AI Credit Scoring Engine

Lending to people without a long, formal credit history is hard: traditional scorecards have little to work with, so lenders either turn good customers away or take on unmanaged risk. For a mobile-first micro-lending product, decisions also have to be near-instant — a customer applying in an app won't wait days for an answer. Tappo.ai needed a way to assess each applicant's risk automatically, in real time, using the data it actually has, and to turn that into a clear decision and a sensible lending amount.

We built an AI credit-scoring engine that assesses each loan applicant in real time. Instead of relying on traditional credit-bureau history alone, the model learns from behavioural signals — an applicant's history and activity on the platform — to estimate their risk. It produces an easy-to-read credit score and risk level, plus a risk-adjusted view of how much the customer can responsibly borrow, so the lending engine can make fast, consistent decisions. The model is served as its own containerised service using BentoML, so scoring is fast, isolated and scalable, and it slots into the loan-application flow: when a customer applies, the application is scored automatically and the result feeds straight into approval and limit decisions.

Tappo.ai can assess applicants and return a lending decision in real time, automatically — turning a slow, manual underwriting step into an instant part of the app experience. Because the model works from behavioural data, it can responsibly serve customers who a traditional scorecard would simply reject, while keeping risk in view through a clear score, risk level and a risk-adjusted borrowing amount. Packaged as its own service, the scoring engine scales independently of the rest of the platform and can be improved without disrupting it.

• Real-Time Scoring: Each application is assessed instantly as part of the loan flow.
• Behaviour-Based Risk: Uses platform history and activity, not just traditional credit history.
• Clear Score & Risk Level: Turns model output into an easy-to-read score and risk category.
• Risk-Adjusted Limits: Suggests how much a customer can responsibly borrow, not just yes/no.
• Independently Scalable: Served as its own containerised service, isolated from the main platform.

• Python: The scoring service and its data-preparation pipeline.
• XGBoost: The machine-learning model behind the risk assessment.
• BentoML: Packages and serves the model as a fast, standalone service.
• Docker: Containerised deployment so the engine scales on its own.

We approached scoring as a real-time service that the lending engine could call on every application. We built a data-preparation pipeline that turns an applicant's profile and platform activity into the inputs the model needs, trained a machine-learning model to estimate risk from those signals, and translated its output into a clear score, risk level and a risk-adjusted borrowing amount that operators and the lending engine can act on. We packaged the model with BentoML and containerised it so it runs as its own scalable service, then integrated it into the loan-application flow so every application is scored automatically the moment it's submitted.