Open to senior credit risk and decision science roles. Open to relocation.

I design credit policy and PD models for a multi-billion-dollar business lending book, including Open Banking underwriting launched in the UK and US.

My work decides which businesses are approved, at what price and credit line, and what each loan earns over its life. I report early-delinquency and guardrail metrics to the credit risk committee.

Eligibility rate, at a 2% lower vintage loss rate than the prior strategy: +3%
Lower vintage loss rate from a new default definition, with a realised profit uplift: 2.5%
Open Banking underwriting launched in the UK and US: 2 markets
In data science and credit risk strategy: 5+ years

Book a call Download resume See the work

If you are hiring from abroad

Based in: Bengaluru, India (UTC+5:30)
Relocation: Open to relocation
Work authorisation: Needs employer visa sponsorship
Experience: Over 5 years in data science and credit risk strategy
Roles: Senior Credit Risk Analyst, Senior Decision Scientist, Senior Data Scientist (Credit Risk), Credit Strategy Manager, Underwriting and Credit Policy Manager (SME lending)

Selected work

Five pieces of work, each with the decision it changed.

Each one starts with the context, so you do not need to know the product to follow it. Figures are shown without baselines, and details covered by confidentiality are left out.

01 · Business lending · Credit policy

More approvals at a lower loss rate

Eligibility up 3% and more loans booked, at a 2% lower vintage loss rate than the previous rule strategy.

Context: Every change to credit policy trades approvals against losses. Approving more is only worth it if the loans swapped in perform at least as well as the loans swapped out.
Diagnosis: Eligibility after the bureau check was set by the previous rule strategy. Raising it safely needed a view of exactly which applicants a new strategy would add and remove, and how each group would perform.
Intervention: I redesigned the eligibility rules and built the models behind them. I ran swap-set analysis to compare the applicants swapped in and swapped out against the old strategy, used balance control to determine gross credit loss (GCL), and assigned risk grades.
Outcome: Post-bureau eligibility rose 3% and more loans were booked, with a vintage loss rate 2% lower than under the previous rule strategy.

How the new strategy was tested

Swap-set analysis against the old rules
Gross credit loss under balance control
Risk grades assigned

02 · Business lending · UK and US

Open Banking underwriting rules

Rules adopted into production unchanged and launched. Data failures eliminated within weeks of going live.

Context: Many small businesses have a thin credit bureau file. Open Banking lets a lender read their bank transactions directly, with consent, and underwrite on actual cash flow.
Diagnosis: For thin-file applicants, bureau-based rules have too little to work with. Approve and decline rules had to be built on bank data, and a new data source brings new ways for a decision to fail.
Intervention: For the UK product, I designed the approve/decline underwriting rules. For the US product, I launched Open Banking underwriting and monitored it after launch, tracing each data-extraction failure to its root cause.
Outcome: The UK rules went into the production rule engine as I drafted them and launched. In the US, extraction failures were eliminated within weeks of launch.

From design to production

Rules designed
Adopted unchanged into the rule engine
Launched

03 · Business lending · Portfolio risk

Multi-horizon default definition

Vintage loss rate down 2.5% against the prior policy, with a realised profit uplift.

Context: A lender has to judge whether a loan is good long before it is repaid. The usual shortcut is one early delinquency indicator a few months in.
Diagnosis: I showed that a single early indicator was statistically insufficient to separate good loans from bad in a portfolio with long-tail risk.
Intervention: Designed a default definition that reads delinquency at several horizons across the first year, and tested it against the existing approach in a champion-challenger setup with population stability monitoring.
Outcome: Adopted as the production rule layer. Vintage loss rate fell 2.5% relative to the prior policy, and the profit uplift was realised, not only projected.

Loan performance read at several points, not one

Early read
Mid-term reads
Twelve-month read

04 · Business lending · Portfolio economics

What is a borrower actually worth?

A multi-year lifetime-value model showing which segments earn back their acquisition cost, and which do not.

Context: In fixed-fee lending, the first loan is rarely where the money is made. Most customers borrow again, so the value of an account sits in its renewals.
Diagnosis: Loan-level profit misleads when most value comes from renewals. Decisions on price, credit limits and acquisition spend needed a per-account, multi-year view of value, by segment.
Intervention: I built a multi-year profit and loss per acquired account, by customer type and credit-score band. It walks balances, fee revenue, credit loss, reserves, funding and acquisition costs through to discounted return, with a delinquency stress test. Risk and behaviour curves are forecast by chain-ladder on origination-month cohorts, and every analyst override is logged.
Outcome: Renewals, revenue, loss, cost and return by score band and customer type, over short and long horizons. It sets the score-band cut-off strategy and guides pricing and acquisition decisions.

From loan history to a decision

Cohort triangles by origination month
Chain-ladder forecast of each curve
Every analyst override logged
Profit and loss per account
Cut-off strategy by segment

05 · Business lending · Underwriting

Underwriting decision engine

One engine in place of overlapping legacy rules, with every path resolving to price and credit line.

Context: A decision engine is the set of rules that turns risk scores and applicant data into an approval, a price and a credit line.
Diagnosis: Underwriting ran on legacy rules that had become redundant and overlapping, which made outcomes hard to explain and to change.
Intervention: Designed one engine that consolidates the legacy rules and combines several internal PD models with external bureau scores. It has three paths: bureau-based, Open Banking with documents, and no-document automation.
Outcome: Every path resolves to pricing and line assignment, so one policy governs approval, price and credit line.

What the engine combines

Internal PD models and bureau scores
Bureau, Open Banking and no-document paths
Price and credit line

Approach

Three things the work has taught me.

Model risk is usually an infrastructure problem.

After an Open Banking launch, the failures I had to chase were in data extraction, not in a model. A sound model on a broken feed still makes bad decisions, so I check the pipes before I tune the score.

Bureau data is a floor, not a ceiling.

A bureau file says how a business paid in the past. Bank-transaction data shows the cash coming in now. That is why I have built underwriting rules on Open Banking data for two markets.

Cut-offs belong to the P&L, not the scorecard.

A low probability of default does not make a loan worth booking. In repeat lending most of the value sits in renewals, so I set score-band cut-offs and guide pricing from each segment’s lifetime profit and loss, not from default risk alone.

Experience

Where I have worked.

Nov 2024 – Present
Data Scientist, Global Credit Risk
PayPal, Bengaluru
- Credit policy, PD calibration and loan-level profitability for US and UK business lending. Report early-delinquency and guardrail metrics to the credit risk committee.
- Direct 2 external consultant data scientists.
Feb 2024 – Oct 2024
Senior Manager, Analytics, Risk & Data Science
Liquiloans, Mumbai
- Led a team of 5 (3 full-time, 2 interns).
- Built application, behaviour and propensity scorecards: 10% uplift in approval rate and 0.5% reduction in non-performing assets.
- Built the early-warning system for collections, raising collection efficiency 30%.
Oct 2022 – Dec 2023
Business Analyst, Strategy, Growth & Risk
Jodo, Bengaluru
- Built the company’s first in-house credit risk model, from zero, cutting underwriting turnaround 40%. Built the data warehouse pipeline that automated 80% of ad-hoc reporting. Managed 2 interns.
Jan 2021 – Oct 2022
Data Analyst, Data Science Consulting
Accelera Eloquent, New Delhi
- Segmentation and language models for retail and eCommerce clients. A Bayesian A/B testing framework lifted new-product sales 10%.
2019
B.Tech, Electronics & Communication Engineering
IIIT Guwahati, India

Contact

Hiring for credit risk?

I am based in Bengaluru and open to relocation. I need visa sponsorship, and I am glad to talk through timing on a first call.