Builder project · Built at NYU SPS X Pulse Foundry
SPS X Pulse Foundry Hackathon
Athena Agent

Tip: click @Pulse AI in the post to tag us.
From idea to reality
What we built
We built an AI growth agent that shows a restaurant where AI assistants ignore it, then fixes the gaps and tracks the results. Our target was Café Athena, a Greek restaurant in Pacific Beach, and the agent follows the same loop the challenge describes: Discovery, Analysis, Action, Measurement. We started by asking Claude and Gemini 111 real San Diego diner questions, grouped by request type and weighted by what a booking is worth. Café Athena was recommended in only 11 of them (Claude 17%, Gemini 4.7%), and it was never recommended for private events, which are worth about $2,000 per booking. Catering, worth $1,200, appeared only in Claude’s answers, at 25%. The agent then explains why rivals win: Café Athena’s site has no Restaurant, FAQ or Menu schema and no llms.txt, competitors have dedicated event pages, and its listings contradict each other on happy hour days and Friday and Saturday closing times. When Gemini was given Café Athena’s own site to read, it ranked the restaurant #1 in 7 of 12 test questions, so the content is fine and the real problem is being found. To fix that, the agent generates 14 ready-to-use files built only from facts the restaurant already publishes, including an events and catering page, happy hour, date night, and vegan and gluten-free pages, an FAQ, schema markup, llms.txt, a Google Business Profile description, and outreach emails to the guides assistants cite. Everything is ranked by booking value, and the owner approves every change before it goes live, with each change reversible and logged. To prove it worked, the same question bank is re-run weekly and compared against the saved baseline. We claim no improvement yet, because a lift can only be shown after a real re-check following publication. This matters because small restaurants can’t see this channel and can’t afford specialist help. The dashboard turns it into a dollar-ranked to-do list the owner approves one item at a time, and the weekly re-check keeps working after the hackathon ends.