Solution case study

RealityAI

Floor-plan-aware multimodal property discovery.

Role: Floor-plan POC, multimodal retrieval, ranking, backend API & interaction architecture
Period: 2025
Status: Completed
Stack: GPT-4 · CLIP · Qdrant · Vector Retrieval · Reranking · Backend APIs

Reality AI product concept showing a natural-language property query, floor-plan-aware retrieval and ranked recommendations — Reconstructed case-study visual based on the implemented pilot and its interaction design; it is not an original production screenshot.

One-line summary

A property-search pilot that turns natural-language needs, listing content, photos and floor plans into ranked, explainable recommendations that users can refine over time.

At a glance

01Problem: Fixed property filters capture budget and bedroom count but struggle with softer needs such as natural light, quiet streets and spatial layout.
02Context: An 8-person startup team built and internally tested an early property-discovery pilot, with a 3-person backend team.
03My contribution: Validated floor-plan search signals and designed the CLIP retrieval, vector search, rule filtering, explainable reranking, recommendation APIs and interaction architecture.
04Outcome: A pilot that supported natural-language search and iterative preference refinement through limited internal use before development stopped after the founder exited.

Pilot: product maturity
8: people in the startup team
3: people in the backend team

Three-minute walkthrough

Why fixed filters miss how people describe a home

People can usually state hard constraints such as location, price and bedroom count, but their real preferences also include natural light, quiet surroundings, room proportions and relationships between spaces. Traditional search makes them translate those needs into filters, inspect floor plans manually and repeatedly restart the process.

Diagram showing the gap between a user's property preferences and traditional fixed filters — Figure 1. Reconstructed explanatory visual. The product question was how to make soft, spatial and lifestyle preferences searchable without weakening hard constraints.

A recommendation loop, not a one-shot prompt

The pilot interpreted a natural-language request, separated hard constraints from softer preferences, retrieved candidate properties using text and image signals, explained the ranking, and let the user refine the preference state instead of starting over. The goal was to turn housing discovery into an iterative conversation grounded in actual listing data.

Treating floor plans as a feasibility question

The floor-plan proof of concept tested whether room types, orientation cues, relative proportions and layout relationships were reliable enough to help search. It did not aim for CAD-grade reconstruction. The narrower product requirement was to extract signals that could improve retrieval and ranking while keeping their limitations visible.

Connecting multimodal retrieval to a usable product

CLIP placed query text, listing text, property photos and floor-plan images in a shared representation space. Qdrant supplied vector retrieval, while structured filters enforced budget, location, property type and bedroom or bathroom requirements. An explainable reranking layer combined the valid candidates and returned both an ordering and a concise reason for each match.

Reality AI backend architecture highlighting floor-plan validation, multimodal embeddings, retrieval, reranking and recommendation APIs — Figure 2. Reconstructed architecture visual. The highlighted workflow shows the **floor-plan** POC, CLIP representations, vector retrieval, filtering, **explainable reranking** and product-facing APIs within the contributor's scope.

How the contribution fit into the complete system

The retrieval and recommendation layer sat between property data and the user experience. Query parsing and product orchestration supplied intent; the highlighted layer validated floor-plan features, created multimodal representations, retrieved and reranked candidates, and exposed search, recommendation and explanation results to the interface.

End-to-end Reality AI product architecture with the multimodal retrieval and recommendation contribution highlighted — Figure 3. Reconstructed architecture visual showing how property data, user intent, retrieval, recommendation, UI results and preference updates formed a closed loop.

Explaining why each property matched

The mobile interaction concept joined ranked properties with floor-plan-aware match reasons and controls for updating preferences. It represents the intended product experience around the implemented pilot rather than a claim that this exact interface was released publicly.

Reconstructed mobile interface for natural-language property search and explainable ranked recommendations — Figure 4. Reconstructed case-study UI, not an original production screenshot. Property details are illustrative and demonstrate the interaction model.

Pilot evidence and why the project ended

The 8-person startup team reached a pilot that was internally tested and saw limited internal use; three people worked on backend-related development. The project concluded after the founder exited, not because an external adoption or performance study had established success or failure. The evidence here is the implemented pilot, its architecture and the attributable technical work—not a public release, external user adoption or quantified business outcome.