personal.cafe

Your cafe remembers you.

Comment

Inspiration

The best dining experiences are personal. The host who remembers your name. The server who knows you're allergic to shellfish. The corner table that's always yours.

That only happens at expensive restaurants with long-time regulars. Every other restaurant treats you like a stranger - every single visit. You could be a weekly regular for three years and still get asked "first time here?"

We built personal.cafe to fix that. The domain said it all: make every cafe personal.

What it does

personal.cafe gives any café, restaurant, or hawker stall an AI front desk that actually remembers its customers.

Customers interact via WhatsApp in English, Mandarin, or Singlish. The AI handles bookings, FAQs, and menu questions - but what makes it different is memory. Every preference a customer mentions is saved silently: dietary restrictions, favourite tables, special occasions, past visits.

Next time they message:

"Welcome back Zach - window seat for 2 on Saturday? Still avoiding shellfish?"

For restaurant owners, onboarding takes 30 seconds. Paste your website URL - personal.cafe scrapes your menu and extracts your hours, location, halal status, and key dishes automatically. Zero manual setup. Zero spreadsheets. Zero crying.

Customers can also send WhatsApp voice notes in any Southeast Asian language - Arpo transcribes and replies in kind. Already live in production.

How we built it

  • Claude Opus (Anthropic) - conversational AI handling bookings and customer memory
  • Nimble - live web scraping to auto-extract menu and restaurant info from any website
  • Meta WhatsApp Cloud API - customer communication channel
  • Next.js 14 + Vercel - web app and API hosting
  • Neon Postgres - customer preference and booking storage
  • Upstash Redis - conversation memory across sessions
  • Arize Phoenix - LLM observability and tracing

Challenges we ran into

Getting Nimble to extract clean, structured data from restaurant websites was the main challenge - menus come in every format imaginable, from elegant PDFs to JavaScript-heavy SPAs that seemed personally offended by the idea of being scraped. We used Nimble's render mode for JS-heavy sites and fed the cleaned HTML to Claude to parse into structured JSON.

The other challenge was making the AI feel genuinely natural rather than robotic - especially across languages. Singlish in particular required careful prompt tuning so the AI could code-switch authentically. "Can book table for 4 lah?" should get a warm reply, not a 404.

Accomplishments that we're proud of

The onboarding flow: paste a URL, get a fully briefed AI front desk in under 30 seconds. We demoed it live with real Singapore restaurant websites and it just works.

The preference memory feels genuinely magical in practice. A customer mentions once that they're vegan - and every future conversation reflects that, without them ever having to repeat it. Like having a really good waiter who never quits.

What we learned

Nimble made the hardest part trivial. Web scraping used to mean fragile CSS selectors and constant maintenance. With Nimble, we pointed it at any restaurant website and got clean, rendered HTML back immediately - the real work was downstream parsing, not the scraping itself.

We also learned that the domain constraint is actually a gift. "personal.cafe" forced us to keep asking: is this personal enough? It shaped every product decision from the first line of code.

What's next for Arpo

  • Personalised discounts - returning customers get offers tailored to their history. Your 10th visit? Free dessert. Haven't been in 3 months? Here's a reason to come back. Discounts that feel like the restaurant noticed, not a mail merge.
  • Customer lifecycle management - track every customer from first message to loyal regular. Identify at-risk customers before they churn, surface your VIPs, and trigger the right outreach at the right moment. A full CRM layer, built entirely from WhatsApp conversations.
  • No-show reminders - automated confirmation and reminder sequences that actually reduce no-shows.
  • Review management - AI-drafted responses to Google reviews. The good ones and the spicy ones.
  • Self-improving AI - Arpo reads its own conversation traces, identifies failures, and rewrites its own instructions automatically. An AI that gets better every shift without anyone touching the code.
  • Multi-restaurant - one platform, hundreds of restaurants, each with their own personalised AI that knows their regulars by name.

Built With

Share this project:

Updates

Submission history