The product idea worked. Everything around it ran through the founder.

A proof of concept that did its job.

Our client is an early-stage real estate technology company in the United States. Its proof of concept worked. Brokers could sit down with it and recognise their own working day on the screen.

It was real software, not a click-through prototype. The product idea was right, which is the rarest part to get right.

Then the bug list started to grow.

The platform had built up a long list of open defects. Workflows held up on the demo path and broke one step off it. Different modules felt as if they had been designed with different ideas of what the product was.

Features shipped when they felt ready. There was no QA gate, no regression testing and no release notes. Each release could quietly break something an earlier one had fixed, and nobody could say when a behaviour had changed.

Everything ran through one person.

Engineering was organised around the founder. Anyone who needed a decision got one by asking. There was no written roadmap, so priorities lived in one head, and product management was whatever the founder had time for that week.

That works at the very start. It stops working once the product goes in front of brokerages running real transactions, and larger, compliance-minded buyers start asking what happens when the software is wrong.

We started with how work gets done.

Our team took on engineering, product management and the marketing work. The first change was not clever. It was putting a delivery system in place of one person's memory.

• Quarterly planning that turns the founder's goals and real customer feedback into a written, sequenced roadmap.
• Two-week sprints, each with one clear ship goal.
• Written acceptance criteria, and QA inside the sprint instead of after it.
• Release notes on every deployment, including the boring ones.
• A daily status comment on every live ticket, so anyone can see where the work stands.

Then we put the AI where it could be checked.

The proof of concept already had AI: a chat box that answered in general terms about a market it could not actually see. It demoed well, but a broker had no way to check a word of it.

We rebuilt the AI as GenAI modules that read the same cleaned-up listing data as the rest of the product. An answer about a market can be traced back to the listings it came from.

• Market analytics you can ask in plain English, over live MLS data.
• Listing copy generated from the listing record itself.
• Email automation that responds to what a buyer actually did in the product.
• Document processing that pulls fields out of transaction paperwork.

And we hardened what sits underneath.

AI features are only as good as the data under them. An email about a property that went under contract yesterday is worse than no email at all.

So the live MLS data path was rebuilt for throughput. The infrastructure went under monitoring and a regular patching routine. Security issues were tracked until they were actually fixed. None of this shows in a demo. All of it matters when a brokerage arrives with a security questionnaire.

Decisions now come from data.

At the start there was no usage tracking. What to build next came from instinct and whichever customer had called most recently.

We added behavioural tracking, interaction timelines and product-usage dashboards. The roadmap now has evidence under it, and the founder has real numbers to take into customer and investor conversations.

What changed.

The proof of concept is now a live, multi-tenant SaaS platform running on live MLS data, with GenAI modules built into the product rather than bolted on.

It has a release process that does not depend on one person. Brokers have left verified reviews on G2, which are public.

The platform is also built to be multi-MLS-ready: the handling for each data feed is kept separate, so adding a new market should be a matter of mapping and configuration rather than a rebuild.

Where they are now.

A year in, the work continues. The first additional MLS market is still ahead, and it will be the real test of that multi-market design.

The usage data is also starting to answer a bigger question: whether the AI modules are why people buy, or why they stay. There is no settled answer yet.