Case Studies

The Challenge

A housing association managing 750+ properties needed to move from reactive to proactive management. Emergency repairs were costly, compliance gaps went undetected, and vulnerable tenants were being identified too late.

Our Solution

We built a comprehensive predictive analytics platform that monitors risk across five domains: property condition, damp and mould, rent arrears, tenant welfare, and regulatory compliance. The system processes 170,000+ data points to generate real-time risk scores for every property and tenancy.

Key Results

• 5-domain risk scoring engine with 20+ predictive indicators
• Real-time alerts for high-risk properties and tenancies
• Automated compliance monitoring for Decent Homes Standard and Awaab's Law
• Predictive maintenance modelling reducing emergency repair costs
• Role-based dashboards for managers, officers, and executives

Technology

The Challenge

Financial markets generate thousands of news articles daily across multiple languages and sources. Manual analysis is impossible at scale, and existing sentiment tools lack the nuance needed for actionable trading intelligence.

Our Solution

We developed a multi-source intelligence pipeline that ingests 82+ RSS feeds across 8 languages, combined with 4 API sources. Articles are triaged using cost-optimised LLM processing, extracting entities, sentiment, causal chains, and market events. The system generates automated daily briefings with PDF reports.

Key Results

• 82+ news sources across 8 languages processed in real-time
• Cost-optimised LLM triage at approximately £0.003 per article
• Entity extraction, sentiment scoring, and causal chain analysis
• 5 automated daily story editions with executive summaries
• Sub-7GB memory footprint for production deployment

Technology

The Challenge

Predicting market movements requires combining diverse signals — news sentiment, technical indicators, macro events, and market data — into a coherent forecasting model that explains its reasoning.

Our Solution

We built an ensemble regression model combining LightGBM, XGBoost, and Ridge regression across a 41-feature prediction engine. The system monitors 127 tickers across equities and crypto, with automated weekly retraining using walk-forward backtesting. Every prediction includes SHAP-based explainability showing the top contributing factors.

Key Results

• 41-feature ensemble model across 3 algorithms with weighted averaging
• 127-ticker watchlist with auto-expanding universe detection
• SHAP explainability for every prediction (top-5 feature attribution)
• 16 Grafana dashboards and 20 Power BI views for monitoring
• Automated walk-forward backtesting and weekly model retraining

Technology