Opendoor — Production Architecture on AWS

System Diagram

Full Production Architecture

Five layers: upstream sources → AWS ingestion → Snowflake warehouse → agent tool layer → output delivery

Upstream Sources

AWS Ingestion

Snowflake Warehouse

Agent Tool Layer

Output

📋 MLS / RETS Feed

List price, sale price, DOM, inventory per market & ZIP

CoreLogic · ATTOM · Redfin

💰 Internal Offer DB

Every offer made, amount, accepted/rejected, seller response time

PostgreSQL / RDS

🏠 AVM Service

Opendoor's automated valuation model — market_estimated_value per property

Internal microservice

👥 CRM / Seller Funnel

Lead count, lead-to-offer rate, offer-to-close rate per market & segment

Salesforce / internal

🔨 Acquisition Pipeline

Homes under contract, purchase price, close date, days held

Internal ops DB

🛠 Renovation Tracker

Reno scope, cost actuals, completion date, contractor status

Internal ops DB

📈 Resale / Listing DB

Final resale price, days to sell, actual realized margin per home

Internal transaction DB

🌎 Macro Data Feed

30yr mortgage rate, affordability index, builder cancellation rate

Freddie Mac · MBA · Census

events / batch

⟶

⚡ Amazon Kinesis

Real-time streaming for offer events, acceptance decisions, AVM updates

real-time

📦 Amazon S3

Raw data lake — MLS batch drops, CRM exports, reno cost files

daily batch

🔄 AWS Glue / dbt

Transform raw events into clean, typed, de-duped tables in Snowflake

scheduled

🔒 AWS Secrets Manager

Stores ANTHROPIC_API_KEY, Snowflake credentials, MLS API tokens

security

⏰ EventBridge Scheduler

Triggers monitor-agent daily at 6am. Triggers weekly feedback recalibration.

cron

Snowflake connector

⟶

❄ Snowflake Data Warehouse

raw.mls_listings

Full MLS history — list price, sale price, DOM, ZIP, segment, date

raw.offer_events

Offer price, market_estimated_value, seller response, acceptance boolean

raw.funnel_activity

lead_count, offer_count, accepted_offer_count, close_count per week/market

raw.inventory_status

homes_owned, days_held, reno_status, estimated_hold_cost per deal

mart.market_metrics

Clean weekly snapshot — replaces housing_market.csv exactly

mart.deal_pl

Per-deal P&L: ARV, acq cost, reno cost, hold cost, net margin

mart.feedback_log

Decision Packets + action taken + observed outcome — replaces feedback_log.json

SQL → dict

⟶

🧠 Agent Tool Layer · Lambda

data_loader.py

get_market_summary() · get_market_trend() · detect_anomalies()

analyzer.py

score_market_risk() · rank_all_markets()

deal_scout.py

get_top_deals() · estimate_renovation()

capital_light.py

detect_inventory_surges() · get_contribution_margin_forecast() · rank_top_100_deals()

pricing_engine.py

analyze_pricing_accuracy() · analyze_funnel_drop() · generate_pricing_actions() · estimate_business_impact()

feedback_tracker.py

log_recommendation() · record_outcome() · recalibrate_confidence()

● Claude Opus 4.6

Orchestrates all tools · Native tool use · No wrappers

briefing / alert

⟶

📤 Output Delivery

💬 Slack / Email

Daily briefing + CRITICAL/HIGH alerts pushed to acquisitions channel

💻 Web Dashboard

Decision Packets · Risk rankings · Deal pipeline — live UI

📋 Deal Pipeline JSON

Top 100 deals pushed to acquisition tool for offer generation

📝 Briefing Archive

briefings/ → S3 bucket, version-controlled, Slack-ready markdown

🔄 Feedback Write-back

Action taken + outcome → mart.feedback_log → recalibrate weights

Upstream Systems

All 8 Data Sources

What each system provides, what Opendoor field it maps to, and how often it refreshes.

📋 MLS / RETS Feed

External

list_price, sale_price — what sellers ask vs what homes close at

days_on_market — time from list to sale or delisted

inventory_count, homes_sold — supply/demand balance per market

list_to_sale_ratio (LSR) — acceptance behavior proxy

Refresh: Daily batch → S3 → Glue → mart.market_metrics

💰 Internal Offer DB

Internal

opendoor_offer_price — exact dollar amount offered per property

seller_accepted_offer — boolean outcome per offer

days_to_accept — seller response time (pricing signal)

accepted_offer_count — weekly acceptance volume per market/segment

Refresh: Real-time events → Kinesis → raw.offer_events

🏠 AVM Service

Internal

market_estimated_value — Opendoor's AVM output per property

price_per_sqft — used for ARV and deal scoring

pricing accuracy — offer vs AVM delta triggers misalignment detection

Refresh: Per-property API call at offer time → raw.offer_events

👥 CRM / Seller Funnel

Internal

lead_count — seller inquiries per week per market

lead_to_offer_rate — funnel efficiency at top of funnel

offer_to_close_rate — downstream conversion health

Refresh: Daily batch → S3 → raw.funnel_activity

🔨 Acquisition Pipeline

Internal

acquisition_price — what Opendoor paid per home

days_held — time owned, drives hold cost accrual

inventory_count — homes owned per market right now

inventory_days_on_market — Z-score surge detection input

Refresh: Daily batch → S3 → raw.inventory_status

🛠 Renovation Tracker

Internal

reno_cost_actual — realized cost per property (vs estimate)

reno_condition — cosmetic / moderate / full_gut tier

reno_completion_date — feeds hold cost calculation

Refresh: Daily batch → S3 → mart.deal_pl

📈 Resale / Transaction DB

Internal

resale_price — actual sale price when Opendoor exits the home

resale_margin_actual — closes the feedback loop vs predicted

days_to_sell — post-reno sale velocity, key capital-light metric

Refresh: On transaction close → Kinesis → mart.feedback_log

🌎 Macro Data Feed

External

30yr mortgage rate — rate lock-out seller signal (Ken Zener framework)

builder cancellation rate — new construction competition (Ivy Zelman framework)

affordability index — demand ceiling context per market

Refresh: Weekly batch → Freddie Mac / MBA / Census APIs → S3

Production Data Flow

How Data Moves End-to-End

From upstream event to Decision Packet delivered to the acquisitions team.

▶ Real-Time Path — Offer Events

Seller submits to Opendoor

AVM service evaluates property → returns market_estimated_value. Pricing microservice sets offer band. Offer event fires.

AVM ServicePricing Microservice

Kinesis captures the event

offer_price, market_estimated_value, seller_accepted, days_to_accept streamed in real-time → raw.offer_events in Snowflake.

Amazon Kinesisraw.offer_events

dbt materializes mart views

Aggregates to weekly market/segment snapshots. Computes LSR, acceptance rate WoW delta, funnel conversion rates. mart.market_metrics refreshed.

dbt / Gluemart.market_metrics

Tool layer queries Snowflake

analyze_pricing_accuracy() runs a SQL query against mart.market_metrics instead of reading CSV. Returns same structured dict. Agent behavior unchanged.

pricing_engine.pyClaude Opus 4.6

Decision Packet delivered

If severity HIGH/CRITICAL → Slack alert fires to #acquisitions within minutes of issue detection. Pricing team sees action in under 1 hour.

Slack Webhookmart.feedback_log

▶ Scheduled Path — Daily Briefing

EventBridge fires at 6:00 AM

Triggers monitor-agent Lambda (or ECS task). No human input needed — proactive by design.

EventBridge SchedulerLambda / ECS Fargate

Agent pulls fresh Snowflake snapshots

All mart views reflect last 24 hours of MLS + offer + funnel + inventory data. Tools query via Snowflake Python connector — same interface as CSV.

Snowflake connector

Proactive loop runs all tools

Same 4-step logic: analyze_pricing_accuracy → detect_inventory_surges → CM forecast → rank_all_markets. Claude chains tools autonomously.

monitor.pyNative tool use

Feedback loop recalibrates weekly

Outcomes from mart.feedback_log feed recalibrate_confidence(). Weights updated. Next day's briefing uses improved confidence scores.

feedback_tracker.py

Briefing saved + delivered

Markdown saved to S3 (briefings/ bucket). Slack message posted to #acquisitions-daily. Web dashboard updated. Deal pipeline JSON pushed to acquisition tool.

S3SlackWeb UI

Field Mapping

Demo → Production

Every synthetic field maps to a real Snowflake table. The tool interfaces don't change — only the data source behind them.

Synthetic field	Demo source	Production source	Snowflake table	Refresh
median_sale_price days_on_market homes_sold	housing_market.csv	MLS / RETS feed	mart.market_metrics	Daily
list_to_sale_ratio price_per_sqft	housing_market.csv	MLS feed + computed	mart.market_metrics	Daily
opendoor_offer_price seller_accepted_offer days_to_accept	synthetic_data.py	Internal Offer DB	raw.offer_events	Real-time
market_estimated_value	synthetic_data.py	AVM microservice	raw.offer_events	Per offer
lead_count offer_count accepted_offer_count close_count	synthetic_data.py	CRM / Salesforce	raw.funnel_activity	Daily
inventory_days_on_market inventory_count	synthetic_data.py	Acquisition pipeline	raw.inventory_status	Daily
acquisition_margin_estimate reno_cost, hold_cost	deal_scout.py tiers	Acq pipeline + Reno tracker	mart.deal_pl	Daily
resale_margin_actual days_to_sell	synthetic_data.py	Resale / Transaction DB	mart.deal_pl	On close
prediction_accuracy confidence_weights	briefings/feedback_log.json	Feedback write-back	mart.feedback_log	Weekly

Key production principle

Every tool returns a structured dict. The data source behind the tool is an implementation detail — switching from CSV to a Snowflake query is a one-line change per tool. Claude's reasoning layer, the agentic loop, the Decision Packet format, and the feedback mechanism are all unchanged in production.