PropTech Platform Architecture

01

Reference architecture. Multi-tenant operator platform

Most operator platforms are multi-tenant by accident before they're multi-tenant by design. This section is the diagram + tradeoffs we ship with the platform-dev engagements.

• Tenant-per-row vs tenant-per-schema vs tenant-per-database. When each is right
• Postgres RLS for isolation; pgvector for AEO embeddings shared across tenants
• Edge-cached static + per-tenant dynamic via Next.js route groups
• Per-tenant cost telemetry from request 1. Never bolt it on later

02

Multi-locale data + content

Locale-aware platforms are 70% data modelling, 30% UI. The mistake teams make is starting with translation. By then the data model is wrong.

• Per-entity locale-aware fields (vs separate translation tables)
• Per-locale URL structure: /en/coliving vs /coliving with hreflang
• Locale fallback strategy when a translation is missing
• RTL support for Arabic / Hebrew without a CSS rewrite

03

AI-native platform patterns

Bolting AI onto an existing platform produces toy chatbots. Designing AI-native means the agents have first-class access to platform data with auditable boundaries and cost guardrails.

• Per-route + per-tenant cost caps with model fallback (GPT-4o → Haiku → cached)
• Eval suite as part of CI. Prompts can't ship without grade ≥ X
• Tool-use boundaries. Agents can read inventory but never directly mutate finance
• RAG on tenant-scoped pgvector with relevance scoring you can debug

04

Integration sprawl. And the patterns that survive it

MLS, IDX, RESO, Hostaway, Mews, RERA portals, Stripe, Razorpay, Telr, Twilio, Meta Cloud API. Most operators end up with 25+ integrations. The architecture decision is: where do they live, who owns the schema, and how do you swap them?

• Anti-corruption layer. Every external schema gets normalised to your domain
• Webhook ingestion → message queue → idempotent processors
• Per-integration health dashboard with SLA + alerting
• Fallback / degraded mode when an integration is down. Never block ops

05

Avoiding the microservice trap

Conway's Law sells microservices to small teams who can't operate them. The right architecture for most PropTech platforms is a modular monolith with clean module boundaries. Until the org actually warrants microservices.

• Modular monolith. Separate modules per bounded context, single deployable
• Extract a service only when team-of-teams structure demands it
• Async messaging (Redis Streams / RabbitMQ) for cross-module concerns
• Resist deploying 12 services on team of 4. You'll spend more on ops than on product.

06

Observability + cost guardrails

If production cost is a surprise, the architecture is wrong. AI cost especially. Most PropTech AI bills are 3–10× higher than they need to be because nobody set guardrails.

• Per-tenant AI cost dashboard from request 1
• Per-route cost cap with model downgrade fallback
• Sentry / Datadog APM with trace ID propagation through agents
• Quarterly cost reviews tied to per-tenant unit economics