Building Homebound: scoring real estate leads with rules and an LLM

Most lead-scoring tools give you a number and no reason. That is fine until an agent asks "why is this a 91?" and the honest answer is "the model said so."

Homebound scores every lead twice. A deterministic rules pass handles the things that are genuinely rules — source, pipeline stage, recency, budget fit, first-touch urgency. Then an LLM pass reads the lead's actual history and returns a qualitative verdict. The two get blended.

Why not just the LLM

Because it cannot be trusted with arithmetic it does not need to do, and because the whole thing has to keep working when there is no API key. If OPENAI_API_KEY is missing, Homebound falls back to rules-only and says so, rather than failing closed.

Why not just the rules

Rules cannot read "we're pre-approved now, calling Sunday" sitting in a note from three weeks ago. That is exactly the signal an agent wants surfaced.

The output an agent sees is never a bare number: it is a score, a summary, a plain-language reason, and a recommended next action.