Platea: AI Agency That Sells Theater Tickets

Gemini agents run a real marketing agency for live theater: Meta/Google ads, CPA at the box office, WhatsApp bot. One human approves. Proof: El Gorila season in Mexico City.

Comment

Inspiration

I’m Osterlen Dupeyrón. My father, Humberto, has performed the same Kafka monologue—El Gorila—in Mexican theaters for 37 years. This season we’re in a 325-seat house in Mexico City. I produce. He acts. There is no marketing department, no agency retainer, and no budget that looks like a Silicon Valley SaaS deck.

What we do have is a real problem: sell tickets without burning money, and know—honestly—whether the ads are working. Not “platform-reported purchases.” Tickets paid for in our own box office.

That’s why I built Platea: an AI-operated marketing agency for live theater. Gemini agents do the daily work of media buying, analytics, PR drafting, and ops. I stay in the loop for anything public or irreversible. One human. Real revenue. Real stage.

(AdPilot is the media-buying layer inside Platea—Meta and Google agents—not a separate product.)

What it does

Platea runs the marketing stack for El Gorila Season 2:

  • Media Buyer agent pulls Meta (and Google) spend, crosses it with real box-office sales from our Stripe checkout, and computes CPA against a hard ceiling ($350 MXN per purchase). If CPA blows past the kill threshold, it alerts—no vanity dashboards.
  • Funnel Monitor found our real leak: people reached checkout but didn’t convert (far below a healthy benchmark). That diagnosis came from live APIs, not a slide.
  • WhatsApp bot (n8n + Vertex/Gemini) answers real customers about dates, prices, and the show—while a CEO path escalates to me.
  • PR agent drafts press pitches; with my explicit “go,” we send them (Resend). Presskit, UTMs, logbook.
  • Ops layer: scheduled jobs every morning (CPA check, occupancy, funnel email), a decision logbook for the hackathon, and a sanitized public repo of the agency.

The product isn’t “another ads autopilot.” It’s an agency that already has a client in production—us—and a theater filling seats.

How we built it

Stack in production:

  • Google Cloud / Vertex AI — Gemini for agent reasoning (CEO, assistants, PR, analysis).
  • Meta Marketing API + Google Ads API for spend and campaign reality.
  • Cloudflare Workers + Stripe for our own ticketing (no Ticketmaster middleman).
  • n8n for WhatsApp workflows; Resend for transactional and ops email.
  • Python agents with launchd jobs so the agency wakes up without me opening a laptop.
  • Public evidence repo: github.com/dupeyronosterlen/platea

What is new for this hackathon (after May 19, 2026): Platea itself—the agent roster, autonomy rules, CPA-as-truth loop, funnel monitor, decision logbook, Vertex migration for the bot, PR send path, Graphify knowledge graph for the agency.

What pre-existed (declared honestly): the show, the venue relationship, and the ticketing website. El Gorila is the customer / proof, not the Project. We say that out loud so judges aren’t guessing.

Challenges we ran into

  1. Attribution lies. Ad platforms love optimistic conversion counts. We forced every serious decision through box-office CPA. That meant building reporting Workers, snapshots, and the discipline to pause spend when reality disagreed with the dashboard.

  2. “Autonomous” vs responsible. Theater is reputation. Golden rule: I approve before anything public goes out. Agents propose, alert, draft, and execute within guardrails. Full autopilot that tweets nonsense would kill the brand faster than a high CPA.

  3. SSL, tokens, and Tuesday night ops. On this Mac, Python HTTPS sometimes dies; we learned to route critical calls through curl. Google tokens expire. WhatsApp OTP stalls approval workflows. Production AI is 40% models and 60% boring reliability.

  4. Language and packaging for XPRIZE. We built in Spanish for a Mexican audience. The submission must speak English. Translating without flattening the story—father, Kafka, 37 years, 325 seats—is its own craft.

  5. Not looking like every other “AI for Meta Ads” demo. The crowded pitch is AdPilot-in-a-vacuum. Our edge is Gemini agents operating a real agency with real ticket revenue. That’s harder to fake and harder to ignore.

What we learned

  • Gemini is strongest when it has tools + a ledger: APIs, a kill rule on CPA, and a written decision trail.
  • Small cultural businesses don’t need another dashboard. They need something that sells and tells the truth about cost.
  • Human-in-the-loop isn’t a weakness for this prize track—it’s how you ship AI into a business that can’t afford a scandal.

What’s next

Finish Season 2 with the agents running. Package Stripe revenue (arms-length vs related-party, cleanly), ship the <3 min demo video, and keep morning jobs + logs as continuous proof that AI is live—not a weekend prototype.

If Platea works for a 37-year monologue in Mexico City, it can work for other independent productions that will never afford a traditional agency.

Built With

Share this project:

Updates