DealKept
An always-on AI assistant for real estate teams that catches every promise made to a client across calls, texts and email, and turns it into a one-click follow-up.
What it does
DealKept was an experiment in a particular shape of AI product: an always-on agent that works in the background and only surfaces what needs a human. It listens across a team's calls, texts and email, works out what was promised to each client, ranks the follow-ups, and lets the agent act with one click. It's for US real estate teams of 2 to 10 people, and it is deliberately not a CRM.
The brand is built around Reid, a robin mascot, with master poses, expressions, animation frames and stickers produced from a versioned image-generation prompt library.
How it's built
A Cloudflare-native TypeScript monorepo (Workers, Durable Objects, D1, Queues, Workflows, Vectorize) using Effect, Hono and oRPC. Ninety architecture decision records, and a rule that every screen passes design review before CI will ship it.
The privacy model
Client messages are full of names, so I trained a small name detector to run inside a Cloudflare Worker: a 6-layer BERT distilled from a commercial PII model plus GLiNER, about 15 MB at 4-bit. It lifted recall on third-party names from 88.7% to 94.9%, but precision on non-Latin text fell from 92.1% to 83.5%, so it isn't shipped yet.