Property Listing: AI Description Generation on GPU

The Challenge: 1,800 Listings and 45 Minutes Each

An estate agency group operating 42 branches across the South East of England lists approximately 1,800 new properties per month — a mix of sales and lettings. Each listing requires a polished property description averaging 250-400 words, highlighting key features, location benefits, and lifestyle appeal. Negotiators currently spend an average of 45 minutes per listing writing and refining descriptions, often after a site visit when details are freshest but time is shortest. With each branch processing 40-50 new listings monthly, description writing consumes roughly 1,350 hours of negotiator time per month — hours that should be spent winning instructions, conducting viewings, and closing deals.

The quality of descriptions varies wildly between branches. Some negotiators write compelling copy; others produce bland lists of rooms. Inconsistent quality affects portal performance (Rightmove click-through rates vary 35% between branches), and the marketing director has no scalable way to enforce the house style across 120 negotiators. Property details, vendor circumstances, and pricing strategies are competitively sensitive — the group will not route them through external AI services where competitor agencies might access the same platform.

AI Solution: LLM-Generated Property Descriptions

A self-hosted open-source LLM fine-tuned on the group’s best-performing property descriptions generates listing copy from structured input: property type, bedrooms, bathrooms, key features, location details, EPC rating, and any additional notes from the negotiator. The model is trained to write in the group’s house style, emphasising lifestyle benefits over dry specifications and maintaining an aspirational but accurate tone.

A vision model processes listing photographs to identify features the negotiator may have omitted — a period fireplace, bi-fold doors opening to a garden, a recently fitted kitchen. These visual insights feed into the LLM prompt, producing descriptions that accurately reflect what the property photographs show. The entire pipeline runs on a dedicated GPU server with vLLM.

GPU Requirements

Property description generation involves processing 8-15 photographs through a vision model, then generating 250-400 words of text. Each listing takes seconds rather than minutes on GPU.

The workload is lightweight — any GPU handles the volume with enormous headroom. Real-time generation (negotiator uploads photos and data, gets a description back in 30 seconds) requires minimal sustained compute. Private AI hosting keeps property and vendor data within GDPR-compliant infrastructure.

Recommended Stack

• vLLM serving Mistral 7B or LLaMA 3 8B fine-tuned on the group’s top 5,000 best-performing descriptions.
• LLaVA or CLIP for extracting property features from listing photographs.
• CRM integration (Reapit, Dezrez, Alto) for pulling property data directly into the generation pipeline.
• Portal API integration for pushing descriptions directly to Rightmove, Zoopla, and OnTheMarket.
• A/B testing framework for measuring click-through rates on AI-generated versus human-written descriptions.

For processing property particulars from acquired estate agencies, add document AI to extract structured data from legacy PDF brochures. Deploy an AI chatbot for applicant enquiries and viewing bookings.

Cost Analysis

Negotiator time spent on descriptions costs the group approximately £40,500 per month (1,350 hours at average negotiator employment cost). AI-generated descriptions reduce this to approximately 300 hours of review and refinement per month, saving £30,000 monthly (£360,000 annually). The recaptured negotiator hours — redirected to revenue-generating activities — are projected to generate an additional £180,000 in annual commission revenue through more viewings and faster instruction wins.

Standardised quality across all 42 branches improves average portal click-through rates, with the group targeting a 15% uplift in enquiry volume from better-performing descriptions. At average conversion rates, this translates to approximately 270 additional sales per year.

Getting Started

Export your 5,000 highest-performing descriptions (measured by Rightmove click-through rate) along with the structured property data and photographs. Fine-tune the LLM on this dataset. Pilot with three branches for one month, comparing AI-generated descriptions against negotiator-written ones on portal performance metrics. Roll out group-wide once AI descriptions match or exceed human performance on click-through rates.

GigaGPU provides UK-based dedicated GPU servers for property technology workloads. Add Stable Diffusion for virtual staging, or scale infrastructure as your branch network grows.