Work
E-commerce risk, credit underwriting and payments infrastructure. Each one comes down to reading behavioural and transactional signals and making a call you'll be held to.
Data Scientist IAUG 2025 — PRESENT
Data Science InternSEP 2024 — JUL 2025
RISK
Cold-start RTO risk
Customer- and merchant-agnostic return-to-origin model that gates cash-on-delivery at checkout. See the case study below.
CREDIT · BANK-STATEMENT INTELLIGENCE
Recurring transactionsSIGNALIQ ↗
Rebuilt recurring-transaction detection with trigram similarity and union-find clustering. It is more precise, finds more, and costs less to run.
Salary detectionSIGNALIQ ↗
BERT-based pipeline that finds salary and income in raw, noisy UPI narrations.
Income estimation
Models on alternate data for thin-file and new-to-credit borrowers.
INFRA · DOCUMENTS
Traffic forecasting
Benchmarked XGBoost, quantile regression and TFT for payment traffic; the chosen model cut provisioning cost.
Document extraction
LLMs and vision-language models for structured extraction from financial documents.
AS INTERN
Annotation review
Fine-tuned Qwen-2.5 to review annotations, taking load off human reviewers.
Underwriting error analysis
False-positive / false-negative analysis across tens of millions of transactions.
[ CASE STUDY ]
Cold-start RTO risk at checkout
A return-to-origin model that doesn't need to know the customer or the merchant.
CATBOOSTPINE LABSIN PRODUCTION
01 · PROBLEM
Refused COD orders are expensive
When a cash-on-delivery order is refused, it ships back. The merchant pays shipping twice and the stock is tied up in transit. The hardest orders to judge come from shoppers and merchants with no history.
02 · APPROACH
Customer- and merchant-agnostic
A CatBoost model built only on session-, order- and pincode-level features. It can score a first-time shopper at a newly onboarded merchant as reliably as a regular.
03 · DEPLOYMENT
A gate at checkout
The model outputs a calibrated probability, bucketed into high, medium and low risk. High-risk COD orders are offered prepaid only; everyone else keeps COD.
04 · OUTCOME
Fewer returns, same GMV
Return-to-origin dropped with no loss in GMV. Most flagged shoppers switched to prepaid instead of abandoning the cart.
SESSION · how the checkout happened
ORDER · what's in the cart
PINCODE · where it's going
[ PLAYGROUND · COD RISK ]
Should this order ship cash-on-delivery?
The model never sees who the shopper or merchant is. It scores the session (including where the visit came from), the order and the pincode, so a first-time shopper at a brand-new store gets a score like everyone else. Try it.
CUSTOMER HISTORYMERCHANT HISTORYSESSION · ORDER · PINCODE
SESSION
TIME TO CHECKOUT3m
ADDRESS ENTERED
UTM SOURCE
TIME OF DAY
ORDER
ORDER VALUE₹1,800
ITEMS IN CART2
CATEGORY
PINCODE
HISTORICAL RTO IN THIS PINCODE
RTO RISK BANDCATBOOST · AGNOSTIC
MODEL OUTPUTMEDIUM
CALIBRATED P(RTO)0.20
LOWMEDIUMHIGHP = 1.0
CHECKOUT POLICYCOD STAYS ON
Medium risk: COD stays available. Biggest push: order value. Longer checkouts, bigger orders and fuller carts pull the risk down.
WHAT'S PUSHING THE SCORE
ILLUSTRATIVE TOY MODEL. INPUTS AND WEIGHTS ARE MADE UP; THE PRODUCTION FEATURES STAY PRIVATE.SESSIONTime to checkout
+0.65
SESSIONUTM source
+0.60
SESSIONAddress quality
+0.00
SESSIONTime of day
+0.00
ORDEROrder value
+0.75
ORDERItems in cart
+0.43
ORDERCategory
+0.50
PINCODEPincode RTO history
+0.60
UNNATI INDUSTRIAL TRAINEE · MAY — JUL 2024Built a campus vehicle-movement analysis system that turns video and sensor data into operational insights.READ THE WRITE-UP ↗