Agentic Realty: WhatsApp AI Platform

• Static prompt audits: scripted n-gram and banned-phrase analysis of the prompt against itself. Every robotic tic in live chats was traced to the prompt's own example dialogues, not missing rules
• Exemplars beat instructions: rules banned phrases like 'on my end' while example replies taught them, so patch scripts now lint every new exemplar against the ban list before shipping
• Personas are swappable per agency (name, agency brand, review links) via whole-word replacement scripts. Same brain, rebranded per client
• Debounce plus typing indicators: rapid message bursts are collected into one coherent reply instead of five overlapping ones
• Budget honesty rules: tell Ama your budget is 15k and she drops the 22k listing and offers what 15k actually gets. She never oversells

• A · Communications router: webhook intake, delivery-receipt filtering, debounce, conversation locks, routing
• B · Lead intake & qualification, Ama's home: profiling, qualification, inventory matching, photo sending, escalation triggers
• C · Nurture: follow-up sender (Day 3/7/14/30), viewing scheduler with Google Calendar, and 24h/2h viewing reminders
• D · Matching engine: a new listing automatically finds every lead whose profile fits and alerts them on WhatsApp
• E · Review harvester: post-deal review requests with a 7-day reminder, converging on deal completion
• F · Escalation SLA: unanswered human escalations re-alert at 2h/4h/6h so a hot lead is never silently dropped
• Sync workflows: Twenty CRM ↔ operational DB property/lead/CRM sync, plus the Telegram agent and shared wa-send / claude-call / groq-call subworkflows

// Every workflow sends through wa-send. Nobody calls Meta directly.
const hoursSinceInbound =
  (Date.now() - new Date(lead.last_inbound_at)) / 36e5;

if (hoursSinceInbound < 24) {
  // Inside the window: Ama replies free-form, in persona
  return sendFreeText(lead.wa_id, message);
}

// Outside the window: only Meta-approved templates are legal
return sendTemplate(lead.wa_id, "property_followup", {
  name: lead.first_name,
  property: match.title,
});

-- Burst-proofing: one reply pipeline per conversation.
-- Acquire: SET lock:conv:{lead_id} <token> NX EX 60

-- Heartbeat (every 15s): only the owner may extend
if redis.call("GET", KEYS[1]) == ARGV[1] then
  return redis.call("EXPIRE", KEYS[1], 60)
end
return 0

-- Release (on every exit path): compare-and-delete
if redis.call("GET", KEYS[1]) == ARGV[1] then
  return redis.call("DEL", KEYS[1])
end
return 0

• Dictate a reply: the boss voice-notes 'tell him 13k flat if he pays 6 months upfront, lock a viewing Saturday 11am'. It lands on the lead in Ama's warm voice, and the viewing is booked
• Create a listing by voice: 'add a 3-bed for rent in Spintex, 8,500 a month'. The property appears in the CRM, and photos sent to the bot attach automatically
• Query the pipeline in natural language: 'show me all leads looking in East Legon'. Instant answer from the operational DB
• Escalation alerts arrive here too, so negotiation requests reach the boss's pocket in seconds

• Voucher scripts: every external API (WhatsApp, Claude, Twenty, Groq, Calendar) is smoke-tested with a standalone script before any workflow is built against it. Integrations are never debugged inside n8n
• Idempotent operations and read-back assertions on every database write
• Ghana phone normalization at lead creation: strip +/0, prepend 233, validate against MTN/Telecel/AirtelTigo/Glo prefixes
• Zero PII in logs: lead IDs only, never names, numbers, or message content
• Append-only versioned migrations for both the operational DB and Twenty's custom objects (Property, Lead, Viewing, Deal, Landlord)

#1WhatsApp message bursts causing racing AI replies

#2The AI's worst habits were taught by its own prompt

#3Being helpful outside WhatsApp's 24-hour window