Real estate looks like a clean ML problem, you predict a price, score a deal, match a buyer. In practice, it's one of the messiest domains in applied AI; every property is unique, ground truth lags, photos are unstandardized, and human judgment dominates the last mile. Even with frontier foundation models freely available, applying them to actual workflows like underwriting and condition assessment is far harder than it looks — generic LLMs hallucinate on domain reasoning, vision models don't understand "needs rehab," and single-model approaches rarely survive contact with real investor decisions. I'm planning to go through why this domain breaks the standard AI playbook, and share what we've shipped here at Rehouzd - a multi-agent LLM workflows for underwriting, multimodal vision for property condition assessment, and ML valuation pipelines tested against real money.