From mismatched spreadsheets to repeatable conversions: a desktop app where AI does the thinking and your data stays on your computer

The challenge

Recurring data files are a quiet tax on many businesses: a supplier export arrives every week in one column layout and has to be reshaped into another before it can be used. The people we built this for are concrete - a data analyst spending 10 hours a week on data preparation who wants that under 2, an operations manager waiting 3 to 5 days for IT to convert a file, a small business owner spending 4 hours on manual data entry for each supplier import.

An online converter fails exactly the people who need it most: many are not allowed to upload business data to a third-party service. And handing the whole file to an AI means sending sensitive data away and trusting whatever comes back.

So we built TF Convert around a stricter idea: the AI does the thinking, your own computer does the work, and nothing the AI writes runs unchecked.

What we built

A desktop app that walks you through a conversion step by step. You pick the file you have and the layout you need. The app looks at both and proposes how the columns map onto each other - which match directly, which need to be combined or split - along with how confident it is in each suggestion. It then writes the conversion itself, shows a preview on a few sample rows, and once you approve, converts the full file on your own computer, with progress shown and the option to cancel. Six common file types are covered, including Excel and CSV, and files where the data sits in nested layers rather than flat tables are handled as well.

Only the column headings and 5 sample rows are ever shared with the AI - the file itself never leaves your computer. The AI also flags columns that look like personal information, such as names, emails and phone numbers, and those are masked on your machine during the conversion.

The app also remembers. Every approved conversion is saved, and when a file with the same layout arrives again - next week's export, say - the app recognizes it and reruns the saved conversion with no AI involved and no internet connection at all. And because real exports are messy, the app first works out where the actual column headings are, ignoring title rows and grand totals.

How we built it

Three rules shaped the build. The AI gets the minimum it needs: a glimpse of the file's structure, never the file. Everything heavy happens locally: reading, converting and saving all run on your computer, and repeat conversions need no internet at all. And anything the AI writes is treated as untrusted: each generated conversion is checked against a list of forbidden actions before it is allowed to run, and the generated conversion steps then run in a sealed-off environment that cannot touch anything else on the computer - not your other files, not the internet - with strict limits on how long they may run.

What this build demonstrates

TF Convert is an early version and still in development, but it is a real, installable product on Mac, Windows and Linux - roughly 41,000 lines of code written between October 2025 and June 2026, with the Mac version passing Apple's security verification. More importantly, it answers a question clients ask us constantly: how do you put AI into a product when the data must not leave the building? The pattern - give the AI the minimum, do the heavy work locally, check everything the AI produces, and reuse what it has learned offline - carries directly into internal tools and products that face the same constraint.