Model ROC-AUC
0.9388
Held-out test set
CREDIT RISK INTELLIGENCE
Good morning, Analyst.
A clear view of your credit risk model and assessment workspace.
PR-AUC
0.8939
Positive class: default
Brier Score
0.0552
Lower is generally better
Decision Threshold
55%
Configured model threshold
Model performance
Saved held-out test metrics.
Metrics have different definitions; interpret them in context.
Decision threshold
Classification uses the configured cutoff.
This threshold is a model classification rule, not a universal lending standard. A lower-risk prediction does not guarantee repayment.
Recent assessments
Predictions made during this browser session.
| Assessment | Applicant profile | Probability | Class | Risk label |
|---|---|---|---|---|
| No assessments yet. Create one to see it here. | ||||
APPLICANT WORKSPACE
New credit assessment
Enter a sample applicant profile and request a prediction from your live FastAPI model.
MODEL TRANSPARENCY
Model insights
Held-out test metrics and development-time global SHAP summaries.
ROC-AUC
0.9388
Higher is generally better
PR-AUC
0.8939
Default is positive class
Brier Score
0.0552
Lower is generally better
Log Loss
0.2037
Lower is generally better
Global SHAP feature importance
Mean absolute SHAP values measured on a validation sample.
Loan / income
0.0837
Annual income
0.0607
Home: RENT
0.0406
Loan grade A
0.0292
Loan grade D
0.0266
Interest rate
0.0245
Loan grade C
0.0212
Responsible interpretation
Important limitations of model explanations.
How to read these values: Higher mean absolute SHAP values indicate greater average influence on model output in the evaluated sample. They do not show whether a feature raises or lowers risk, and they do not establish causation.
Test metrics summarize a held-out evaluation; real-world performance can differ due to population shift and data quality.
Before real use, evaluate fairness, subgroup calibration, data protection, drift, governance, and applicable regulations.
These are saved development-time SHAP summaries. This UI does not calculate per-applicant SHAP values.
CONNECTIVITY
API status
Check whether FastAPI and the trained model are available to this browser.
Checking backend...
Requesting GET /health.
Connection settings
Local Docker API defaults to port 8000.
If frontend and backend have different origins, configure FastAPI CORS for this frontend origin.
PROJECT OVERVIEW
About CreditGuard AI
Educational end-to-end machine learning for credit default risk estimation.
Technology stack
Components powering this project.
ModelCalibrated Random Forest
APIFastAPI + Pydantic
ExplainabilitySHAP
ContainerizationDocker + WSL2
FrontendHTML · CSS · JavaScript
Model disclaimer
Use responsibly.
Educational demonstration only. This is not a standalone lending decision tool. Predictions may be inaccurate or biased and must not be the sole basis for decisions affecting individuals. Do not enter real sensitive applicant data.
HOW IT WORKS
From applicant data to risk estimate
Understand the main stages behind a CreditGuard AI assessment.
Applicant details
Enter the requested financial and loan-related features in the assessment form.
Model inference
The API validates the input and passes supported features to the calibrated model.
Risk estimate
The model returns an estimated default probability for the supplied profile.
Threshold comparison
The probability is compared with the configured 0.55 threshold to produce the demo classification.
HELD-OUT TEST RESULTS
Model performance at a glance
Saved evaluation metrics for the current calibrated Random Forest model.
Measures how well the model ranks positive cases above negative cases.
Summarizes precision and recall across classification thresholds.
Measures squared error in predicted probabilities; lower is generally better.
Penalizes inaccurate probability estimates, especially confident errors; lower is better.
Metrics describe this evaluation dataset and do not guarantee performance on new applicants.
RESPONSIBLE AI
Use predictions with care
This project is an educational demonstration, not a production lending system. Model predictions can be inaccurate, biased, or unsuitable for a different population.
- Do not use a prediction as the sole basis for a lending decision.
- SHAP can help inspect feature contributions, but does not prove fairness or causality.
- Use synthetic or non-sensitive example data in demonstrations.
- Real-world deployment requires validation, privacy safeguards, fairness testing, monitoring, and human oversight.