id8Labs
Back to Essays
Essay··10 min read

From Phone Call to Production: Building a Hospitality AI in 5.5 Hours

How pattern recognition mastery + AI collaboration = enterprise systems in hours, not months

The Challenge

4:30 PM EST, February 3rd, 2026. I'm catching up with my mom over the phone—she's in Costa Rica running her hotel, I'm at my desk in Miami surrounded by the usual chaos of building AI tools. We're just talking, you know, normal family stuff. How's the weather, how's business, what's new with the guests.

Then she mentions something that makes me pause. Hotel guests keep asking the same questions: Where should we eat? What activities are worth doing? Which beaches are the best? Her staff answers them fine, but there's this gap—guests want instant recommendations with photos, booking links, the works. Like a digital concierge that never sleeps.

I'm listening to this and thinking: I've built this before. Not exactly this, but the pieces. The chat widget, the database patterns, the visual recommendations. It's all sitting there in Homer, our real estate AI. What if we could adapt that architecture for hospitality?

Without even thinking about it, I grab my phone and hit record on voice notes. "Keep going," I tell her. "Tell me exactly what guests ask about. What would make your life easier?"

By 10:30 PM—six hours later—we had a production-ready hospitality AI deployed and ready for real hotel guests. Not a prototype. Not a demo. A working system that her guests could start using immediately.

This is the story of how that happened, and more importantly, how you can replicate this methodology for your own AI-powered product development.

The Breakthrough: Voice Notes + Existing Architecture

The speed wasn't luck. It was the result of two critical insights that changed how we approach AI development:

First, authentic customer development through voice recording. Instead of making assumptions about what hotel guests want, I recorded my mom—a real hotel owner—describing real problems in her own words. No product manager filter, no interpretation layer. Just unfiltered customer insight flowing directly into the product pipeline.

Second, foundation leverage over ground-up builds. We didn't start from scratch. We took Homer's proven chat widget architecture—months of development, 600+ tests, production-grade error handling—and customized it for hospitality. Why reinvent the wheel when you can put racing tires on the one you already built?

These two insights, combined with a systematic AI collaboration pipeline, compressed what normally takes months into a single evening sprint.

The Pipeline: Mobile → Milo → Claude Code

Here's the three-stage system that makes this kind of velocity possible:

Stage 1: Mobile Planning (Anywhere)

  • -Input: Natural language request via Telegram
  • -Output: Comprehensive implementation plan with research
  • -Benefit: Capture ideas anywhere, maximize idle time

Stage 2: Milo Research & Architecture (OpenClaw)

  • -Input: Voice recording of mom's conversation + raw pain points
  • -Processing: Analyzed customer needs, built comprehensive implementation plan
  • -Output: Detailed technical plans with datetime naming (20260203-1630-costa-rica-concierge.md)
  • -Benefit: Customer-driven requirements, cost-optimized planning, structured handoff

Stage 3: Claude Code Execution (Local)

  • -Input: "Take parts from Homer (existing chat widget), check Milo's plan directory"
  • -Processing: Leveraged existing Homer architecture, customized for hospitality use case
  • -Output: Working software with tests and documentation
  • -Benefit: Built on proven foundation, enterprise patterns, production-ready code

The Timeline: 5.5 Hours End-to-End

4:30 PM - Phone Call + Voice Recording

  • -Started as casual catch-up with mom in Costa Rica
  • -Digital concierge idea emerges naturally in conversation
  • -Breakthrough move: Immediately grabbed voice notes app and started recording
  • -Let mom guide the conversation through her pain points and desired features
  • -Pure customer development: Real hotel owner describing real problems in her own words
  • -Decision: Leverage existing Homer chat widget architecture as foundation

6:00 PM - Rich Media Database

  • -Loaded 29 Costa Rica locations with visual content:
    • -10 restaurants (food photos + booking links + Google Maps)
    • -12 activities (tour photos + operator websites + booking URLs)
    • -7 beaches (sunset shots + travel guides + location info)
  • -System architecture: Database → System prompt → ReactMarkdown rendering

8:30 PM - MVP with Visual Recommendations

  • -Chat widget showing actual photos instead of text descriptions
  • -Example response: "Try Anders Restaurant" becomes signature dish photo + website + Google Maps
  • -Parallel work: While stages executed, also created NotebookLM content packages (podcasts, videos, slides) for mom's marketing
  • -Ready for first user testing

9:45 PM - Production Hardening

  • -Error categorization system (src/lib/errors.ts)
  • -Loading timeouts with user feedback (30-second threshold)
  • -Input validation (1000-char message limit)
  • -Mobile-responsive design
  • -76 tests passing

10:30 PM - Deployed & Ready for Guests

The Secret Weapons: Voice Notes + Existing Architecture

Voice Recording as Product Discovery

The breakthrough wasn't just speed—it was authentic customer development. Instead of assuming requirements, I recorded mom talking through her actual pain points in her own words. No product manager filter, no interpretation layer—just real hotel owner describing real problems.

Why voice recording changed everything:

  • -Unfiltered insight: Mom described what guests actually ask about
  • -Natural conversation flow: Let her guide the discussion, not my assumptions
  • -Rich context: Heard the emotion and priorities in her voice
  • -Perfect handoff: Gave Milo the exact customer voice to analyze and plan from

Foundation Leverage: Homer Architecture

We didn't build from scratch—we customized existing proven systems. Homer's chat widget architecture already solved the hard problems (real-time chat, database integration, error handling). We just added hospitality-specific features.

Why building on Homer accelerated everything:

  • -Chat infrastructure: Real-time messaging already solved
  • -Database patterns: Supabase integration and queries proven
  • -Widget embedding: iframe and responsive design working
  • -Error handling: Production-grade patterns already implemented

What Made This Fast: Enterprise Patterns Applied

1. Clear Scope Fence

  • -Max 5 features for pilot
  • -Explicit "NOT building" list to prevent feature creep
  • -Vertical slices: Ship complete features one-by-one, not half-builds

2. Existing Foundations

  • -Homer widget architecture provided chat infrastructure
  • -Supabase data layer handled storage and queries
  • -Vercel deployment gave instant hosting and CI/CD

3. AI Collaboration Patterns

  • -Plan-first development: Comprehensive research before coding
  • -Context preservation: Plans survive model switches and session resets
  • -Parallel execution: Multiple AI agents working different aspects simultaneously

4. Production-Grade Discipline

  • -Error handling patterns from enterprise software development
  • -User experience research applied: 30-second timeout matches user abandonment patterns
  • -Test-driven validation: 76 tests ensure reliability
  • -Documentation-first: Admin guides written before deployment

The Pattern Recognition Breakthrough

The real insight: This is the same pattern I've used for 20+ years as a drummer.

Music learning pattern:

  • -Listen to song 100-1000 times
  • -Internalize structure and rhythm
  • -Play by ear (no sheet music needed)
  • -Trust the pattern recognition, iterate until it "sounds right"

Applied to AI development:

  • -Study the domain deeply (hospitality pain points)
  • -Understand the system architecture (chat widgets, databases, UX)
  • -Build by feel using AI tools (no rigid frameworks)
  • -Ship iteratively until it "feels right" for users

Same cognitive muscle, different domain. Pattern recognition mastery transfers across disciplines.

Business Impact: Real Validation

This wasn't a demo or prototype—it's a real hospitality AI for real hotel guests.

  • -Pilot customer: Mom's Hotel Conchal in Costa Rica
  • -Market validation: Hotel owner testing actual solution with actual guests
  • -Scale potential: Costa Rica hospitality industry (500+ hotels)
  • -Revenue model: Pilot → subscription → channel partnerships

The critical insight: Start with someone who has real skin in the game. Mom's reputation depends on guest experience, so she's the perfect validation partner.

Technical Architecture: Why It Scales

Database Design

-- Rich media stored as JSONB arrays
{
  "images": ["sunset-shot.jpg", "daytime-view.jpg"],
  "booking_url": "https://example.com/book",
  "google_maps": "https://maps.google.com/?q=Location"
}

System Prompt Injection

API route automatically builds knowledge from database, including URL fields. When AI references a restaurant, it has immediate access to photos and booking links.

ReactMarkdown Rendering

Chat widget renders [links](url) and ![images](url) natively. AI picks most contextual photo automatically.

Result: Visual Travel Companion

Instead of text recommendations, guests get rich visual experiences with instant booking capabilities.

Lessons for AI-Leveraged Development

1. Pipeline Beats Point Solutions

Don't optimize individual tools—optimize the handoff between stages. Mobile planning → research → execution creates compound leverage.

2. Scope Discipline Enables Speed

The "NOT building" list was as important as the feature list. Clear constraints prevent endless iteration.

3. Start with Real Users

Mom's hotel guests are the ultimate product validation. Real pain points, real usage, real feedback.

4. Foundation Investment Pays Off

Homer's widget architecture enabled this sprint. Previous infrastructure work creates future velocity.

5. Voice Notes for Customer Discovery

Recording natural conversations captures unfiltered customer insight. Let customers guide the discussion, don't lead with your assumptions. The customer's voice becomes your product requirements document.

6. Parallel Work While Waiting

While AI agents executed development stages, I simultaneously created NotebookLM content packages (podcasts, videos, slides) for mom's marketing. Maximum leverage of idle time.

7. Pattern Recognition Transfers

The same cognitive skills that make someone good at music, sports, or filmmaking apply to AI development. Trust your ear.

What's Next: Stage 11 Iteration

Immediate (Week 1):

  • -Monitor first guest interactions
  • -Gather feedback from mom and hotel staff
  • -Track usage patterns and popular queries

Short-term (Month 1):

  • -Add Spanish-first content for local preferences
  • -Integrate real booking systems (not just links)
  • -Custom branding for Hotel Conchal

Long-term (Quarter 1):

  • -Scale to other Costa Rica hotels
  • -Channel partner program with local tourism boards
  • -Multi-language support (German, French for European tourists)

The Bigger Picture: AI as Cognitive Leverage

This sprint demonstrates AI as cognitive extension—not replacement, but amplification of human pattern recognition and domain expertise.

The human provides:

  • -Domain knowledge (hospitality industry pain points)
  • -Pattern recognition (what feels right for users)
  • -Strategic direction (pilot → scale pathway)
  • -Quality judgment (ship vs. iterate decisions)

AI provides:

  • -Research velocity (database of Costa Rica locations in hours)
  • -Implementation speed (production code with tests and docs)
  • -Error handling patterns (enterprise-grade reliability)
  • -Documentation generation (admin guides, API docs)

Together: Enterprise software built in hours with human judgment and AI execution working in harmony.

Conclusion: The New Normal

5.5 hours from phone call to production-ready hospitality AI.

This isn't a one-off achievement—it's a reproducible methodology for turning domain expertise into software systems using AI collaboration.

The future of software development isn't humans vs. AI. It's humans + AI creating systems at the speed of thought, with enterprise quality, deployed globally.

Pura Vida! 🇨🇷

Eddie Belaval is the founder of id8Labs, an ideation laboratory building AI-powered tools for creators and infrastructure for builders. Follow the journey at id8labs.app.

Back to all essays