How we built a proprietary AI core for a luxury real estate platform
The context
An international luxury real estate platform, active in 12 markets with more than 200 affiliated agencies, had a precise objective: AI infrastructure of its own, which competitors could not obtain with a subscription.
Its stack relied entirely on commercial APIs. Every AI task, from property descriptions to market analysis to client communications, went through a single premium model. Costs grew in line with usage, quality varied from one type of task to another, and competitors using the same tools had the same capabilities.
AI Discovery and the PMAI framework
Before any code, the engagement went through AI Discovery, the PiirZ analysis program that precedes technology choices. In three weeks we mapped processes, data flows, pain points and organizational priorities. The output identified 8 high-value AI projects, each with estimates of cost, impact, risk and return, and a prioritized implementation roadmap.
That analysis decided which tasks to optimize, which models to prioritize and where the value was. The whole engagement followed PMAI, the four-phase PiirZ framework (Context, Intelligence, Integration, Evolution) that takes a project from strategic assessment to production.
The architecture
The core of the platform is an Intelligent Prompt Router that works in three steps. It classifies each request by type of task: chat, reasoning, creative writing, data analysis, translation, image generation. It scores the available models on accuracy, speed and cost, with weights set for each task. It sends the request to the best-scoring model, with automatic fallback.
The strategy is OSS-first. Open-source models handle the vast majority of requests, and proprietary models are called only when a quality threshold requires it. The routing decision takes 0.03 ms.
On top of the router runs a mix of large language models and models fine-tuned for luxury real estate: property descriptions in the language of the high-end market, translation across the platform's 8 languages, comparative market analysis trained on regional price patterns, classification of the visual assets in property portfolios.
The roadmap
The core is the foundation for the applications that produce value for the platform and its affiliates. The roadmap includes 14 applications on three horizons.
In production: a property description generator that writes copy calibrated for the luxury segment in 8 languages within seconds, and a market positioning assistant that helps agents price a property against comparable sales.
In development: an affiliate coaching system that analyzes agent performance patterns and proposes improvements, and a client communication tool that adapts tone and content to the buyer's profile and interaction history.
Planned: predictive market intelligence that flags emerging opportunity zones before they show up in traditional data, and a visual staging assistant that proposes presentation changes based on successful sales in the same segment.
Results after 90 days
After the first 90 days in production, AI infrastructure costs were 8 times lower than with the previous approach based only on commercial APIs. The quality score on internal benchmarks stayed between 0.89 and 0.93. All standard tasks go to open-source models; the proprietary fallback is used only for edge cases that need complex reasoning.
The estimated saving is over 140,000 dollars a year on AI infrastructure alone. The platform has moved that budget to the applications on the roadmap.
The platform now owns its AI infrastructure. It does not depend on one provider's pricing decisions, and a competitor cannot obtain the same capabilities by subscribing to the same API.
What the project shows
Luxury real estate, like many traditional sectors, tends to treat AI as a purchase: which tool to buy, which API to subscribe to. In this project the advantage came from the architecture, designed around the business processes.
OSS-first is an engineering decision based on cost, performance and independence. When open-source models reach comparable quality on most tasks at a fraction of the cost, routing follows, and premium models are kept for the cases where they make a measurable difference.

