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

The challenge

Repeated data files cost many businesses a lot of time. A supplier export arrives every week in one column layout, and somebody must reshape it into another layout before use. We built the app for concrete people:

• a data analyst who spends 10 hours a week on data preparation, and wants under 2
• an operations manager who waits 3 to 5 days for IT to convert a file
• a small business owner who spends 4 hours on manual data entry for each supplier import

An online converter fails exactly the people with the greatest need. Many of them cannot upload business data to a third-party service. A transfer of the whole file to an AI also sends sensitive data away, and forces trust in the result.

So we built TF Convert around a stricter idea. The AI does the analysis. Your own computer does the work. Nothing from the AI runs without a check.

What we built

A desktop app guides you through a conversion, step by step. You select the file you have and the layout you need. The app examines both and proposes a map between the columns. It marks a direct match, and a column that needs a combination or a split. It also gives a confidence level for each suggestion. The app then writes the conversion itself and shows a preview on a few sample rows. After your approval, it converts the full file on your own computer. It shows the progress, and you can cancel. The app covers six common file types, among them Excel and CSV. It also handles a file with data in nested layers instead of a flat table.

Only the column headings and 5 sample rows go to the AI. The file itself never leaves your computer. The AI also marks a column with probable personal information, such as a name, an email address, or a phone number. Your own machine masks those columns during the conversion.

The app also remembers. It saves every approved conversion. When a file with the same layout arrives again, such as the export of the next week, the app recognizes it. It then runs the saved conversion again, with no AI and no internet connection. Real exports are untidy, so the app first locates the real column headings. It ignores a title row and a grand total.

How we built it

Three rules shaped the build.

The AI gets the minimum. It sees a small part of the file structure, never the file.

The heavy work stays local. The read, the conversion, and the save all run on your computer. A repeat conversion needs no internet at all.

Everything from the AI is untrusted. The app checks each generated conversion against a list of forbidden actions before it runs. The generated conversion steps then run in a sealed environment with no access to the rest of the computer: not to your other files, and not to the internet. A strict time limit also applies.

What this build demonstrates

TF Convert is an early version, still in development. It is also a real, installable product on Mac, Windows and Linux. It holds roughly 41,000 lines of code from the period between October 2025 and June 2026, and the Mac version passed Apple's security verification.

More importantly, it answers a question clients ask constantly. How do you put AI into a product when the data must not leave the building? The pattern is repeatable: give the AI the minimum, do the heavy work locally, check everything the AI produces, and reuse a saved conversion offline. That pattern carries directly into an internal tool or a product with the same constraint.