Andrea

Overview

Andrea is an internally developed AI initiative built for the real-estate ecosystem. Rather than adding a general-purpose assistant beside the business, the objective was to place AI inside it — connected to real-estate knowledge, business data, CRM workflows and customer interaction, so that its output is grounded in how the operation actually works.

Challenge

Real-estate enquiries arrive in natural language and rarely map cleanly onto structured search. A client describes a lifestyle, a location, a budget range and a set of constraints in one message. Meanwhile the business holds property information, lead records and operational workflows in systems that were never designed to answer that kind of question. The problem was translation: turning open-ended intent into something the business systems can act on, and returning an answer that reflects real availability rather than a plausible-sounding guess.

Role

Technology ownership — architecture, integration design, security review and delivery of the initiative alongside the engineering team.

Architecture

1. 01

   Intent & Context

   Incoming requests are interpreted for intent and enriched with conversational context so a follow-up message is understood in relation to what came before.

2. 02

   Knowledge Layer

   Real-estate knowledge is represented so that queries can be resolved against domain concepts — location, property type, positioning — rather than keyword matching alone.

3. 03

   Business Data Integration

   Retrieval is bound to live business data through an integration layer, so responses reflect the current state of the platform rather than a static snapshot.

4. 04

   Workflow Automation

   Outcomes route into operational workflows — lead handling, CRM records and follow-up processes — so the interaction produces business action, not only a reply.

5. 05

   Response Composition

   Responses are assembled from retrieved context with grounding constraints, keeping output tied to information the business can stand behind.

Conceptual only. No real topology, configuration or internal architecture is described.

Solution

• Conversational property search grounded in live business data.
• Contextual information retrieval across real-estate knowledge and platform records.
• CRM intelligence — enquiries connected to the records and workflows they belong to.
• Lead processing and workflow automation triggered from conversation.
• API integration binding the AI layer to existing business systems.

Security

• AI access to business data is scoped and mediated through the integration layer rather than granted directly.
• Inputs are validated before they reach downstream systems.
• Operational and customer data handling follows the organisation’s data protection practices.
• Prompts, model configuration, private endpoints and business logic remain internal.

Outcome

• AI embedded inside real operational workflows rather than deployed as a standalone assistant.
• Search and enquiry handling connected to live business information.
• A reusable integration pattern for extending AI into further business processes.

Lessons

• The model is rarely the hard part. Integration, data access and grounding are where the value and the difficulty both sit.
• Business-specific AI needs business-specific evaluation — general benchmarks say very little about whether an answer is useful to a client.
• Treating the AI layer as an untrusted client of internal systems keeps the security model simple as the surface grows.