Selected Projects
Causal inference · Criteo Uplift Prediction Dataset · 14M users
Does an ad campaign cause incremental conversions — and whose? A complete causal-inference project: a network-interference simulation, a randomization audit, and uplift modeling on 14 million real users, with uncertainty on every number.
Experimentation · RetailHero (X5) SMS campaign · 200,000 customers
How many users does a test really need, and what happens when you check it early? Variance reduction and sequential testing on a real 200,000-customer experiment, verified against simulations where the true effect is known.
Marketplace experimentation · Simulation built from 19.7M NYC ride-hail trips
When riders compete for the same drivers, a standard A/B test of a promotion measures the wrong thing. A simulation built from 19.7 million real NYC ride-hail trips shows how large the error is, and what a time-window (switchback) design recovers and costs.
Open-source tooling · Python + Streamlit · Diagnostic checks for A/B tests
Most A/B tests aren't invalid because the math is wrong — they're invalid because of what happened around the math: a logging bug that skewed the split, a few extra looks at the dashboard, a shared resource, a segment mix that shifted. Built a tool that runs five checks for exactly those problems — sample ratio mismatch, peeking, interference, Simpson's paradox, missed CUPED opportunity — and returns a plain-English verdict instead of a wall of p-values.
arXiv · 2026
TL;DR: Oja's algorithm for streaming PCA is noisy because every step uses a rank-one gradient. VR-OJA keeps a bias-corrected moving average of those gradients, with a parameter-free momentum default β⋆ = 1 − k/p. It cuts final subspace error by 22–27% over OJA-QR on synthetic streams, and on the S&P 500 benchmark gives 24–47% error reduction at 10× the base learning rate.
Interactive explainers
DS Case Study · Simplified model
A navigation app suggesting a faster route, shown with a simplified model. Drag the slider.
Navigation apps suggest the faster route to each driver. If only a few drivers follow, the shortcut really is fast. But every driver who takes it slows it down, and clears the main road behind them. A test on a small share of drivers measures the left edge of this chart, not what happens when everyone follows the advice. It is the same interference problem I simulate in the marketplace project above: one user's treatment changes the conditions for the others.
Before the app suggested the shortcut, everyone took the main road: 35 minutes.
The shortcut still saves 17.5 minutes over the main road. A small pilot would report exactly this number, and it would be right for the pilot.
Simplified two-road model with made-up travel times (shortcut: 15 min + 40 min × share of drivers on it; main road: 25 min + 10 min × share still on it). It illustrates the idea. It does not describe how Google Maps routes drivers and uses no Google data.
DS Case Study · Simulated numbers
A food-delivery promotion, shown with simulated numbers. Toggle the view.
When a promotion launches, orders go up. But some of those orders would have happened anyway, and some are only pulled forward from the following weeks. The way to tell is a holdout group: users who were eligible but deliberately left out. It is the same idea behind the ad experiment above: compare against what would have happened anyway, not against last week.
| Week | Promo group | Holdout | Extra orders |
|---|---|---|---|
| 1 (promotion) | 14,000 | 12,500 | +1,500 |
| 2 | 11,800 | 12,500 | −700 |
| 3 | 12,300 | 12,500 | −200 |
| 4 | 12,500 | 12,500 | 0 |
| Total | 50,600 | 50,000 | +600 |
Compared with last week, orders are up 17%. But the holdout group, who got no promotion, also ordered more than last week (12,500), so the promotion's real effect that week was +12%: 1,500 extra orders, not 2,000.
Over four weeks the promo group placed only 600 more orders (+1.2%). Many customers moved orders they would have placed anyway into the promotion week, so the following weeks came in below the holdout. Free delivery waived $3 on 14,000 orders, which is $42,000, or about $70 for each order the promotion actually created. Whether that is worth it is a business question; "orders went up" does not answer it.
Simulated illustration with made-up numbers, not data from any company. In a real test the holdout is chosen at random before the promotion starts.
DS Case Study · Simulated numbers
A new answer box above the search results, shown with simulated numbers. Toggle the metric.
A search page can show a direct answer at the top (a height, a date, a conversion). Fewer people then click through to a website. If clicks were the success metric, that looks like a failure. The better questions are whether people got what they came for, and whether that holds for every kind of search. The toggle shows the same simulated experiment under two metrics.
Clicks fall from 77% to 42% of searches, down 46%. If clicks were the goal, the answer box looks like a disaster. For quick facts the drop is dramatic (74% to 22%): people read the answer and leave.
Searches where people got what they needed rise from 70% to 80%. But look inside: for quick facts it jumps from 71% to 91%, while for research questions it falls from 69% to 63%, because a short answer is not enough and some people stop too early. The average hides a feature that helps one kind of search and hurts another, so the decision needs both the right metric and a breakdown.
Simulated illustration with made-up numbers, not data from Google or any company. In practice, "got what they needed" is measured with signals such as no quick re-search, time on a result, or a survey, and each of those is imperfect.
Work
2021 - 2024
SQL data pipelines for brokerage and insurance claims data · Informatica PowerCenter, AQT, Autosys
Degrees
2025 - 2027
Computational Data Analytics Track · GPA 3.83/4.00 · Expected Aug 2027
2017 - 2021
Kolkata, India · CGPA 8.58/10.00
