Case Study · AI Engineering · Property Market
London Housing Market Intelligence
A production-grade dashboard turning raw property transactions into forecasts and natural-language answers.
Executive Summary
PERRA built this demonstration using 100,000+ Greater London property transactions from the HM Land Registry (2021–2026). The goal: demonstrate what applied AI engineering looks like in practice. This is not a chatbot bolted onto a report, but a robust data pipeline, dual forecasting models, and an LLM integrated into a unified dashboard.
While the dataset is public and the project serves as a showcase, the underlying engineering pattern is identical to PERRA's client solutions: a structured data layer first, statistical/machine learning models on top, and an AI agent querying the data layer directly rather than hallucinating from memory.
Tech stack: Python, DuckDB, scikit-learn, statsmodels, Google Gemini, Streamlit.
The Dashboard
Key figures, price forecasts and the AI assistant sit on a single screen, fully filterable by London district.
- Key Figures
- Live transaction volume, average & median prices, and quarterly percentage change for any selected district.
- Trend & Forecast
- Historical quarterly price movements alongside a 4-quarter outlook, complete with a live model toggle.
- Ask the Data
- A plain-English query interface powered by deterministic SQL execution behind the scenes.
The Challenge
HM Land Registry data is freely available, but raw CSV exports with hundreds of thousands of rows are inaccessible to non-technical stakeholders. Decision-makers often end up waiting on analysts or static quarterly PDF reports for basic market insights.
PERRA set out to make this data immediately actionable without requiring an analyst in the loop. Four key questions guided the build:
- Market Trends
- How have property prices actually shifted across London boroughs over the past five years?
- Outlook
- Where are average prices heading over the next four quarters, and what is the range of uncertainty?
- Instant Answers
- Can non-technical users query the dataset directly using natural language and receive answers in seconds?
- Trustworthy AI
- How do you guarantee the LLM delivers real, verified database figures instead of inventing plausible numbers?
The PERRA Approach
A modular three-layer architecture, built in strict sequence.
-
01
The Data Foundation Layer
Raw transaction logs are ingested into a DuckDB analytical database with a strict schema. Non-market transactions — family transfers recorded at £1, for example — are filtered out at the pipeline stage, so downstream averages reflect true market value rather than data artefacts.
-
02
Multi-Model Forecasting
Rather than relying on a single prediction, two models run side by side for every district: a linear regression baseline, capturing the macro, long-term market trend, and an ARIMA time-series model, capturing short-term momentum. Displaying both explicitly keeps forecast uncertainty visible and actionable, rather than hidden behind a black box.
-
03
Grounded AI Agent (Text-to-SQL)
This is the core AI engineering feature. When a user asks a question, the LLM translates it into a DuckDB SQL query; that query executes against the actual database, so the model never answers from internal memory; the retrieved result set is fed back to the LLM to compose a concise, human-readable response. If a query fails or returns no results, the agent says so explicitly instead of guessing.
Before: "Ask an analyst, wait for the next scheduled report."
After: "Type the question, get a verified answer traceable to the database in seconds."
Key Capabilities
Data-Grounded Accuracy
Every response is backed by a live, executable SQL query.
Interactive Forecast Toggles
Switch seamlessly between baseline and momentum-based forecasting models.
Filtered & Comparable Metrics
Pre-cleaned datasets ensure fair, apples-to-apples district comparisons.
Zero SQL Barrier
Any team member can explore complex transactional data, effortlessly.
From Open Data to Enterprise Decision Support
The same architecture applies seamlessly to proprietary client data.
Estate agencies, mortgage providers, property funds and municipal authorities all sit on vast amounts of transaction and operational data. By implementing these three layers — clean data layer, predictive models, grounded AI agent — organisations eliminate reporting bottlenecks and give their teams self-service intelligence.
Want AI engineering applied to your own data?
Book a free 30-minute chat to talk through what's possible with the data you already have.
Need an interim AI engineer in London or the UK? Get in touch — PERRA works on fixed-term and project engagements as well as ongoing retainers.