About Paperswift
Paperswift started as a loop I kept repeating. I had data in a database, questions about it, and no analyst to hand either one to. So I did what a lot of people do: pulled the data out with AI-written SQL, dropped the export into Claude or ChatGPT, and asked away. It worked well enough to keep doing and badly enough to keep bothering me.
The loop
- 1Get the data out
I'd ask an AI for the SQL, run it against the database, and get an error back. Then paste the error into the chat, take the corrected query, paste that into the database client, and run it again. Usually more than once.
- 2Hand the file to a chatbot
Once the export was clean it went into Claude or ChatGPT, which meant a copy of real customer and revenue data sitting on someone else's servers.
- 3Wonder whether the number was real
The answers came back confident and neatly formatted. I couldn't tell which figures had been computed across every row and which had been read off a sample, and checking one properly meant rebuilding it in a spreadsheet.
- 4Watch the tokens go
Long files and long threads cost real money, and every follow-up question paid again to re-read the same data.
- 5Lose the query
Months later the query I wanted was buried somewhere in a chat thread. Even when I found it, rerunning it on a fresh export and getting the identical logic wasn't something I could count on.
What Paperswift does instead
The starting point is the same: an export you already have. What changes is that the parts I was doing by hand are the parts the product does.
The answer is computed, not described. Your question becomes a query that runs over every row of your file, on your own computer. No sampling step, no summarising step, nowhere for a number to drift away from your data.
You can check it without redoing it. Every answer shows the steps it took in plain English and the exact query it ran, so verifying an answer is reading rather than rebuilding. See how that works.
Broken queries fix themselves. When a generated query fails, the error goes straight back to the model for a corrected one, up to two attempts. You never read a database error, and you never paste one anywhere.
Questions aren't metered. Every plan includes unlimited questions, so a follow-up costs nothing and a long afternoon of them doesn't end in a bill. See what each plan includes.
A saved insight is reproducible. Save the logic once and re-running it replays the stored query exactly, with no AI model involved, against next month's export. The same question months later gets the same logic, not a fresh guess at it.
Why it runs in your browser
Financial and customer data is exactly the data people are most reluctant to upload, and the old loop began by uploading it. Running the analysis in the browser removes that trade-off: the file is opened where it already sits, and the rows stay there. When a question needs an AI model, what we send is small and specific: your column names and whether each column holds text, numbers, or dates — never the values in your rows. The privacy and security page sets the boundary out in full.
What it isn't
Paperswift does not connect to your database. Step 1 above is still yours to do: you run the export, and Paperswift takes it from there. There are also no notebooks, no Python to edit, and no statistical modelling.
It sits between a spreadsheet and a data team. A spreadsheet is fast and private and stalls somewhere past a few thousand rows. A warehouse and an analyst will answer anything, months from now, at a cost most companies can't justify yet. Paperswift is for the export in your downloads folder and the question you have about it today.
Who builds it
I'm Stefan, and I build Paperswift. If an export of yours doesn't work, or you have a privacy question the security page doesn't answer, email me at stefan@paperswift.io.