What I built at Fired Fest

Gauntlet AI ran a hackathon called Fired Fest on June 12 and 13, and the brief was short: build a harness for an AI. Harness meaning the structure around a model that makes its output trustworthy, as opposed to another chatbot. I liked the framing because I already had a problem that fit it.

Here's the problem. An AI that drafts marketing copy will happily invent things. Ask it for a promo post and it can hand you a discount that doesn't exist or a feature you never built. On a personal account that's embarrassing. On a company account it's a liability. Lumora Sleep sells real products at real prices, and I'd like our posts to keep matching reality.

So that's what I built. The harness checks Lumora Sleep's social posts before they publish. At the center is a source-of-truth file with the actual product facts: Essence is $89, Max is $249, Caelum inserts are $24. When the model drafts a post, the harness verifies every claim and price against that file. It checks the links. It flags anything the model wrote that the facts don't support. The rule is blunt. Prove the claim or lose it.

Verification is the unglamorous half of AI. Generation gets the demos and the applause. Verification is a file of prices and a function that says no. But the boring half is the one that makes the whole thing usable for a business. A model that usually writes a fine post is still a bad hire for a company account if it sometimes promises a sale you never planned. The harness exists for the sometimes.

Two days isn't much time, so nothing about the build is polished. It doesn't need to be. It needs to be strict, and it is (the model doesn't get the benefit of the doubt). Watching it flag a claim the model invented, then refuse to let the post through, was the best moment of the two days.

The room was the other half of the experience. I was the youngest one there by quite a bit, in a hackathon full of working engineers, and I got to talk with real investors and with the Gauntlet AI team, including the CEO. Those conversations taught me a different kind of thing than the build did, mostly what the people who fund and run AI companies actually worry about.

As for Lumora's account, the arrangement is simple now: the model writes, the harness checks. Anything the model can't back up with the facts file never gets posted.