New files run with 0 model calls
What Nemotron learned from three imports
Each recipe is tested before you can approve it

Inspiration

A workshop coordinator gets the same registration CSV from a partner every week, and cleaning it by hand takes about 20 minutes. An AI agent could do it, but it would cost money every week and might do it differently each time. We wanted it to learn the job once and then just run it.

What it does

You import the file yourself three times. again sees it's the same job, and Nemotron writes down the rules you followed. It asks you about anything it's not sure of. Then it writes a recipe, a second model tries to break it, and again tests it on cases the models never see. You approve it, and from then on new files run as normal code with no model calls. If the file format changes, it stops and suggests a fix for you to approve.

How we built it

Nemotron 3 Ultra on Nebius Token Factory finds the rules, writes the recipe and fixes it
Nemotron 3.5 Lightning runs the edge-case tests (fast and cheap)
Strands Agents SDK, Python, FastAPI, SQLite, React, Google Cloud Run

A full learning run is 14 to 25 requests, about 15 to 25 cents.

Challenges we ran into

Getting Nemotron to write recipes our checker accepts. We raised the output limit, added an example of the format, and spelled out the order of the checks. We also moved the tester from Nano to Lightning because Nano was too slow.