Estate agency runs on conversations: the first message from an owner thinking about selling, the follow-up a week later, the reminder nobody got round to sending. You would build the agents that handle them.
About Cultiv
A conversation widget sits on an estate agency website. It qualifies a homeowner in a handful of messages, values their property, and drops a lead into the CRM that the agent can pick up the phone about. Follow-up over WhatsApp and email, the valuation report, and the pipeline through to a signed mandate all run on the same platform.
The platform is Next.js and TypeScript on MongoDB, running on AWS. The AI side spans Claude on Bedrock, Azure OpenAI, and a set of automations in n8n. Agencies in Belgium and France use it daily.
The role
You own the agentic layer: the widget conversation, the valuation pipeline behind it, and the follow-up that runs without anyone pressing a button. Expect to spend as much time on evaluation, and on what happens when a model returns something confidently wrong, as on the prompts themselves.
What you’ll do
- Run the widget conversation. It has to qualify a seller in a dozen messages, in Dutch, French or English, without feeling like a form.
- Own the valuation pipeline. Three independent methods (comparable sales, land plus construction, rental yield) each produce a number, and something has to reconcile them into one estimate with a range an agent can defend to a client.
- Write the tools the agents call: cadastre lookups, comparable sales, EPC and permit documents, the CRM itself.
- Build the evaluation harness. Today, whether a prompt change helped is answered by reading transcripts, and that does not scale.
- Watch cost and latency. A homeowner abandons a chat that thinks for eight seconds.
Your profile
📈Experience & background
- A year or more of software engineering, most of it near AI.
- You have shipped an LLM system to real users and stayed responsible for it afterwards.
- Computer science, engineering, or the projects to show for it.
- English is enough to work here. Dutch or French helps you read what our users actually write.
💻Technical skills
- Python for the AI services, and enough TypeScript to work inside a Next.js codebase.
- The Claude, OpenAI and Gemini APIs in anger: tool use, function calling, structured outputs.
- Multi-agent orchestration, plus an opinion about when one well-built prompt beats it.
- Evaluation you can point at. Saying a change helped should not require reading fifty transcripts.
- RAG where it earns its place. Much of what we handle is structured data, not documents.
🧠Mindset
- You are comfortable putting something unfinished in front of a user in order to learn from it.
- You want to be on call for what you built.
- You can tell a model that demos well from one an agency relies on daily.
- You keep up with model releases and can judge which ones are worth acting on.
- You would rather watch an estate agent work for an afternoon than read a spec about one.
Why join us
- You own the AI layer outright. There is no ML team above you to defer to.
- Property data is a genuine mess: cadastre records, EPC certificates, listings that contradict each other.
- Small team, so what you ship is in front of agencies within days.
Our process
- A first interview online with Miguel, Cultiv’s CTO (30 min).
- The technical interview, at our office.
- You meet the rest of the team over lunch.