SME Advisor: AI-Augmented Credit Scoring
Context
Italian SMEs face a structural information asymmetry: banks rely on static PDF balance sheets with 12–18 month latency, creating 'blind spots' that lead to false negatives and latent risk across 4.4 million businesses. UniCredit's legacy scoring model lacks the digital operational context needed for accurate SME creditworthiness assessment.
Task
Design an AI-augmented credit scoring engine integrating alternative data — digital footprint signals, real-time transaction data, and macroeconomic sector indicators — to produce a holistic, explainable, and GDPR-compliant credit score for Italian SMEs.
My Contribution
Architected the 3-Pillar Hybrid Engine: Traditional Financial (50%), Digital Footprint (25%), and Growth Potential (25%), with adaptive dynamic weighting that auto-adjusts per company profile (Established Business, Digital Leader, High Potential Startup, Tech Startup). Built a digital scoring module ingesting Google/Trustpilot reviews, NLP sentiment analysis, and web professionalism signals. Implemented K-Means clustering for peer benchmarking so companies are scored against sector peers, not cross-industry. Designed an anti-fraud seismograph detecting review volatility spikes. Integrated ISTAT macro multipliers and supply chain contagion logic. Applied Knock-Out Gates (Bankruptcy Flag, Revenue < €30k). Implemented the Glass Box Protocol — deterministic weighted-average scoring with full JSON audit trails ensuring EBA compliance.
Outcome
Projected 10–20% reduction in Cost of Risk (15% mid-point), translating to €40–80 million per year in avoided credit losses for UniCredit. Built and demonstrated a live dashboard covering 100 synthetic Italian SME profiles with full scoring transparency.
Key Insights
- —The 'Startup Penalty' problem: static financial scoring penalises young companies with no historical assets — solved via dynamic weighting that shifts emphasis to Digital Footprint (40%) and Growth Potential (35%) for Tech Startups
- —Peer benchmarking via K-Means eliminates the 'Apples to Oranges' problem: a manufacturer with 20 reviews scores at the 90th percentile within its cluster, vs. the bottom 10% if benchmarked against hospitality
- —Anti-fraud seismograph detects manufactured reputation: review volatility > 0.6 triggers a −5 penalty, while organic growth volatility < 0.3 earns a +10 bonus
- —Supply chain contagion propagates sector headwinds: a management consultant whose primary client is in a contracting sector receives a −5 point deduction on the final score
- —The Glass Box Protocol ensures full auditability: every score generates a JSON artifact detailing exactly why each point was awarded, satisfying EBA Explainable AI guidelines