The whole answer, first

Anik Singal runs Lurn and was building a compliance product: software that reads marketing copy and flags claims the FTC would call an earnings claim. Some claims are explicit. Most are implied. His team had been at it for a while and the results wouldn’t hold still.

We got on a call in May 2024. I asked to see the failures. Within the hour I’d proposed a different shape: let plain rules catch the obvious cases first, send only the uncertain sentences to the AI, and ship the stable version before chasing the last few percent. I walked him through the whole thing. I didn’t hold anything back to sell later.

He hired me that week. The engagement ran four months.

One honest note. I told his team at one point that the detector was “100%.” Later examples proved it wasn’t. I don’t say numbers like that anymore. The architecture held; the boast didn’t.