Rini Machine · Internal Overview
How our credit scoring engine makes decisions
Rini Machine combines a transparent rule-based layer with machine learning models to estimate credit risk for micro‑finance borrowers. This page explains the building blocks so that product, risk, and operations teams can clearly articulate how the system works.
Rules, logistic regression, and Bayesian reasoning work together.
Can swap in new ML models (GBM, trees, neural nets) as performance improves.
Rini Machine Overview
Rini Machine is an AI‑powered credit scoring system designed to help banks and microfinance institutions assess the creditworthiness of borrowers in low‑documentation environments. Instead of relying only on traditional bureau scores, Rini ingests application data, behavioural data and alternative data sources to estimate the probability that a borrower will default.
The engine is intentionally layered. A rule‑based layer encodes expert policies and regulatory constraints. On top of that, logistic regression and Bayesian models estimate risk using historical repayment data. Thresholds and anomaly detectors sit at the end of the pipeline to turn scores into clear approve / review / decline decisions.
1. Rule‑Based Layer (Expert Policy)
The rule‑based layer forms the foundation of every decision. It encodes expert knowledge from micro‑finance practitioners and risk teams into simple, auditable rules.
How the rules work
- Data collection: application forms, income and expense data, previous borrowing history, digital behaviour.
- Rule definition: conditions such as minimum income, maximum existing debt, employment stability, sector blacklists, etc.
- Execution: each application is tagged as low / medium / high risk based on the rules before any ML model is applied.
Why this layer matters
- Speed: instant, explainable filters for clearly high‑ or low‑risk cases.
- Interpretability: easy to explain to regulators, auditors and front‑line staff.
- Guardrails: ensures ML models never approve applications that violate hard business or compliance rules.
2. Logistic Regression Layer (Core Risk Score)
Logistic regression is the primary machine learning model in the first production version of Rini. It estimates the probability that a borrower will default on a loan in the next 12–24 months.
How logistic regression is used
- Inputs: income, expenses, household size, existing loans, repayment history, sector, region and behavioural signals.
- Model: a weighted combination of these features passed through a sigmoid function to produce a probability between 0 and 1.
- Interpretation: values close to 1 indicate high risk of default, values close to 0 indicate low risk.
Benefits in Rini
- Transparent: each coefficient shows how a feature impacts risk.
- Stable: well suited for highly regulated environments where interpretability is important.
- Baseline for upgrades: future models (GBM, trees, neural nets) are compared against this baseline using metrics like ROC‑AUC and KS.
3. Bayesian Layer (Probabilistic Adjustment)
The Bayesian layer brings in prior knowledge and continuously updates beliefs as new information arrives. It is especially useful in micro‑finance contexts where data can be noisy or sparse.
Conceptual formula:
Posterior risk ∝ Prior risk × Likelihood(repayment behaviour given this risk level)
As we observe more repayments and defaults in a given segment (e.g. sector or region), the posterior risk moves up or down accordingly.
4. Thresholds & Anomaly Detection
Once scores are produced, Rini converts them into operational decisions using cut‑offs and anomaly checks.
Thresholds
- Scores below a low‑risk threshold are auto‑approved (within policy limits).
- Scores above a high‑risk threshold are declined automatically.
- Scores in the middle band go to manual review.
Anomaly checks
- Sudden income drops, inconsistent employment history.
- Conflicting answers across the application form.
- Behaviour that looks very different from the borrower's peer group.
5. How the Layers Work Together
For each application, Rini orchestrates the layers in a consistent pipeline:
- Apply rule‑based policies and eligibility checks.
- Compute logistic regression default probability.
- Adjust the risk estimate with the Bayesian layer.
- Apply thresholds to assign approve / review / decline.
- Run anomaly checks and, if needed, route to manual review.
6. Flexibility, Monitoring & Evolution
Rini is designed so that the underlying models can evolve without breaking trust or compliance.
Continuous monitoring
- Track accuracy, AUC‑ROC and default rates across segments.
- Detect drift when performance degrades for a region, sector or product.
- Schedule retraining and back‑testing before promoting new models.
Path to advanced models
- Gradually introduce models like random forests, gradient boosting and neural networks.
- Always compare to the current logistic‑regression baseline.
- Keep the rule‑based layer in place so explanations remain simple for bankers and regulators.