Skip to content

DEMO-SCRIPT.md

Latest commit

 

History

History
117 lines (84 loc) · 5.9 KB

DEMO-SCRIPT.md

File metadata and controls

117 lines (84 loc) · 5.9 KB
Raw

Concierge AI — Demo Script (3–4 minutes)

Opening (30 seconds)

"Every night in Nairobi, hotel owners like Mary — who runs a 12-room guesthouse in Westlands — lose bookings because guests message on WhatsApp at midnight and get no reply until morning. By then, the guest booked somewhere else.

Concierge AI is an always-on hotel assistant that handles bookings, speaks Swahili, and collects M-Pesa payments — all inside the chat apps guests already use."

Problem (30 seconds)

Show the Before vs After comparison:

Before (Manual) After (Concierge AI)
Guest messages at 11pm Guest messages at 11pm
Owner checks availability manually (next morning) AI checks availability instantly
Owner quotes price, sends paybill number AI quotes price, sends STK push
Guest switches apps, types paybill + amount Guest enters PIN on their phone
Guest screenshots receipt → sends back Payment auto-reconciled. Done.
6 steps. 15+ minutes. Often abandoned. 4 steps. Under 60 seconds.

Live Demo (90 seconds)

Flow 1: Guest Hotel Demo Page (30s)

  1. Open browser → navigate to /demo (public hotel showcase)
  2. Show: hero with "Savanna Suites Westlands", room cards (Standard KES 4,500 / Deluxe KES 8,500 / Suite KES 15,000)
  3. Type in embedded chat: "Habari, nataka chumba kwa watu 2 usiku moja" (Swahili)
  4. Show AI response: room options, prices, M-Pesa prompt
  5. Highlight: no login required, guest pays via phone number

Flow 2: Telegram Booking (30s)

  1. Open Telegram → message @conciergeSME_bot via deep link
  2. Send: "Niko na budget ya 5k kwa usiku moja"
  3. Show AI detecting budget tier + Swahili → offers max 3 rooms
  4. Reply: "Nachukua standard room"
  5. Show M-Pesa STK push notification (sandbox)
  6. Show receipt generated in chat

Edge cases to demo if asked:

  • Send a sticker → bot replies "I can only process text messages right now"
  • Spam 13 messages quickly → bot replies "You're sending messages too quickly"
  • Send from a group → silently ignored (bot only works in private chat)

Flow 3: Agentic Dashboard (30s)

  1. Switch to browser → open Dashboard (/dashboard)
    • Highlight: SSE "Live" indicator (green dot), tooltip on KPI cards explaining each metric
    • Show actionable insight bar: "Payment success rate is 78% — check failed transactions"
  2. Switch to Agent Chat (/agent)
    • Show owner talking to AI: "Check availability for this weekend"
    • Highlight inline tool-call cards (Availability Check → result)
  3. Switch to Workflows (/workflows)
    • Show the active booking pipeline with step progress (Search ✓ → Select ✓ → Pay ⏳ → Confirm)
  4. Switch to Approvals (/approvals)
    • Show escalated item: "Large payment — KES 34,000"
    • Hover risk badge → tooltip explains "Flagged for review — amount exceeds KES 15,000"
    • Show AI recommendation + one-click Approve/Reject (calls real API)
  5. Switch to Billing (/billing)
    • Show plan selection → Paystack inline popup
  6. Quick flash: Live Activity (/live) — real-time SSE event feed
  7. Show Reconciliation (in Payments or admin API):
    • Explain: "Cron runs every 5 minutes, auto-matches payments to bookings"
    • Show /admin/tenants/:id/reconciliation/summary — matched vs disputed
    • Show dispute resolution: owner marks unmatched payment as valid

Codex Role (45 seconds)

"I built this with OpenAI Codex — not as a bolt-on, but as my engineering team."

Show the .codex/ folder:

  1. Planning: "I use AGENTS.md to give Codex full project context — architecture, constraints, patterns. It knows the hotel domain, the M-Pesa rules, the testing standards."

  2. Building: "Codex wrote the Swahili NLU parser, the M-Pesa tracker, the receipt generator. Each commit is a Codex task."

  3. Reviewing: "I have 4 custom subagents — a reviewer that catches security issues, a test writer, an explorer, and a docs researcher. They run in parallel."

    • Show: .codex/agents/reviewer.toml (brief flash)

  4. Quality Gates: "A Stop hook automatically blocks Codex from finishing until all 112 tests pass and TypeScript compiles clean."

    • Show: .codex/hooks/stop_verify.py

What's Next (30 seconds)

"With another month, I would:"

  1. Go live with 3 pilot hotels in Nairobi (Westlands, CBD, Karen)
  2. Add WhatsApp Business API (currently Telegram + Web — WhatsApp is the #1 channel in Kenya)
  3. Multi-language — expand Sheng/Kikuyu intent detection
  4. Connect to live PMS — sync room availability from Opera, RoomRaccoon, or Google Sheets
  5. Open-source the Codex config — so other developers can fork and adapt for their markets

Closing (15 seconds)

"Concierge AI solves a real problem for real hotel owners in East Africa. It speaks their language, uses their payment rail, and it was built by an AI engineering team — Codex — configured to ship production-grade code.

Thank you. Questions?"

Backup Talking Points (if judges ask)

  • "Why not just use ChatGPT?" — "ChatGPT can't collect payments, doesn't integrate with M-Pesa, doesn't have tenant isolation for multi-hotel deployment, and doesn't persist conversation state."
  • "How does Codex compare to Copilot?" — "Copilot suggests lines. Codex builds features end-to-end: plans, implements, tests, reviews — autonomously in a sandbox. I configure the guardrails, it ships the code."
  • "What about security?" — "Tenant isolation at DB + route + DO level. Never store PINs. Parameterized queries only. Webhook signature verification. Per-chat rate limiting on Telegram. E.164 phone validation. Atomic D1 batch for payment state updates. 112 tests including security-focused ones."
  • "Cost to run?" — "Cloudflare Workers free tier handles 100K requests/day. D1 is free up to 5GB. Total: ~$0/month for a single hotel, $5/month at scale."