WhatsApp Assistant: Resolving Rigidity via Hybrid AI
The bot reads natural WhatsApp messages (EN · VI · ZH), extracts trip details, clarifies missing parameters, and hands off to human staff for pricing. Odoo ERP retains rates, rooms, and invoices — this system acts as the front concierge.
Diving package pricing fix was pending prod sync yesterday. Now built and deployed directly to Production.
Bot Re-asks Stated Details (Over-Logicalized)Resolved in Code
Guest writes: '1 person dives day 1, 5 people dive both days'. Old slot logic failed to collapse to one integer divers and re-asked: 'How many people are diving?'Now completely resolved!
Zero Naturalness ScoringPhase 2 Target
Rigidly measured hallucination (0%), but never scored conversational charm. Needs 30 real transcripts from Eloa to establish objective grading.
2. How Does Hybrid AI Solve This?
Neuro Layer (AI): 5-Star Concierge VoiceShipped
Eliminated static robot templates. AI acts as a warm 5-star concierge: greets politely and explains itineraries warmly in EN, VI, or ZH based solely on verified facts.
verifySynthesizedReply audits every AI sentence: blocks any currency quote ($, ₱, PHP), false confirmation, or count mismatch — auto-rolling back to safe templates.
New Decision Rule: NEVER RE-ASK STATED FACTSShipped (442 Tests)
When guests describe nuances, the bot never re-asks and preserves guest counts. It stores the verbatim schedule in diveNotes and triggers getStaffAlerts for human handoff!
Slide 2 / 2 · Delivery Roadmap & Decisions
Upcoming Phases Roadmap & Critical Blockers
Focusing on 3 consecutive phases: Phase 1 is now 100% completed in code & production; Phase 2 unblocks Odoo and naturalness scoring; Phase 3 delivers full quotation automation.
Phase 1 · Completed Today (100% Live)
✓ Hotfix Deploy & Guardrails
✓ Local pricing uses the customer's real rate card: room by occupancy, one dive line per day at diveTiers, meals 1,500, 30% off rooms only — deployed to production.