From MLS feed to multi-platform listings in under 2 minutes.
Monday: 3 prompts for listing variations. Tuesday: automated code pipeline. Wednesday: agent/broker workflows. Thursday: complete technical architecture. Agent orchestration, MLS integration, variation generation, and distribution to 50+ platforms with quality checks and compliance.
Key Assumptions
- •Process 100-10,000 listings per day per brokerage
- •MLS data updates hourly via RETS/RESO API
- •Support 50+ distribution platforms (Zillow, Realtor.com, social, email, print)
- •Fair Housing Act compliance required (no discriminatory language)
- •Multi-tenant SaaS: 1,000+ brokerages, each with 10-500 agents
- •99.9% uptime SLA for listing publication
- •GDPR/CCPA compliance for consumer data
System Requirements
Functional
- Ingest MLS data via RETS/RESO API (hourly sync)
- Generate 50+ platform-specific variations (web, social, email, print)
- Validate Fair Housing compliance (no discriminatory terms)
- Distribute to platforms via APIs (Zillow, Realtor.com, etc.)
- Track performance metrics (views, leads, conversions)
- Support custom branding per agent/brokerage
- Handle image optimization and watermarking
Non-Functional (SLOs)
💰 Cost Targets: {"per_listing_usd":0.15,"per_variation_usd":0.003,"storage_per_gb_month_usd":0.023}
Agent Layer
planner
L4Decompose listing publication into tasks, select tools, route to specialized agents
🔧 task_decomposer, platform_selector, agent_router
⚡ Recovery: Retry with backoff (3x), Fallback to default plan, Alert ops team
executor
L3Execute primary workflow: generate variations, orchestrate sub-agents
🔧 template_engine, llm_api (GPT-4), image_optimizer, watermarker
⚡ Recovery: Partial success (publish what succeeded), Retry failed variations, Queue for manual review
evaluator
L2Validate output quality, check completeness, score variations
🔧 quality_scorer, completeness_checker, style_validator
⚡ Recovery: Flag low-quality variations, Trigger regeneration, Human review queue
guardrail
L1Policy checks, Fair Housing compliance, PII redaction, safety filters
🔧 fair_housing_checker, pii_detector, profanity_filter, legal_validator
⚡ Recovery: Block publication on violation, Alert compliance team, Log for audit
template
L2Generate platform-specific variations using templates and LLMs
🔧 template_renderer, llm_api (GPT-4), image_selector, seo_optimizer
⚡ Recovery: Fallback to base template, Use cached similar listing, Generate generic version
distribution
L3Publish approved variations to 50+ platforms via APIs
🔧 zillow_api, realtor_api, social_api (FB, IG, Twitter), email_api (SendGrid), crm_sync
⚡ Recovery: Retry with exponential backoff, Queue failed platforms, Alert agent on critical failures
ML Layer
Feature Store
Update: Hourly for listings, daily for neighborhood stats
- • listing_price_zscore (normalized price)
- • days_on_market
- • price_per_sqft
- • neighborhood_avg_price
- • school_rating
- • walkability_score
- • recent_sales_velocity
- • agent_performance_score
- • listing_quality_score
- • image_quality_metrics
Model Registry
Strategy: Semantic versioning, shadow mode for 7 days before promotion
- • gpt-4-turbo
- • quality-classifier
- • fair-housing-detector
- • image-ranker
Observability
Metrics
- 📊 listing_ingestion_rate
- 📊 variation_generation_time_p95_ms
- 📊 quality_score_avg
- 📊 fair_housing_violation_rate
- 📊 platform_publish_success_rate
- 📊 llm_latency_p95_ms
- 📊 llm_tokens_per_listing
- 📊 cost_per_listing_usd
- 📊 agent_satisfaction_score
Dashboards
- 📈 ops_dashboard
- 📈 ml_dashboard
- 📈 cost_dashboard
- 📈 agent_performance_dashboard
Traces
✅ Enabled
Deployment Variants
🚀 Startup
Infrastructure:
- • AWS Lightsail or Heroku (simple deploy)
- • Managed PostgreSQL (RDS or Heroku Postgres)
- • Redis Cloud (free tier)
- • OpenAI API (pay-as-you-go)
- • S3 + CloudFront (images)
- • Serverless workers (Lambda or Cloud Run)
→ Single region (us-east-1)
→ Single tenant (1 brokerage)
→ Manual onboarding
→ Basic monitoring (CloudWatch)
→ Cost: $150-500/month for 100-1K listings/day
🏢 Enterprise
Infrastructure:
- • EKS/ECS (container orchestration)
- • Aurora PostgreSQL (multi-AZ, read replicas)
- • ElastiCache Redis (cluster mode)
- • Multi-LLM (OpenAI + Anthropic + Azure fallback)
- • Dedicated VPC per tenant
- • Private networking (VPC peering)
- • BYO KMS/HSM for encryption
- • SSO/SAML integration
- • Audit trail (7-year retention)
- • Multi-region (US + EU)
→ Multi-tenant with tenant isolation
→ White-label branding
→ Custom SLA (99.9%+)
→ Dedicated support
→ Data residency compliance (GDPR, CCPA)
→ Cost: $8K-20K/month for 10K+ listings/day
📈 Migration: Start with startup stack. At 1K listings/day, migrate to queue-based workers. At 5K listings/day, containerize and add multi-tenant support. At 10K listings/day, move to Kubernetes with multi-region and enterprise features.
Risks & Mitigations
⚠️ Fair Housing violation (discriminatory language)
Medium✓ Mitigation: Multi-layer guardrails: rule-based filter + ML classifier + human review queue. 100% compliance required before publication. Regular audits of published listings.
⚠️ LLM hallucination (fake property features)
High✓ Mitigation: Cross-reference all factual claims with MLS data. Validate neighborhood stats against public APIs. Flag low-confidence outputs (<0.9) for human review. A/B test variations to measure accuracy.
⚠️ MLS API downtime (no new listings)
Low✓ Mitigation: Retry logic with exponential backoff. Queue for manual sync. Cache recent listings for 24 hours. Alert ops team on prolonged outage.
⚠️ Platform API rate limits (Zillow, Realtor.com)
Medium✓ Mitigation: Implement rate limiting (5 req/sec per platform). Queue excess requests. Spread publications over time. Monitor rate limit headers. Escalate to enterprise API tier if needed.
⚠️ Cost overrun (LLM API costs)
Medium✓ Mitigation: Set cost caps per listing ($0.15 target). Monitor token usage. Cache common variations. Use cheaper models for drafts (GPT-3.5), expensive models for final (GPT-4). Alert on cost spikes.
⚠️ Quality degradation over time (model drift)
Medium✓ Mitigation: Weekly quality monitoring (score distribution, agent feedback). A/B test new prompts/models. Retrain classifiers quarterly. Human review sample (10%) for ground truth.
⚠️ Data breach (agent/buyer PII exposed)
Low✓ Mitigation: Encrypt all data at rest (AES-256) and in transit (TLS 1.3). PII redaction before LLM processing. Role-based access control. Regular security audits. Incident response plan.
Evolution Roadmap
Phase 1: MVP (0-3 months)
Months 0-3- → Launch with 1 brokerage, 10 agents
- → Support 5 platforms (Zillow, Realtor.com, email, Facebook, Instagram)
- → Generate 10 variations per listing
- → Achieve 0.90+ quality score
- → Process 100 listings/day
Phase 2: Scale (3-6 months)
Months 3-6- → Onboard 10 brokerages, 100 agents
- → Support 20 platforms (add Twitter, LinkedIn, SMS, print)
- → Generate 30 variations per listing
- → Achieve 0.95+ quality score
- → Process 1,000 listings/day
Phase 3: Enterprise (6-12 months)
Months 6-12- → Onboard 100+ brokerages, 1,000+ agents
- → Support 50+ platforms (all major real estate, social, email, print)
- → Generate 50+ variations per listing
- → Achieve 0.98+ quality score
- → Process 10,000+ listings/day
- → Multi-region (US + EU)
Complete Systems Architecture
9-layer architecture from presentation to security
Sequence Diagram - Listing Publication Flow
Real Estate Listing System - Agent Orchestration
6 ComponentsReal Estate Listing System - External Integrations
10 ComponentsData Flow
MLS → 50 platforms in 90 seconds
Scaling Patterns
Key Integrations
MLS Integration (RETS/RESO)
Zillow/Realtor.com APIs
CRM Sync (Salesforce, HubSpot)
Image CDN (Cloudinary)
Security & Compliance
Failure Modes & Fallbacks
| Failure | Fallback | Impact | SLA |
|---|---|---|---|
| OpenAI API down | Switch to Anthropic Claude (failover LLM) | Slight quality variation, no downtime | 99.9% |
| MLS API timeout | Retry 3x with exponential backoff, then queue for manual sync | Delayed listing ingestion (up to 1 hour) | 99.5% |
| Fair Housing violation detected | Block publication, alert agent, suggest edits | Listing not published until fixed | 100% compliance |
| Platform API rate limit (Zillow) | Queue for later, spread requests over time | Delayed publication (up to 30 min) | 99.0% |
| Database unavailable | Read from replica, queue writes | Read-only mode, delayed updates | 99.9% |
| Quality score too low (<0.9) | Regenerate variation, escalate to human review | Delayed publication, quality maintained | 99.5% quality |
| Image optimization fails | Use original images, skip watermark | Lower image quality, no branding | 99.0% |
Advanced ML/AI Patterns
Production ML beyond basic API calls