Vela Zanzibar
ZURI: How Vela Zanzibar Qualifies International Property Buyers at Scale
Industry
Real Estate / Luxury PropTechProject type
AI Sales Agent, RAG SystemStatus
Live, 2026Team size
4 specialists
At a glance
What ZURI delivers for Vela:
Scale without headcount
At a 2% engagement rate on a 20,000-visitor day, ZURI handles 400 conversations. On a 40,000-visitor peak at 3%, that's 1,200. A 4-person sales team can't do either alongside their closing and follow-up work.Accurate answers to hard questions
Zanzibar condominium law, ZIPA registration, leasehold renewal, payment milestones — answered from actual sale agreements.Classified in 3 messages
Investor or vacation seeker, classified in 3 messages. Visitor classification adjusts depth, tone, and advance direction based on early signals. Vacation seekers get a graceful redirect. Investors get full ROI engagement.No premature CTAs
Contact gating withholds human handoff until buying signals are confirmed. A conversion decision for luxury buyers.5,000+ eval assertions
RAG accuracy, sales behavior, contact gating compliance, frustration scoring, and long-conversation retrieval persistence all machine-verified before code ships.
About Vela Zanzibar
Vela Zanzibar is a pre-construction luxury residential development located in Paje, on Zanzibar's southeast coast. The development offers four property types: Bahari Studios, Upepo Apartments, Anga Apartments, and Asili Penthouses.
The target buyers are international: European investors seeking yield-generating assets, Middle Eastern buyers diversifying into East African real estate, and lifestyle buyers planning a second home or retirement relocation. Vela's sales team operates in an inherently cross-timezone, multilingual, high-consideration environment.
Founded
Pre-launch, 2025Location
Paje, Zanzibar, TanzaniaMarket
International luxury real estate, off-plan pre-constructionBuyer journey
3-9 months, high-consideration, significant legal and financial complexity
The Challenge
Three structural problems made manual inbound impossible
Vela's website sees peak traffic of 20,000-40,000 daily visitors during paid campaigns. Their sales team has 2-4 agents. ZURI is the AI sales concierge built to close that gap without hiring.
Core business problems
The team faced structural challenges that off-the-shelf tools could not solve:
Volume vs. team capacity
Peak traffic of 20,000-40,000 daily visitors during paid campaigns. At the real estate industry average AI chat engagement rate of 3% on a 40,000-visitor peak, that's 1,200 conversations per campaign day. A team of 2-4 agents cannot handle that alongside their sales responsibilities. According to MasterOfCode (2025), 48% of buyer inquiries receive no response at all when managed manually.Complexity beyond off-the-shelf tools
Buyers ask genuinely hard questions about Zanzibar condominium law, ZIPA registration requirements, leasehold renewal costs, and payment milestone calculations. Generic chatbots do not have these answers.The visitor mix problem
A significant share of visitors are vacation seekers looking for short-term rentals. Without qualification, agents field rental inquiries alongside investor conversations — agent time with no revenue potential.Repeat questions consuming agent time
Prices, payment milestones, and ownership structure questions consuming agent time, with different answers depending on which agent responded.Overnight and weekend inquiries
Inquiries arriving outside business hours going unanswered until momentum was gone. No visibility into which questions visitors were asking or what share converted.
What was hard
The RAG architecture was not the difficult part. Calibrating sales behavior was:
Contact gating calibration
Withholding the human handoff until buying signals are confirmed is straightforward in theory. Getting the threshold right took most of those 26 system prompt iterations. Too assertive and ZURI felt like a pushy chatbot, the exact opposite of Vela's brand positioning. Too passive and qualified buyers who wanted to speak with someone closed the tab without converting. The right threshold isn't a number you calculate; it's one you find through testing against real conversation patterns.Long-conversation RAG persistence failure
Past 30 messages, context window pressure caused the agent to start answering from training data rather than from retrieved documents. The fix wasn't complicated. Finding the failure mode required deliberately pushing ZURI to the edge of its context window in test conditions, something most deployments skip. We added a dedicated eval category to catch that regression on every subsequent code change. Neither of these was anticipated in the initial project scope. Both are now documented in the eval suite, which means any similar project we build inherits those lessons automatically.
What we built
Why not an off-the-shelf chatbot
ZURI is live on Vela's website and has been operating in production since early 2026. The system prompt reached version 2.26.0, representing 26 versioned iterations through live testing and refinement.
Generic real estate AI platforms like Realty AI and Ylopo range from $300 to $5,000 per month and deploy in days. For a standard property listing, that's a reasonable starting point. For Vela it isn't, for four specific reasons.
Why HyperSense:
- Generic real estate AI platforms like Realty AI and Ylopo range from $300 to $5,000 per month and deploy in days. No off-the-shelf platform trains on proprietary sale agreements — they answer from LLM training data, not from Vela's actual contracts.
- None implement visitor classification that adjusts depth and sales approach per buyer type, or contact gating calibrated to luxury brand positioning.
- Generic tools push a raw form entry to a CRM. ZURI submits an AI-generated conversation summary with buyer type, detected intent signals, key topics discussed, and lead temperature.
Our approach:
Knowledge Base Construction
We completed the knowledge base before system prompt development began. We sourced the complete Vela document corpus: sale agreements for all four property types, compiled buyer FAQ, full website content, and Zanzibar legal and regulatory context. VoyageAI generated embeddings optimized for the domain.
Sales Behavior Design & System Prompt Development
ZURI's identity, visitor classification logic, advance hierarchy, contact gating thresholds, intent signal detection rules, and response formatting conventions calibrated for a luxury audience. Through 26 iterations of live testing and refinement, the system prompt reached version 2.26.0.
Evaluation Suite & Production Deployment
We built the RAG accuracy eval suite first: 285 test cases, 1,783 assertions. Then the behavioral eval suite: 58 conversation chains. The Observer analytics layer and CRM integration via Strapi went live at production deployment, making results visible from day one.
What we delivered
Core platform capabilities:
Named AI sales concierge
ZURI operates as a named brand representative, not a generic chatbot. Designed specifically for Vela with calibrated luxury sales behavior.Visitor classification engine
Classifies visitors into four types (investor, lifestyle buyer, vacation seeker, casual browser) and adjusts response depth, tone, and advance direction.Document-grounded RAG
Every factual claim retrieved from actual sale agreements, buyer FAQs, and Zanzibar legal context — not generated from training data.Contact gating system
Deliberately withholds human handoff until buying signals are confirmed. Protects luxury brand positioning through appropriate restraint.AI-generated lead summaries
At contact submission, generates structured summary with buyer type, detected intent signals, key topics discussed, and lead temperature for the sales team.Observer analytics dashboard
Real-time conversation metrics, lead quality scoring, geography distribution, and operational cost tracking per conversation.
Supporting Services:
AI agent architecture
Hybrid deterministic + probabilistic design with LLMConcurrencyGate, ProcessAbortManager, and AgenticLoopOrchestrator for production reliability.Knowledge base engineering
VoyageAI embeddings with pgvector, metadata filters for property type, document tag, and source URL. Self-verification step before every response.Sales behavior design
26 versioned system prompt iterations: visitor classification, advance hierarchy, contact gating thresholds, intent signal detection, and response formatting.CRM integration
Strapi CMS with ContactSupportTool for validated contact capture and AI-generated conversation summaries with buyer type and intent scoring.Evaluation & quality assurance
Custom eval suite: 285 RAG test cases (1,783 assertions), 58 behavioral conversation chains (159 checks), long-conversation retrieval persistence testing.
Architecture & technology
ZURI uses a hybrid deterministic and probabilistic architecture: the LLM drives conversation quality and sales judgment; deterministic systems handle retrieval, lead capture, validation, and reliability.
AI / LLM
Technologies
Anthropic Claude (primary), ModelFallbackManager (secondary)
Purpose
Conversation, sales judgment, advance decisions. Hybrid deterministic + probabilistic architecture.
RAG / Vector DB
Technologies
pgvector, VoyageAI embeddings
Purpose
Semantic search over Vela knowledge base with exact metadata filters for property type, document tag, and source URL.
Lead Capture
Technologies
Strapi CMS, ContactSupportTool
Purpose
Validated contact form + AI-generated conversation summary with buyer type and intent scoring.
Concurrency & Reliability
Technologies
LLMConcurrencyGate, ProcessAbortManager, AgenticLoopOrchestrator
Purpose
Parallel request limits, clean cancellation, 120-second timeout with retry budget.
Evaluation
Technologies
Custom eval suite (Jest, LLM-as-judge)
Purpose
5,000+ assertions: RAG accuracy (1,783 checks across 285 test cases), sales behavior (159 checks across 58 chains).
Analytics
Technologies
Observer dashboard (observer-analytics.service.ts)
Purpose
Conversation metrics, lead temperature, geography, operational costs.
Transport
Technologies
Socket.io
Purpose
Real-time bidirectional conversation delivery.
Key Capabilities
Visitor Classification
How it works
ZURI classifies visitors into four types based on early conversation signals. Investor signals: ROI, yield, rental income, exit strategy, portfolio — full engagement with projections, ROI breakdown, payment schedules, and sales team invitation. Lifestyle buyer signals: retire, escape, second home, floor plan, amenities — warm engagement with property details and viewing invitation. Vacation seeker signals: booking, nightly rate, availability, all-inclusive — graceful redirect to nearby hotels, door left open for future ownership interest. Casual browser: brief, accurate answers at exploratory pace, no escalation.Why this matters
Off-the-shelf chatbots treat a vacation seeker and a portfolio investor identically. That wastes engagement depth for the vacation seeker and provides insufficient depth for the investor. Classification is what makes sales behavior possible.Technical approach
The classification persists for the conversation once established. It's not a rigid decision tree. ZURI reclassifies if later messages change the picture. Investors get full ROI engagement; lifestyle buyers get warm property details; vacation seekers get graceful hotel redirects.
RAG Knowledge Base Grounded in Actual Documents
How it works
ZURI's knowledge base consists of sale agreements for all four property types, compiled buyer FAQ, full website content, and Zanzibar legal and regulatory context. Every factual claim is retrieved from this corpus, not generated from memory.Why this matters
Standard chatbots hallucinate when asked about specific contracts, pricing structures, or legal terms. When ZURI says the payment milestone at 30% construction completion is a specific amount, that figure comes from the actual sale agreement.Technical approach
Before composing any visitor-facing response, ZURI runs a self-verification step: confirm that every cited figure traces to a retrieved search result. If it does not verify, it does not appear in the response.
Sales Behavior Engine & Contact Gating
How it works
ZURI uses second-order intent detection to identify the buying signal behind surface questions. "What's the cancellation policy?" signals pre-purchase risk assessment and treats the visitor as high-intent. "Can I share this with my partner?" signals a multi-person decision and triggers an offer of a brochure for sharing. "When is the handover date?" signals planning around a real purchase, so ZURI asks about their timeline. "What do owners say?" signals validation-seeking before committing, so ZURI shares buyer demographics and offers the sales team.Why this matters
Luxury buyers are high-net-worth, financially sophisticated, and researching a significant purchase. Pushing a "schedule a call" CTA after a first message reads as pressure, not service. Contact gating protects brand positioning.Technical approach
Contact gating deliberately withholds the human handoff until the fifth message or a confirmed buying signal, whichever comes first. ZURI builds trust through specific, accurate, relevant answers first.
Evaluation Framework: 5,000+ Assertions
How it works
Two evaluation layers run on every code change. RAG accuracy evals (eval:accuracy): 285 test cases covering every factual question a buyer might ask, with 1,783 individual keyword and source-file assertions. Sales behavior evals (eval:qualification): LLM-as-judge tests across 58 conversation chains verifying: qualifying questions appear in messages 1-2; contact CTA is absent before message 5 unless buying signals are detected; price sensitivity triggers value reframing, not discounting; frustration scores stay below threshold; life event mentions trigger correct buyer-type classification; and second-order intent signals correctly escalate engagement depth.Why this matters
Every AI agent looks good in a demo. The failure modes live in edge cases: the 31st message in a long conversation, the knowledge base update that shifts a key document, the system prompt change that breaks contact gating for one visitor type.Technical approach
Long-conversation RAG persistence eval specifically tests retrieval after 30+ messages. Combined: 285 RAG test cases with 1,783 assertions, 159 behavioral checks across 58 chains — 5,000+ individual assertions.
Observer Analytics: From Deployment to Managed Channel
How it works
The Observer dashboard provides real-time analytics: total conversations, conversion rate, escalation rate, deflection rate, average confidence score, hot leads count, and period-over-period tracking.Why this matters
This turns ZURI from a deployment into a managed sales channel. The team can see which conversation patterns produce hot leads, which questions generate high frustration, and whether conversion rate is trending.Technical approach
Per-conversation lead temperature (hot, warm, cold), visitor type, conversation outcome, frustration score, visitor location distribution, token cost per conversation, message count, and session duration.
Implementation journey
Phase 1: Knowledge base construction
- Sourced complete Vela document corpus: sale agreements for all four property types
- Compiled buyer FAQ from real Vela inquiry history
- Full website content and Zanzibar legal and regulatory context
- VoyageAI generated embeddings optimized for the domain
- pgvector schema with metadata filters for property type, document tag, and source URL
Phase 2: Sales behavior design & system prompt development
- ZURI identity, visitor classification logic, and advance hierarchy design
- Contact gating thresholds and intent signal detection rules
- Response formatting conventions calibrated for luxury audience
- 26 versioned iterations through live testing and refinement to version 2.26.0
- Live conversation pattern testing for calibrating sales behavior thresholds
- Brand positioning alignment: assertiveness vs. passivity balance for luxury context
Phase 3: Evaluation suite
- RAG accuracy eval suite: 285 test cases, 1,783 assertions covering every factual question
- Behavioral eval suite: 58 conversation chains testing qualification, contact gating, and intent signals
- Frustration threshold testing and long-conversation retrieval persistence eval
- Eval development ran in parallel with system prompt iteration
- Machine-verifiable quality gate running before every code deployment
Phase 4: Observer & lead pipeline
- Observer analytics layer with real-time conversation metrics and lead quality scoring
- Strapi CMS integration with ContactSupportTool for structured lead submission
- AI-generated conversation summaries: visitor type, intent signals, topics discussed, lead temperature
- Geography distribution tracking for multi-region campaign targeting
- Token cost per conversation, message count, and session duration monitoring
Projected Capacity
These figures apply industry benchmarks to Vela's confirmed first-party traffic data. They are directional estimates based on comparable deployments, not Vela-specific measured results. Observer dashboard actuals will replace them as performance data accumulates.
Pipeline Scale at Peak Traffic
At a 2% engagement rate on a 20,000-40,000-visitor peak day: 400-800 conversations handled by ZURIConservative estimate: 40% MQL rate surfaces ~160 MQLs, 20% SQL rate surfaces ~80 SQLsBest case (Dashly benchmarks, Dec 2025): 76% MQL rate (~304 MQLs), 44% SQL rate (~177 SQLs)These are directional estimates based on comparable deployments, not Vela-specific measured results. Observer dashboard actuals will replace them as performance data accumulates.Speed-to-Lead Advantage
78% of real estate buyers go with the first agent who respondsAverage human response time exceeds 15 hours; teams responding within 1 minute achieve 391% more conversions than those waiting five minutesZURI responds in under 30 seconds, around the clock. 62% of inquiries arrive outside business hours.Comparable Deployments
Sterling Estates (Singapore): 45% increase in inquiry-to-viewing conversion, 60% reduction in sales team workload (as reported by Yogreet Global, the implementation partner)Notar (Norway): 40% FAQ deflection, 15% increase in booked property viewings (as reported by MasterOfCode, 2025)Either pipeline scenario is workable for a 4-person sales team. Without ZURI, capacity caps the pipeline regardless of traffic.
Key Takeaways
AI accuracy is a knowledge base problem, not an LLM problem
A better LLM will not tell your agent about your specific contracts or payment milestones. What makes ZURI accurate for Zanzibar-specific questions is the knowledge base: actual sale agreements, not training data. The retrieval architecture is what grounds the answers. Switching models helps at the margins. Building the right knowledge base is what makes accuracy possible at all.Your eval suite is the only signal between demo and deployment
Every AI agent looks good in a demo. The failure modes live in edge cases: the 31st message in a long conversation, the knowledge base update that shifts a key document, the system prompt change that breaks contact gating for one visitor type. 5,000+ assertions running on every code change are the only reliable signal between a working demo and a working deployment.Every premature CTA is a trust withdrawal
Pushing for the contact form fast is wrong for any purchase that takes weeks or months to decide. The buyer has not finished their research. Pushing the handoff before they signal readiness tells them you want the contact more than you want to help. Contact gating is not a luxury real estate design choice. It is what appropriate restraint looks like in any high-consideration context.
Tell Us About Your Inbound Qualification Problem
If your sales team is fielding more inbound than they can handle, or if inbound quality is inconsistent, we'd like to hear about it. We'll tell you honestly whether a custom AI sales agent makes sense for your situation or whether an off-the-shelf tool would serve you better.
We're happy to connect you with Vela's team if you'd like to hear their perspective directly.
Tell Us About Your Project
Describe what you're building, your timeline, your KPIs, and any technical constraints. We'll review your project and get back to you with a clear next step.Let's talk about your project