TECHNICAL CASE STUDY

21 Days. One AI.
Full Business Automation.

How a solo mortgage professional and an AI agent built a complete business operating system โ€” CRM, LOS, marketing automation, compliance engine, trading bot, and 16 production websites โ€” from scratch.

21
Days Built
1,986
CRM Contacts
5,699
People Records
3,687
Listings Tracked
26
Cron Jobs
16
Live Sites

System Overview

A multi-layer architecture connecting AI reasoning, business logic, data persistence, and external services โ€” all orchestrated by a single autonomous agent.

๐Ÿ“ฑ Telegram / iMessage / WebUI
โ†’
๐Ÿง  Judy (Claude Opus 4)
โ†’
โšก OpenClaw Gateway
โ†’
๐Ÿ—„๏ธ Supabase + Graph API
๐Ÿง 

AI Agent Layer

Claude Opus 4 running on OpenClaw with persistent memory, 26 scheduled automations, sub-agent spawning, and multi-channel I/O. The agent maintains its own memory files, standing rules, and learned behaviors across sessions.

Claude Opus 4 OpenClaw 200k ctx
๐Ÿ—„๏ธ

Data Layer

Supabase PostgreSQL with Row Level Security, 40+ tables spanning CRM, LOS, People Hub, listings, drip campaigns, trading, and analytics. Service-role key for agent access, anon key for frontend auth.

Supabase PostgreSQL RLS
๐Ÿ“ง

Communications Layer

Microsoft Graph API for Outlook email (10K/day capacity), Google Workspace for Gmail, Telegram bot, native macOS iMessage via AppleScript. Real-time email monitoring via webhook subscriptions.

Graph API Telegram iMessage
๐Ÿ”„

Automation Engine

26 cron jobs handling email monitoring, drip campaigns, rate scraping, listing ingestion, compliance checks, trading, CRM sync, and proactive outreach. Mix of 15-minute intervals and daily/weekly schedules.

OpenClaw Cron Python Bash
๐ŸŒ

Web Properties

16 production websites on Netlify with full SEO optimization (JSON-LD schema, OG tags, sitemaps), WCAG 2.2 accessibility, security headers, and AI crawler optimization for SGE/Perplexity/ChatGPT.

Netlify HTML/JS GoDaddy DNS
๐Ÿ”’

Security & Networking

Tailscale mesh VPN for secure remote access. Token + password gateway auth with rate limiting. SPF/DKIM/DMARC email authentication. Bot protection on all public forms. Encrypted credential storage.

Tailscale SPF/DKIM HTTPS

What Was Built

Every system built, deployed, and maintained by a human-AI team of two.

๐Ÿ“Š RealtorCRM

Full-featured CRM with 1,986 agent contacts, interaction logging, pipeline tracking, tag-based segmentation, and automated outreach. React + TypeScript frontend, Supabase backend with RLS.

React TypeScript Supabase

๐Ÿฆ Loan Origination System

Custom LOS tracking loan pipeline from application through funding. Needs list system with 7 templates and 71 checklist items. TRID tracking (broker mode โ€” informational, not compliance-owning).

Supabase React TRID

๐Ÿ‘ฅ People Hub

5,699-record unified people database with de-duplication, merge/unmerge, relationship mapping, activity logging, and cross-system identity resolution. 7 normalized tables.

PostgreSQL RPC SECURITY DEFINER

๐Ÿ’ง Drip Campaign Engine

15 active campaigns with automated email sequences, engagement tracking, unsubscribe handling, and hot lead alerting. Sends via Graph API with HTML templates and compliance footers.

Python Graph API CAN-SPAM

๐Ÿ“ˆ Rate Watch Engine

Daily mortgage rate scraping from LenderPrice, automated rate-watch alerts for clients, co-branded rate sheets, and rate comparison tools. Target rate monitoring with 0.125% increments.

Python Playwright Supabase

๐Ÿค– Paper Trading Bot

Momentum-based equity trading on Alpaca paper account ($100K). 15-minute execution cycles during market hours. Position logging, risk management, and portfolio snapshots to Supabase.

Alpaca API Python Momentum

๐Ÿ“ง Email Intelligence

Real-time inbox monitoring with Microsoft Graph webhooks. Campaign response classification (hot/warm/unsubscribe/OOO). VIP sender detection. Automated bounce processing with CRM flagging.

Webhooks NLP Graph API

๐Ÿ  Listings Pipeline

3,687 active listings ingested via Redfin scraper. Listing agent enrichment for outreach. Per-listing co-branded payment flyers generated with Playwright + Chromium.

Redfin Playwright PDF Gen

๐Ÿ“‹ Compliance Engine

RESPA, CAN-SPAM, TCPA, Reg Z, and state licensing compliance baked into every outreach system. Equal Housing Lender disclaimers, NMLS/DRE numbers, time-window enforcement, opt-out management.

RESPA CAN-SPAM TCPA

26 Autonomous Processes

Running 24/7 without human intervention. The agent monitors, acts, and only escalates when human judgment is needed.

Process Schedule Function Status
Email Monitor */15 min Inbox triage, VIP detection, campaign response classification OK
Calendar-People Sync */15 min Sync calendar events with People Hub contacts OK
Outbox Scanner */30 min Log outbound emails as CRM interactions OK
Trading Bot */15 min (market hrs) Momentum scan, position entry/exit, risk checks OK
Drip Engine hourly (9a-5p) Campaign email sends, engagement tracking OK
Bounce Cleanup hourly Delete bounces, flag CRM records OK
Rate Scraper daily 11am (M-F) Scrape LenderPrice for current mortgage rates OK
Listings Scraper daily 6am Ingest new Redfin listings into pipeline OK
Daily Backup daily 4am Full workspace + database backup OK
Morning Brief daily 6am Daily action items, hot leads, calendar preview OK
Marketing Intel daily 10am Market news, competitor monitoring, content ideas OK
Activity Digest daily 3pm End-of-day summary of all automated actions OK
FB Post weekly Tue Auto-generate and post to Facebook Business Page OK
Security Healthcheck weekly Mon Host hardening audit, vulnerability scan OK
Webhook Renewal every 2 days Renew Microsoft Graph webhook subscriptions OK

Showing 15 of 26 processes. Additional jobs include past-client outreach, agent referral campaigns, web intake summaries, and one-time reminders.

16 Production Websites

All SEO-optimized with JSON-LD structured data, WCAG 2.2 accessibility, security headers, and AI crawler optimization. Compliance-audited for mortgage industry regulations.

Technology Stack

Zero cloud vendor lock-in. Everything runs on a Mac mini in a closet.

๐Ÿง  AI / LLM

Claude Opus 4 (Anthropic) ยท 200K context ยท Adaptive reasoning

โšก Agent Runtime

OpenClaw ยท LaunchAgent daemon ยท Cron scheduler ยท Sub-agent spawning

๐Ÿ—„๏ธ Database

Supabase PostgreSQL ยท Row Level Security ยท 40+ tables ยท Service role API

๐Ÿ–ฅ๏ธ Compute

Mac mini (Apple Silicon) ยท macOS 26.2 ยท Node 25.7 ยท Python 3.9

๐ŸŒ Hosting

Netlify (16 sites) ยท GoDaddy DNS ยท Cloudflare (select domains)

๐Ÿ“ง Email

Microsoft Graph API ยท Google Workspace ยท SPF/DKIM/DMARC

๐Ÿ’ฌ Messaging

Telegram Bot API ยท macOS iMessage (AppleScript) ยท OpenClaw WebUI

๐Ÿ”’ Networking

Tailscale mesh VPN ยท Serve (private HTTPS) ยท Zero-trust access

๐Ÿ”ง Automation

Python ยท Playwright ยท Chromium ยท Bash ยท Microsoft MSAL

๐Ÿ“Š Trading

Alpaca Markets API ยท Paper trading ยท Momentum strategies

๐Ÿ“ฑ Social

Facebook Graph API ยท Page posting ยท Content generation

๐Ÿ—๏ธ Frontend

React ยท TypeScript ยท Vite ยท Bolt.new ยท Vanilla HTML/JS

21 Days of Building

From zero to a fully autonomous business operating system.

Day 1 โ€” Feb 10, 2026
Bootstrap
OpenClaw installed. Telegram connected. Gmail webhooks live. Supabase CRM operational. 37 agents emailed. First warm lead (Krissy Graham) within hours.
Days 2-4
CRM + Email Infrastructure
Microsoft Graph API integrated. SPF/DKIM/DMARC configured. Email signature built. Agent pipeline and interaction tracking live. Drip campaign engine v1.
Days 5-7
Rate Watch + Trading Bot
LenderPrice rate scraper built. Rate Watch portal deployed. Alpaca paper trading bot running. Portfolio tracking in Supabase. Co-branded rate sheets generating.
Days 8-10
People Hub + Listings Pipeline
People Hub v1 deployed (7 tables, merge/de-dupe). Redfin listings scraper ingesting 3,600+ properties. Listing agent flyer generation with Playwright.
Days 11-14
LOS + Compliance + Social
Loan Origination System built. Needs list system (7 templates, 71 items). Full compliance engine. Facebook Graph API integration. Content automation.
Days 15-18
Email Intelligence + Campaigns
Campaign response processor (NLP classification). 15 drip campaigns active. Bounce cleanup automation. Past-client outreach batches. Hot lead alerting.
Days 19-21
SEO Sprint + Migration
All 13 customer-facing sites SEO-optimized in one session. Full compliance audit. Migrated entire system from Linux VPS to Mac mini. iMessage integration live. Tailscale remote access configured.

The Human-AI Operating Model

This isn't a chatbot answering questions. It's an autonomous agent that operates a business.

๐Ÿงญ Persistent Memory

The agent maintains its own long-term memory (MEMORY.md), daily logs, standing rules, and learned behaviors. It wakes up fresh each session but reconstructs context from its files โ€” like a human reviewing their notes. After 21 days, it has 22 daily log files and 340+ lines of curated institutional knowledge.

โš–๏ธ Safety Architecture

Destructive actions require per-item human confirmation. The agent can read, search, and organize freely but must ask before sending external communications. Mass deletion is architecturally prevented, not just instructed against โ€” safety rules are stored in 3 redundant locations to survive context compaction.

๐Ÿ”„ Self-Improving Systems

The agent updates its own documentation, creates standing rules from mistakes, and evolves its processes. The inbox cleanup process was formalized into a repeatable standard during this session. The compliance engine was built after the agent identified regulatory gaps in outreach templates.

๐Ÿ’ฐ Economics

Total infrastructure cost: ~$50/month (Supabase free tier, Netlify free tier, Tailscale free tier, Anthropic API usage). The Mac mini is the only hardware. No SaaS subscriptions. No engineering team. One mortgage professional and one AI agent.

# What a typical day looks like for the agent: 04:00 Daily backup (workspace + database snapshots) 06:00 Generate action items, scrape new listings, morning brief 06:30 Hot leads brief โ†’ Telegram alert if anything urgent 09:00 Drip campaigns begin sending (hourly through 5pm) 10:00 Marketing intel scout โ€” market news, competitor monitoring 11:00 Mortgage rate scraping (weekdays) *:00 Email monitoring every 15 min โ€” classify, triage, alert *:00 Trading bot every 15 min during market hours *:30 Outbox scanner โ€” log sent emails as CRM interactions 15:00 Daily activity digest โ€” summary of everything that happened 17:00 Follow-up checks on pending items 22:00 Quiet mode โ€” only urgent alerts forwarded # Human involvement: ~30 min/day reviewing alerts and approving drafts