Check your indicative loan eligibility in seconds.
Enter a few details and get an indicative eligible amount, interest rate and repayment schedule. See the About the Model tab for what's under the hood.
How it works
What Finsta QuickLoans is
Finsta QuickLoans is a concept for an instant personal-loan app: check eligibility, accept an offer, and (in the real product) get verified and disbursed in around 10 minutes, without paperwork. This demo implements the eligibility check for real; the verification and disbursal steps are simulated so you can see how the full journey would work.
Check eligibility
Enter your income, credit score and existing obligations. The credit-risk model estimates your PD instantly and shows an eligible amount and rate.
Accept the offer
If the requested amount is within your eligible limit, you can accept it and move to verification.
Automatic verification
In production, the app would fetch your identity (PAN/Aadhaar), credit bureau (CIBIL) report and bank statement automatically via approved data partners, and only ask you to upload documents when it can't fetch something automatically.
Cross-verification
The fetched CIBIL, bank and identity data would be cross-checked against what you entered, to confirm the application before money moves.
Disbursal
Once verified, the loan is disbursed to your bank account — the whole journey aiming for under 10 minutes end to end.
Check Loan Eligibility
These are the application-time details used by the Finsta QuickLoans website-compatible credit model.
How the eligible amount is calculated
Why is my PD what it is?
The factors below had the biggest effect on your PD, compared with a typical applicant in the training data.
These are computed directly from the model for this specific application. For features the model relies on less, the effect can be small or not entirely intuitive, since the training data is synthetic.
My Loans
Your accepted applications and completed loans will appear here.
No active loan
Accept an eligible offer to create an application.
Completed Loan — Demo
₹30,000 • 6 months • Fully repaid
About the Model
What this project is, what model powers it, and how it was built and checked.
What this is
Finsta QuickLoans is an academic/portfolio project that demonstrates an end-to-end credit-risk workflow: a machine-learning model estimates an applicant's Probability of Default (PD), and a lending policy turns that PD (together with credit score) into an eligible loan amount, an interest rate, or a decline. It is built for learning and demonstration only and is not a real lending product.
The model
The model takes in details such as age, income, credit score, card utilisation, existing loans and EMI, employment and education, and outputs a probability of default. It was deliberately kept free of gender as an input, since it added no measurable accuracy and pricing by gender is best avoided. A handful of fields on the form (application channel, prior default) are recorded but do not feed the model; a reported prior default is instead an automatic decline under policy.
A note on "other monthly income": it is the least influential of the 20 inputs (0.68% of total feature importance). Across the whole dataset, applicants who report any other income actually default slightly less (7.2% vs 8.7%) — but that is mostly because people with side income also tend to have higher salaries and net worth, and it's those factors doing the real work. Once those are held fixed, the model's own residual response to this one field can run the other way for some applicants. Because of that, this field is left out of the "Why is my PD what it is?" explanation on the results page, even though it remains an input to the PD calculation itself.
How it was validated
On accuracy specifically: only 8.5% of applicants in the data actually default, so a model that simply guessed "no default" every time would already be 91.5% accurate without learning anything. That is exactly what the 91.5% figure above is — it matches guessing, so it is not evidence the model works. Flagging an applicant as high-risk at a lower, 15%-probability cutoff (closer to the real default rate) gives a fairer read: 84.6% accuracy, correctly catching 47.7% of the applicants who go on to default. ROC-AUC and the calibration figures below don't have this problem and are the more meaningful checks for this kind of model.
The model was benchmarked on data it never saw during training (a held-out 20% split) and again with 5-fold cross-validation for stability. It was also compared against a plain logistic regression, a gradient-boosted model, and credit score alone — the simplest of these scored close to the deployed model, which suggests the Random Forest is near the ceiling of what this dataset supports, not that a much better model was left on the table.
From PD to an offer
The model only outputs a probability of default. A separate, fixed policy then converts that PD (and credit score) into a risk band, an eligible amount — (total monthly income − existing EMI) × 50% × 12 × a risk factor drawn from the PD — and an interest rate between 12% and 36% p.a. Applications with a PD above 40% are not accepted. These policy numbers are reasonable assumptions for a demo, not figures calibrated against real lending losses.
| PD (risk score) | Risk band | Amount risk factor | Rate impact |
|---|---|---|---|
| Below 5% | Very Low Risk | 1.00 | Closest to 12% |
| 5% - 10% | Low Risk | 0.90 | Slightly above 12% |
| 10% - 15% | Moderate Risk | 0.75 | Mid-range |
| 15% - 25% | High Risk | 0.50 | Rising toward 36% |
| 25% - 30% | Very High Risk | 0.25 | Approaching 36% |
| 30% - 40% | Very High Risk | 0.25 | Fixed at 36% |
| Above 40% | Not accepted | 0 (declined) | — |
Within the 12-36% range, the exact rate also depends on credit score: at the same PD, a lower credit score pushes the rate up (see "How it was developed" for the formula).
Built with AI assistance
This project was built with the help of an AI assistant (Anthropic's Claude), used as a development and analysis tool throughout — from writing and debugging the training code and this website, to proposing and running the validation checks above (cross-validation, calibration tables, comparisons against simpler models). All decisions on what to build, which model to keep, and how to price loans were reviewed and made by the project's author; the AI did not choose the outcomes on its own.
How it was developed
- Started from a synthetic dataset of 100,000 customers with a known outcome (defaulted or not).
- Excluded identifiers and any column that is a direct output of the outcome, to avoid the model "cheating" by seeing the answer in disguise.
- Split the data 80/20, trained the model on the 80%, and checked its predictions against the unseen 20% before ever training on the full dataset.
- Compared it against simpler alternatives (a plain logistic regression, gradient boosting, and credit score alone) to make sure the chosen model was actually earning its complexity.
- Checked that predicted risk matches real-world-style outcomes: on the hold-out data, the model's average predicted PD was 8.4%, close to the actual 8.5% default rate.
- Exported the trained model so the exact same calculation runs in your browser when you use this site — no separate server call, and nothing you enter is uploaded anywhere.