Building a Workflow-Based Property Search Engine with Gemini 2.0 Flash

When Google shipped its Gemini 2.0 Flash models, our team in Dubai did not wait for a roadmap meeting. Within 24 hours we had a working proof-of-concept: a property search engine that interprets how people actually describe the home they want — and answers like an estate agent who never sleeps. This is the story of what we built, the architecture behind it, and why a workflow-based approach is the right way to bring AI into real-estate technology.

The vision: search that thinks like an estate agent

Dubai's property market moves fast, and buyers rarely think in dropdown menus. They describe a feeling: a quiet two-bedroom near the water, walkable to a metro station, with a balcony big enough for morning coffee. Traditional portals force that human language through a funnel of rigid filters — bedrooms, price band, community — and lose most of the nuance on the way.

Our goal was to close the gap between how a person describes a home and what an automated system can actually act on. We wanted a search experience that thinks like a seasoned Dubai estate agent whilst operating at AI speed and scale.

The challenge

Conventional real-estate search has a structural problem: it asks the user to do the translation. You hold a rich mental picture of your ideal home, then flatten it into checkboxes the database understands. The system never sees the intent — only the lossy approximation.

To do better, the proof-of-concept had to meet four requirements:

1. Harness Gemini 2.0 Flash for fast, low-cost natural-language understanding.
2. Orchestrate multiple specialised AI agents rather than relying on one monolithic prompt.
3. Process complex, conversational requirements without forcing them into predefined filters.
4. Deliver results that hold up at production scale, not just in a demo.

Technical architecture: a workflow-based method

Rather than a single giant model call, we designed a three-tier workflow that blends human-style expertise with AI execution. Each tier is responsible for one job, and the output of one feeds cleanly into the next.

The core components

• Intelligent Query Processing — turns a free-text description into a structured understanding of intent.
• Multi-Agent Orchestration — coordinates the specialised agents and routes work between them.
• Keyword Extraction Agent — pulls the signals that matter (location, lifestyle cues, must-haves) from messy input.
• Property Research Agent — searches and gathers candidate properties against those signals.
• Listing Generation Agent — assembles clear, human-readable results from raw property data.
• Result Synthesis Engine — ranks, de-duplicates and composes the final response.

The power of workflow-based processing

Breaking the problem into independent yet interconnected steps unlocks three properties that a single prompt cannot offer: parallel processing of work that has no dependency on each other, incremental refinement where each step sharpens the last, and flexible adaptation when a query takes an unexpected turn.

Each step in the pipeline implements a small, typed contract. Conceptually, every stage is a generic WorkflowStep<T, R> — it takes a typed input, performs one responsibility, and returns a typed result that the next stage can rely on. That uniform shape is what makes the agents composable.

This proof-of-concept leans on the same foundations we use in production engagements: a strong TypeScript core, multi-agent orchestration, and the kind of pragmatic AI engineering that ships. It is the practical edge of our work, not a lab experiment.

Beyond basic search: resilience and safety

A demo that works once is easy. A system that holds up is the hard part, so we built in production-minded guarantees from the start:

• Error resilience — failed agent calls retry with exponential backoff instead of collapsing the whole request.
• Type safety — typed contracts between steps catch malformed hand-offs before they reach the user.
• Performance — asynchronous processing lets independent agents run concurrently, keeping latency low.

Everything runs on a modern, edge-ready stack. The orchestration layer is built with the Vercel AI SDK on Next.js, giving us streaming responses, typed tool calls, and the async primitives needed to run agents in parallel without blocking the user.

Our workflow innovation

Three ideas make the architecture more than the sum of its agents. Dynamic orchestration lets the pipeline adjust in real time as it learns more about a query. Multi-agent collaboration means agents can exchange context rather than working blind. And progressive refinement treats search as a conversation that narrows, the way a good agent reads a client across a viewing.

Integration with modern AI

Today the system combines natural-language processing, vector-based matching and multi-dimensional analysis to connect intent with inventory. With Gemini 2.0 Flash in the loop, the near-term roadmap opens up further:

• Multi-modal analysis — reasoning over floor plans and photography, not just text.
• Complex interpretation — handling layered, conditional preferences in a single query.
• Market prediction — surfacing timing and value signals alongside matches.

Conclusion

Pairing a workflow-based architecture with modern AI is a genuine step change for property search — and we proved it in a single day. The future of real-estate technology is not a smarter filter; it is smart, adaptable systems that meet people in their own language. For a Dubai market that rewards speed and nuance in equal measure, that is exactly the edge that matters.