2026 · Real estate
Fieldnotes
Turns a recorded real estate showing into structured claims, like preferences, budgets and follow-ups. Each claim is checked against the transcript, and the agent approves the exact note before it is exported for the CRM.
private
The problem
Real estate agents run showing after showing, often with a different client at each. The agent I'm building this for put it simply: by evening, it's hard to remember who said what. Whatever they recall is what goes into the CRM that runs their follow-ups. Fieldnotes pulls those details from the recording instead, keeps only what traces back to it, and labels anything the agent adds.
Why showings are different
A showing has several speakers, so each statement is tied to a role: Client, Client 2, Client 3, the agent or Other. Each recording covers one property. Under Fair Housing rules, protected-class details are kept out of the CRM by default.
What it does
- Upload the showing recording, pick the client and the property, and confirm who each speaker is. Claims are saved to the selected client, and a second or third client's comments stay separately attributed.
- Review claims for that property: preferences, budgets, timelines and proposed follow-ups, each linked to the moment in the transcript it came from. Claims the agent adds are labeled as the agent's, not the client's.
- Approve the claims, pick Zoho or Follow Up Boss, and download the note request formatted for that CRM's Notes API. Nothing is sent.
How it's built
- The model proposes typed claims. The server checks that each one quotes a full turn from the transcript, keeps qualifiers and negation, and names the speaker and the property. The model can't mark its own claims as supported.
- Budgets and dates go through a strict parser. Relative dates such as "in 3 days" resolve from the confirmed capture time, and anything ambiguous stays unknown.
- The preview stores the exact request with a hash tied to the transcript, the context and the checked claims. Any edit means a recheck and a new preview.
- A cautious policy check blocks protected-class content from the final note. It reduces risk and is not legal advice.
- An opt-in local model path runs Qwen3 4B on Ollama, pinned by digest, on loopback only with cloud turned off. Postgres keeps tenants apart with row-level security.
