
A single-family rental REIT doing about a billion a year was getting over 1,000 pieces of HOA mail a day, and an 18,000-document backlog had built up. I architected the workflow that reads that mail, led the team that built it (3 engineers, 1 tester and 1 product manager), and reviewed and approved all of the code.
How it works
Paper mail was scanned to PDF, and each new PDF kicked off the workflow. It read the text, sorted the document into one of a few categories the business had defined, and each category ran its own branch. For example, on an HOA violation fine it pulled every mailing address out of the letter (most had several), searched them against the company's property search index to work out which home the letter was about, picked the cause from a fixed list (e.g. weeds, trash, parking), and then looked up whether this was the 1st, 2nd, 3rd or final warning. The result landed on a dashboard the HOA team sorted by whatever columns they needed that day, with Final Warnings always pinned to the top.
What kept it dependable
Every answer from the AI model had to match a fixed format before anything else used it. Low-confidence results went to a review queue for a person. And we ran 200 real mailings from the backlog through the test environment every day, so each day's changes were checked against actual mail.
Results
We built it from Q3 2025 and it went live in Q1 2026. About 98% of mailings were classified correctly and the 18,000-document backlog was eliminated. The business projected $17M a year in operational savings. The HOA team used it every day, and it was still in production when I left.
0 likes · 0 views · Nov 20, 2026
I architected an AI workflow that reads 1,000+ HOA mailings a day for a $1B/yr REIT and led the team that built it. It eliminated an 18,000-document backlog.