Bulk AI runs over a spreadsheet · Lab

BatchPilot takes a CSV and one instruction and runs every row through an AI model, up to 25 rows at a time, with optional web search; I started it and Barış Ülgen built most of it.

What it is

FIG. 1One run over a table. The answer for each row has to be JSON that fits the typed columns, so the output lands as table cells, not free text.
Source: BatchPilot README.md and src/App.tsx (concurrencyLimit = 25), HEAD 14a016a

BatchPilot is a spreadsheet where every row gets its own AI call. You upload a CSV, write one instruction, and add the output columns you want, each with a type and a short description. The tool runs the instruction on every row, up to 25 rows at a time, and the new columns fill in while you watch. A switch lets each call search the web before it answers, and the model has to return JSON that fits the columns. You paste in your own free API key. A small server makes the calls, so the key never goes from the browser to the model provider.

The code is public at github.com/CogniScope/BatchPilot.

What happened

I made the first five commits in April 2026: the project itself, typed output columns, renaming columns in place, and later the field for your own key and the web search switch. Barış Ülgen wrote the other 44. He moved the calls onto the server, added pagination, drag and drop and a drawer with every field of a row, cleaned up the code, named it BatchPilot and released it as open source under an MIT licence in his name.

We used it inside CogniScope for lead research: from a studio to its live games, to the people who work on them, to who might need what we built. The hosted demo is down now; the code still runs locally.