AI humanizer and AI study app for classway.ai and niggle.aiAI humanizer and AI study app for classway.ai and niggle.ai

The founder's first plan was the one most tools followed then. Generate the answer, send it to a detector to measure how much of it read as AI, rewrite it with a humanizing service, then send it back to the detector to show it had passed, which made four paid calls for every question a student asked.

I tested the plan before building it. In those early days the detectors were all over the place, and the same paragraph could score fully human on one and fully AI on the next, so I checked every rewrite against five of them at once. The best humanizing service on the market gave the steadiest results, and when I took it apart I found it wasn't doing anything clever with the text. It was handing it to another language model with a prompt of its own.

That changed the design. If a humanizer was only a model with the right instructions, Classway needed one model and nothing else. I took the founder through four versions of the pipeline, each with fewer calls than the last, down to one call to write the answer and one to rewrite it.

Then came the prompt, and it took longer than everything else. I rewrote it against all five detectors until the pattern showed. Polished, evenly punctuated text gives a model away, and a student's writing doesn't look like that. It slips on its grammar and punctuation in ways a reader passes over, and it never reaches for the words models love. With that written in, the rewrites scored above 95 percent human on every detector I tested, and the detectors could come out of the product too. Every answer started as AI text, so there was no reason to pay a detector to say so, and once the rewrite held up there was no reason to pay one afterwards. The score a student saw was set by design, in the range the humanizer had earned in my testing, and was never a check on their answer.

I designed Classway's brand and every screen, and my team built it as a Chrome extension and a web app.

In the browser a student could highlight a question on any page, and a sidebar answered it and humanized the answer, with a multiple choice mode that picks out the right option in bold.

Every time OpenAI updated its models the prompt aged, and I tuned it again after each one, so Classway kept its score and its costs through every model from GPT-3.5 to GPT-6 Sol.

When I redesigned the homepage and built the humanizer into the site itself, the founder's reply was one line.

Months later the founder came back with a second company. Niggle turns a student's documents, PDFs, YouTube videos and recorded lectures into smart notes, summaries, flashcards and quizzes, and answers questions about any of them, so a course's material becomes something to study from instead of something to get through.

The hard part was the audio and video. Long recordings have to be prepared and sliced before a model can transcribe them, and I'd already built that pipeline on FFmpeg for Audiotextify, so Niggle's recordings run through it and on to Whisper. I designed Niggle's brand and the app, my team built it, and Classway's humanizer later joined it.

Students took to Classway's web app so fully that the team dropped the extension, since nearly all the use had moved to the site. Today classway.ai has more than 222,000 users and has solved more than three million questions, and niggle.ai stands beside it as the founder's second product. Both started earning early on, and both still run and still earn.