Builder project · Built at NYU SPS X Pulse Foundry
Hermes
How to optimise gemini specifically in regards to AISO for Cafe Athena

Tip: click @Pulse AI in the post to tag us.
From idea to reality
What we built
# HERMES: AI search agent for Café Athena ## What we built When a diner asks an AI assistant "happy hour near Garnet Ave?", the assistant gives one confident answer. Café Athena, right on Garnet with a daily 3–6 PM happy hour, isn't in it. HERMES is an agent that finds out why and fixes it, in four steps: **Discover → Analyze → Act → Measure**. ## How it works 1. **Discover.** HERMES asks Claude, with live web search, 15 questions San Diego diners actually ask. It records every answer, every search Claude ran and every source it cited. - **Result:** Café Athena is recommended in **4 of 15** answers and ranks **#3 of 62** restaurants named. - It's #1 for every "Greek" question and missing from all 11 others: date night, events, happy hour, dietary needs, patio. 2. **Analyze.** HERMES checks the sources the AI trusts and audits each restaurant's homepage. - AI cites Yelp in 12 answers and OpenTable in 7, but cafeathena.com in only 2. - The site has **no Restaurant, FAQ or Menu labels** (the structured data AI reads). Slowly, the most-recommended rival at 6 of 15, has the first two. 3. **Act.** HERMES drafts 7 fixes from verified facts only: Restaurant and FAQ structured data, an FAQ page, a landing page, an owner-disclosed community post, a press email and a review request. - **Nothing goes live without the owner.** They approve each fix, approved fixes publish to a staging page, and every change is logged and can be undone. 4. **Measure.** The same questions are asked again on a weekly schedule. - The baseline has held at 4 of 15 across three live runs, so it's a solid "before." - With the fixes, the **projection is 4 → 11 recommendations**. It's clearly labeled as a projection, because it assumes AI reads the new content. Weekly live re-runs after publishing are the real proof. ## Why it matters - **AI answers are the new search.** A small restaurant can't see that it's losing, let alone why. HERMES shows the gap, explains it in plain English, and closes it. - **Every number is real.** They all come from recorded Claude answers. Projections are labeled. No fake reviews, no sock-puppet posts, nothing published without approval. - **It keeps working after the hackathon.** Re-runs are scheduled weekly, the before-and-after is always visible, and the owner just taps Approve. **Live demo:** hit **Play the story** to see all eight steps in about 4 minutes, or open any tile to see one.