Dasa Gang

Credit Passport A person can have a stable income, pay rent on time, send money home and manage a wallet responsibly, yet still have little usable credit history in the country where they are applying. That is the problem Credit Passport addresses. Migrants, students, gig workers and cross-border professionals often leave their financial evidence scattered across bank statements, payroll systems, rent records, wallets, remittance platforms and repayment providers. A traditional bureau file sees only part of that picture. Credit Passport turns permissioned financial records into a portable, explainable assessment that a lender can actually review. How it works The system first converts source records into standardized financial events. Records describing the same activity are reconciled so that a salary deposit and its payroll entry do not count as two separate positive signals. The events are then scored across seven financial dimensions:

Payment track record
Income stability
Room for new payments
Cash buffer
Financial resilience
Income trend
Cross-border consistency The product being considered changes the weights. A credit card gives more weight to payment behaviour and available liquidity. A personal loan gives more weight to income stability and repayment capacity. A student loan gives more weight to affordability and resilience. The score is: [ S_p = \sum_{d=1}^{7} w_{p,d}x_d ]where (x_d) is the applicant’s result in each financial dimension and (w_{p,d}) is the weight for the selected product. The weights for each product sum to 1. The passport score remains transparent. Every result can be traced to a period, a measured value and the records that support it. Alongside the score, we serve a separate 90-day payment-risk signal using a calibrated Histogram Gradient Boosting model. It estimates the probability of an adverse payment event, such as becoming 30 or more days past due. This signal is shown separately from the score so lenders can distinguish an explainable policy assessment from a statistical risk estimate. The product also includes a Transition Guide for situations that commonly disrupt financial continuity: moving abroad, changing jobs, losing employment, returning to India, converting NRE/NRO accounts and updating tax or FEMA-related records. What we built We built the full flow from evidence ingestion to lender review:
Upload a bank statement, spreadsheet, CSV or text-based PDF.
Parse income, obligations, payments, balances, remittances and dates.
Deduplicate records into unique economic events.
Recalculate the product-specific score.
Recalculate the separate 90-day risk signal.
Show the before-and-after result after a new file is uploaded.
Explain which financial dimensions moved and why.
Let a lender review applications in a dedicated queue. A lender can upload a new file and immediately see the effect. For example, increasing salary and on-time payments improved the transparent score. Lowering income, increasing obligations, adding late payments and introducing negative balances reduced it. What we learned The hardest part was not producing one number. It was deciding what the number should mean. We learned that evidence quality and financial behaviour are different things. A complete set of bad records should increase confidence in a negative decision, not improve the applicant’s score. We also learned that product risk matters. The same person can reasonably receive different scores for a credit card, personal loan and student loan because each product exposes the lender to a different repayment pattern. Finally, a risk probability should not be presented as if it were the same thing as an explainable score. The scorecard and the challenger model answer related but different questions. Challenges The main challenges were:
Reconciling duplicate observations from different providers.
Handling records from different currencies and countries.
Distinguishing current evidence from stale evidence after a job or country change.
Preventing irrelevant or invasive signals from entering the score.
Making negative evidence visible without burying the user in technical details.
Supporting editable Excel and PDF inputs without trusting uploaded score fields.
Keeping the system useful when evidence is incomplete.

The prototype was evaluated on 2,434 held-out cases. At a 50% risk threshold, it classified 1,961 correctly, or 80.6%. Its ROC-AUC was 0.852. These are controlled evaluation results, not production underwriting accuracy. Real deployment would require consented lender outcomes, corridor-specific validation, subgroup testing, drift monitoring and governance.