AI-Powered Realtor Collaboration Platform for Localized Real Estate Markets

The platform brought brokers, sellers and back-office teams together in a shared workspace that was shaped by local zoning, permitting and disclosure regulations. Every listing was handled through this platform.

The product quickly caught on and was used more intensively with each new version. Over time, the structure took on more and more responsibility: quotes became workflows, calendars, documents, and interactive maps, all tightly connected. As the platform matured, each new feature request required deeper intervention into this logic, and the pace of deployment became increasingly delicate. Signs of overload slowly began to emerge: first in the pace of development, then in the reliance on product decisions.

They approached Devox Software with a clear goal: to evolve the system while keeping it fully functional. Each new feature had to be integrated without disrupting ongoing transactions.

• Product Designer
• 2 Full-Stack Engineers
• DevOps Engineer
• QA Engineer
• Project Manager

The inherited CRM became the ceiling for structural evolution. Its foundation was fragile: hand-written SQL queries embedded in PHP templates, with no separation between interface, business logic, and data access. Even the slightest update to the collection model triggered complete table migrations and tedious manual validations. Since there was no schema versioning, a rollback meant downtime.

Collaboration between teams was fragmented. Agents, coordinators, and external providers had no common workspace in which to collaborate. Approvals were granted via endless email chains. Key players like photographers and inspectors were brought in too late, leading to inconsistent readiness.

Geographic information was not included in the system. Overlays of development plans, utility corridors, and walkability indices were in scattered municipal portals. Moreover, each region had its own standards and peculiarities.

Administration lagged behind growth. The platform managed sensitive data — personal data, financial information, transaction logs, licensing documents — without tiered permissions, audit trails, or encryption at rest. As teams multiplied and roles diversified, the attack surface increased without the infrastructure being able to contain it.

And the business logic itself resisted segmentation. Key workflows such as publishing quotes, sending vendors, updating properties were all enmeshed in a single global state with hard-coded dependencies. Trying to untangle them for parallel development or testing led to regressions that were hard to track. Delivery progress was slow. Experimentation, once a driver of innovation, became a risk few were willing to take.

• Frontend: React with custom UI state containers for listings, calendar views, and agent workspaces
• Backend: NestJS with modular services for listings, authentication, messaging, AI processing, and vendor coordination
• Database: PostgreSQL with spatial indexing for map overlays and historical zoning data
• Cache & Sync: Redis used for pub/sub messaging, caching, and real-time session state
• Real-time Transport: WebSockets with dynamic channel naming for multi-session collaboration
• AI Services: Python microservice connected to OpenAI API for natural language generation; Markdown output streamed to frontend
• Geospatial Engine: Mapbox GL JS for client-side tile rendering and region-specific cluster logic
• Cloud Infrastructure: AWS ECS for container orchestration, S3 for media storage, Lambda for batch sync jobs

• Strangler-based rebuild with feature flag routing. We introduced modular services behind an internal gateway, gradually offloading core logic from the legacy CRM. New flows such as listing publication, vendor coordination, and AI-generated content were served via React and NestJS, while read traffic continued shadowing into the old system. Regional cutovers were staged by brokerage zone.
• Geospatial clustering optimization on Mapbox. High-density areas were segmented using dynamic tile boundaries and batched marker rendering. This eliminated client-side freezes in core property corridors and enabled smooth panning across regional clusters.
• AI gateway for automated local insights. A Python service pulled zoning, permit history, infrastructure data, and regional disclosures from public sources. This context was processed through OpenAI models to generate plain-language Markdown summaries, injected directly into listing descriptions without disrupting CMS templates.
• Real-time dual-view sync with Redis pub/sub. Sessions active in the new SPA and those still running in the legacy UI received synchronized updates through channelized WebSocket streams. Shared state was preserved across both versions during the transition window.
• Hybrid infrastructure bridge for legacy parity. Data pipelines connected AWS-hosted services with the bare-metal CRM via secure batch jobs on Lambda. Media assets, listing diffs, and user actions were staged incrementally before final cutover, preserving operational continuity.
• Future-ready architecture with integration hooks. The platform was prepared for integration with MLS feeds, automated valuation models, and third-party PropTech tools. Service boundaries were defined around listing logic, vendor coordination, and client-facing analytics to support extension without rework.

BUSINESS OUTCOMES

• Activation speed. Listing go-live time dropped by 42% via AI-enriched vendor and zoning workflows.
• Team uptake. 90%+ agent and vendor adoption achieved in six weeks with zero onboarding friction.
• Research elimination. Zoning and permit lookups cut by 65% through automated local context injection.
• Continuity assurance. Dual-runtime rollout enabled seamless migration during live operations.
• Admin load drop. Structured flows reduced broker coordination overhead by 38% in regional offices.

TECHNICAL OUTCOMES

• Modular Architecture. Five services (Listings, Auth, Vendors, Messaging, AI) refactored on NestJS and GraphQL.
• Context Engine. Python+OpenAI gateway generated property-specific zoning summaries in real time.
• Live Sync. Redis and scoped WebSockets enabled instant cross-UI updates across legacy and SPA.
• Scoped Access. Fine-grained token system enforced visibility rules across roles and brokerages.
• Zero Disruption. Blue-green deployments secured cutovers with zero client-visible impact.
• Secure by Design. Role-scoped auth, audit logging, and rate-limiting embedded across service mesh.