How one bad homework session started a company.

The origin story is not glamorous. In the fall of 2024, one of us was on the phone with a nephew who had just used ChatGPT to write his entire English paper. He was proud of it. He’d gotten an A. He couldn’t tell us, three days later, what the paper had argued.

We spent that weekend arguing. Not with the nephew — with each other. Was this a temporary displacement, like calculators once were? Or was this something more corrosive, aimed at the exact substrate that made learning possible?

We decided it was the second thing. Then we spent a year building the product that was our answer.

What we learned building it

The stance is easier to say than to keep. It took us four months to realize a system prompt was not enough — the output guard, the attempt gate, the sanitizer, the classifier all exist because we watched the base model drift under pressure. Every guardrail is a specific memory of a specific leak.

The other thing we learned: students are not the problem. They are the ones asking the right question. “Why does this feel bad?” They knew. We built the product they were already asking for.