Dipesh Karki argues thin-file borrowers aren’t risky, and AI lending works best with minimal, consent-based data, affordability checks and human oversight.
India’s lenders have no shortage of data. What they often lack is the right data, used responsibly. Bureau scores remain the starting point, not the full picture. For millions of first-time borrowers, gig workers and the self-employed, a thin file says little about their ability to repay.
Dipesh Karki, Co-founder & CTO, Vartis Platforms and LenDenClub, believes AI can close that gap, but only if it improves the entire lending journey, not just underwriting. In this conversation with CIO&Leader, he explains why “minimum necessary data” beats maximum collection, how affordability should anchor inclusion, and where humans must stay in the loop. He also discusses explainability, drift and bias under RBI and DPDP expectations, and his five-year vision for AI-native credit, including near-zero NPAs.
CIO&Leader: Vartis Platforms describes itself as building “AI-powered credit infrastructure.” In practice, what does that mean, and how is it different from a lender that uses AI tools for underwriting?
Dipesh Karki: For me, the distinction is whether AI improves one decision or helps improve the entire lending journey. Underwriting is important, but a lending business also needs to service borrowers, track repayments, identify stress and manage collections. Intelligence should connect all these functions.
At Vartis, OmniCredit is our credit intelligence engine, while VartisOne provides the infrastructure for loan origination and management. My focus is on bringing intelligence and execution together, so that what we learn from repayment behaviour improves future decisions.
Having built and operated lending platforms, I see AI-powered infrastructure as a practical requirement: it should help us make better decisions, respond faster and grow with greater control over risk.
CIO&Leader: Credit bureau data still anchors most lending decisions in India. Where does it fall short, and which borrower segments, such as new-to-credit users and gig workers, feel that gap most?
Dipesh Karki: I see bureau data as an essential starting point. But it largely tells us how someone has handled formal credit in the past. We also need to understand whether they can afford the loan today.
This gap matters most for first-time borrowers, gig workers and self-employed individuals. Someone may earn consistently and manage money responsibly without having a substantial credit history. A thin file tells us that we know less about the borrower; it does not, by itself, tell us that the borrower is unreliable.
Our approach is to supplement bureau information with relevant, consent-based financial data, particularly cash flows and existing obligations. The objective is to understand repayment capacity better while maintaining credit discipline.
CIO&Leader: Which alternative data sources have been most predictive of repayment behaviour in your models, and how do you separate signal from noise?
Dipesh Karki: In our experience, the most useful signals are generally those that have a defensible relationship with repayment capacity rather than those that are merely easy to collect. Cash-flow behaviour can help indicate income regularity and financial stability, while existing financial obligations help establish how much of a borrower’s income is already committed.
But more data does not automatically produce better underwriting. One of the most important disciplines in credit modelling knows what data not to use. Every additional variable should have a clear purpose, demonstrate predictive relevance and be appropriate from a consent, privacy and governance perspective.
The way to distinguish signal from noise is to continuously test model predictions against actual repayment outcomes and portfolio behaviour. A variable that appears statistically interesting but does not improve risk differentiation, or that is difficult to justify, should not be included in a credit decision merely because it is available.
Ultimately, technology should improve credit discipline. The goal is not maximum data collection; it is the minimum necessary data that produce better, explainable and responsible risk decisions.
CIO&Leader: Expanding access to credit and protecting portfolio quality can pull in opposite directions. How do you decide how much risk to take on a thin-file borrower, and what guardrails keep inclusion from turning into delinquency?
Dipesh Karki: We believe financial inclusion has to begin with affordability. Extending credit to someone who cannot comfortably repay it does not create meaningful inclusion; it can place that borrower under greater financial stress.
For a thin-file borrower, limited credit history should not automatically result in rejection. Equally, the absence of negative credit history should never be interpreted as evidence of low risk. The real question is whether there is sufficient evidence of repayment capacity. That means evaluating current financial position, income patterns, existing obligations, affordability and other appropriate risk indicators. Technology can make that assessment more granular by helping to distinguish among a lack of historical information, genuine repayment risk, and potential fraud.
It can also help lenders calibrate their exposure. Inclusion does not have to mean giving every eligible borrower the maximum available loan. Appropriate ticket size, tenure and repayment structure are they important risk controls.
The other important guardrail begins after disbursal. Portfolio monitoring and early-warning indicators can help identify emerging repayment stress before it becomes serious delinquency. For us, inclusion and portfolio quality are not competing objectives. Sustainable inclusion means expanding access without weakening underwriting discipline.
CIO&Leader: Beyond approval decisions, how are you using predictive analytics for repayment forecasting and collections, and what measurable improvements have you seen in recovery rates or collection efficiency?
Dipesh Karki: For us, a key opportunity understands who needs attention, when they need it and what kind of engagement will help.
Two borrowers can miss the same instalment for very different reasons. One may have a temporary cash-flow mismatch; another may be facing sustained financial stress. Repayment-risk prediction can help collections teams prioritise accounts and choose a more appropriate response.
At Vartis, our focus is on connecting that intelligence with portfolio monitoring and borrower engagement. NeuraVoice also has a role in making routine communication more scalable.
I would measure the impact through collection cost, repayments after contact and movement between delinquency buckets. Any improvement we attribute to AI should be validated against a comparable baseline.
CIO&Leader: Automation has cut turnaround times across digital lending. Where do you still keep humans in the loop, and how do you ensure consistency and avoid errors at scale?
Dipesh Karki: With automation and AI, we expect people’s roles and responsibilities to grow. As routine work gets automated, teams should focus more on judgement, oversight and improving outcomes.
We see human intervention in two ways. First, proactive intervention in critical flows, such as reviewing complex credit decisions or evaluating fraud signals before taking action. Second, reviewing outcomes afterwards—for example, using actual collection performance to identify gaps and retrain models.
To maintain consistency at scale, we need clear policies, defined escalation paths and traceable decisions. AI can handle more of the execution, while people take greater responsibility for how the system performs.
CIO&Leader: With the RBI’s digital lending guidelines and the DPDP Act tightening expectations, how do you handle explainability, consent, and data privacy when a machine learning model drives credit decisions?
Dipesh Karki: Our ML models operate within the credit policies, rules and risk boundaries defined by experienced credit underwriters. They help apply this framework more quickly and consistently, reducing individual judgement bias. Explainability is embedded so that our teams can examine the factors behind a decision.
We periodically evaluate these models for performance drift, data drift and unintended bias to ensure they continue to operate within the intended framework.
Consent and data privacy are governed through separate controls and processes; they are not decisions delegated to the credit models. Experienced credit underwriters retain overall credit oversight, reviewing model decisions, exceptions, and any flags for potential violations, with support from our compliance and data-governance teams.
CIO&Leader: Models can drift and carry hidden bias, especially when trained on alternative data. How do you monitor performance over time, and how do you test for fairness across borrower groups?
Dipesh Karki: Our models operate within the credit framework defined by experienced underwriters, and we periodically evaluate whether their performance remains aligned with that framework.
We monitor both data drift—changes in the information coming into the model—and performance drift, where predictions become less reliable against actual repayment outcomes. We also compare approval patterns and prediction errors across borrower groups to identify unintended bias.
We run simulations on past borrowers using their historical bureau information and other alternative data available at the time of assessment. Comparing the model’s predictions with actual repayment outcomes helps us evaluate its performance without relying on information that becomes available later.
Explainability helps our underwriters investigate differences and understand the factors influencing decisions. They retain oversight and determine whether policies, data inputs or models need adjustment or retraining.
CIO&Leader: What does a “risk-first” infrastructure require from a technology standpoint, in terms of architecture, data pipelines, and talent, and what mistakes do fast-growing lending platforms most often make when scaling?
Dipesh Karki: Risk-first infrastructure gives teams a reliable view of the lending business and the ability to act quickly when something changes.
This requires connected systems across origination, underwriting, and collections, supported by reliable data pipelines and consistent definitions. Decisions and transactions should be traceable, with clear ownership when something goes wrong.
It also requires people who understand both technology and lending. Our operating experience with LenDenClub and InstaMoney informs how we build VartisOne and OmniCredit.
A common scaling mistake is allowing acquisition and disbursal to grow faster than servicing, reconciliation and collections capabilities. We believe these functions need to scale together to support sustainable growth.
CIO&Leader: Looking five years ahead, how could AI-native lending change credit delivery in India, and where do you see LenDenClub and Vartis Platforms in that picture?
Dipesh Karki: We expect AI to make credit more responsive to borrowers’ financial circumstances. With appropriate consent and reliable data, lenders can better assess repayment capacity, recognise financial stress earlier and offer more suitable loan structures.
AI will also become more proactive in assisting borrowers. It can help people understand the purpose of credit, their repayment responsibilities and the importance of maintaining a good credit record. It can also help them compare available offers and find the most suitable option for their needs and affordability. This should make borrowers more credit-aware and better equipped to make financial decisions.
For LenDenClub, our priorities are responsible access to credit, stronger risk assessment, a simpler borrower experience and greater borrower credit awareness. Our long-term ambition is to move towards near-zero NPAs and a more consistent, risk-aligned cost of capital for borrowers and returns for lenders.
For Vartis Platforms, our ambition is to turn our operating experience into infrastructure that helps other lending businesses achieve these outcomes at scale.