact before/after table, per-step breakdown, and a safety check that blocks any step that would turn a valid value into an invalid one.
4. Apply. Changed cells are tinted in the grid with a note saying which cleanup changed them — inspectable, undoable, exportable like any other edit.

It gets cheaper the more you use it

An accepted cleanup is remembered for that file layout. Next Monday’s file from the same customer gets a one-click “Your saved cleanup” suggestion that runs instantly — no model call, no AI cost, guaranteed same result. Over time the assistant teaches your import the rules and gets out of the way.

Everything the importer has learned is listed on one admin page — saved cleanups and learned column aliases alike — with per-item “forget” buttons. Right-to-forget is a click, not a support ticket.

	Typical “AI cleaning”	Mildport
Who edits the data	The model, in the vendor cloud	The importer itself, in the user’s browser
Review	After the fact, if at all	Exact before/after preview before anything
Repeat imports	Pay per AI run	Accepted cleanups replay free, identically
Model	Vendor’s	Bring your own
Audit	Opaque	Every proposal and decision logged; one audit page

For developers: cleanups as code

An accepted cleanup can be downloaded as a small JSON file and shipped with your integration — the widget applies it automatically for matching files, your CI can pin its behavior with a dry-run endpoint, and the same file works headlessly over the REST API. See Configure for the embed options.

FAQ

What is AI data cleaning in an import tool? The user describes a fix in plain language and the importer proposes it across the whole file — with an exact preview a person approves before anything changes.

Does the AI edit my customers’ data? No. The model only proposes cleanup steps; the importer executes them in the browser, and only after a person accepts the preview.

Do repeated imports cost AI usage every time? No. Accepted cleanups are remembered per file layout and replay with no model call — identical result, zero AI cost.

Can I turn it off? Yes — per plan, per deployment, per workspace, per embed. And suggested fixes that need no model keep working with AI fully disabled.

---

Try it on your own files — mildport.com — or see how the whole flow fits together in How it works.