Randunu Dimeshan · AI Product Manager · Colombo, Sri Lanka

A full-service AI assistant for real estate agents and agencies. Each agent gets a personal assistant tuned to their brand and tone of voice. It creates social posts, graphics and email newsletters, posts on a schedule, and helps find leads.

When an agent signs up, the system pulls in all of their MLS listings and keeps them up to date, then builds a posting schedule around them automatically.

The product was good and the waitlist was long, with hundreds of brokerages waiting. But the core creative engine wasn't predictable. The same input gave slightly different output every time, and brokerages need on-brand content they can trust.

The way the work ran made it harder:

• Close to a thousand work items with no structure for picking what came next.
• Tight deadlines and frequent demos.
• Manual QA, where people compared each output to its reference design by eye.
• A team spread across time zones 12.5 hours apart.

• Made output deterministic. We rebuilt the engine around a templating system that gives the same result every time and still lets each brand customize it.
• Built an intake-to-delivery process. Every request had a clear path from idea to demo, so the backlog became a plan.
• Moved to small work cycles. They were shorter than sprints, and each one ended in a usable, demo-ready piece.
• Automated QA. We built an internal tool that checks generated designs against the reference designs, replacing manual visual checks.
• Made time zones work for us. Core overlap hours plus a solid async routine kept the team moving around the clock.

You don't always need to run an extremely tight ship. This startup moved at lightning speed. We built solid processes where they mattered, like intake, QA and demo-ready cycles, but we also cut a lot of red tape so people could move fast, especially before demos.

A bit of unorganized chaos is fine when your teammates are experienced. The job is knowing which parts need structure and which parts need room.

Over nine years, each new client had been given its own features, built straight into the core product instead of configured per client. The result was a system with more than 300 features, and the client wanted all of them modernized on a short timeline.

• Years of team changes meant nobody had the full picture, not even the client.
• There was no API or technical documentation.
• Most of the app was server-side rendered with very few REST APIs, so learning how it behaved meant digging through the system itself.

I made AI a core part of how we worked.

• Reverse-engineered the system. I used AI to work out how the backend and its financial calculations behaved.
• Documented 300+ features end to end, where there had been no documentation before.
• Built AI skills for QA runs, PRD writing and design checks.
• Automated a backlog of almost 10,000 Jira tickets, so we knew exactly where the project stood at any moment.

We chose not to do a complete rewrite. The foundations were never built right, and fixing them properly meant changing the underlying infrastructure, which the timeline didn't allow. So we deferred the deep backend rework to a later phase.

Within those limits, we went after the slowest paths that were in our reach: the backend fixes that didn't need new infrastructure. That work alone cut system response times by 70%.

• Fix the backend first. I'd correct the calculations and pull them into one standard library, instead of keeping several versions of the same logic.
• Bring AI in on day one. I'd set up a shared "project brain" the whole team can ask about the system, plus the skills, infrastructure and work plans. We lost months to manual discovery and QA.
• Rebuild less. Not every one of the 300+ features gets used. I'd ship a smaller set that is 100% accurate and much faster, use it to win new clients, and grow from there.