Case Study — Model Fine-Tuning & Custom RAG
JLL: Fine-Tuned Foundation Models and Custom RAG for Property Intelligence
Customising foundation models for property discovery and real-estate insight, combining supervised fine-tuning with a domain retrieval layer.
Industry
Real Estate & PropTech
The challenge
JLL needed AI to support property discovery and real-estate insight, which means the system has to reason in the vocabulary and conventions of the domain rather than in general language. A general-purpose model produces plausible prose but is unreliable on domain-specific structure — the categories, comparisons and attribute relationships that property search actually turns on.
Two separate problems were in play. The first was behavioural: the model needed to reason the way the domain reasons. The second was factual: it needed access to current property data. Those need different techniques — fine-tuning addresses the first and cannot address the second, since it teaches behaviour rather than delivering facts.
Cost was an explicit constraint from the outset. A domain-adapted system that improved output quality but increased compute spend per query would not have been viable in production.
The approach
Supervised fine-tuning for domain behaviour
We fine-tuned an open-source foundation model on domain-specific workflows and terminology using LoRA adapters, keeping training memory requirements low enough to iterate quickly. Supervised fine-tuning established the distributional patterns of property reasoning — how attributes relate, how comparisons are framed, what a useful answer looks like in this domain.
A PropTech retrieval layer for current facts
Because fine-tuning cannot deliver current or verifiable facts, a retrieval layer was built over the property data and paired with the tuned model. The division of labour matters: the tuned model handles how to reason and in what shape to answer, the retrieval layer supplies what is actually true right now.
Measuring before optimising
We established a baseline against the untuned model before optimising anything, so each subsequent gain was attributable rather than assumed. That discipline is what made the final numbers defensible rather than directional.
Quantisation and memory optimisation
Once behaviour was correct, inference efficiency became the lever. Post-training quantisation at 8-bit precision, applied against an evaluation set built from the fine-tuned model's own output distribution, cut memory footprint without measurably degrading the domain-specific behaviour that had just been trained.
The result
- Inference speed improved by 30% over the pre-optimisation baseline.
- Memory footprint reduced by 20% through quantisation.
- Compute cost reduced by 30% while retaining fine-tuned domain behaviour.
- Domain reasoning and current property facts are handled by complementary techniques rather than one model doing both poorly.
Facing a similar problem?
Model Fine-Tuning & Custom RAG is one of the four service lines we operate. We can walk through your traffic profile, data constraints and cost targets directly.