Applied ML · Causal Inference · Production Systems

Hi, I'm Taru Tak.

Senior Data Scientist with 6+ years building and deploying machine learning for targeting, personalization, and decisioning, from causal uplift models at Nike to agentic AI at Virtuous.

I build causal uplift models, predictive models, agentic AI workflows, and identity resolution systems.

See selected work Resume

$19M–$48Mannual incremental revenue from uplift targeting at Nike
1,500+nonprofits served by the ML platform I built at Virtuous
100M+donor households in enriched identity profiles
5.8×lift over random targeting in out-of-time validation

01 · About

ML that ships and moves a number

I'm a Senior Data Scientist in New York. I build applied ML systems (predictive models, causal uplift targeting, identity resolution, and LLM-powered agents) and own them end to end: feature engineering, validation, retraining, batch inference, and delivery into the tools people use every day.

As Virtuous's founding data science hire, I built the ML foundation behind Virtuous Insights / Donor360, serving 1,500+ nonprofits. Before that, at Nike Consumer Data Science, I replaced propensity-based targeting with causal uplift models across 200+ campaigns, generating $19M–$48M in annual incremental revenue.

I care about measurement as much as modeling: holdouts, out-of-time validation, and explaining results clearly to non-technical stakeholders. I've presented causal-inference methodology to 60–80 senior leaders at Nike Data Science Labs.

02 · Experience

Where I've worked

Virtuous
Sep 2023 – Present · Remote

Senior Data Scientist
- As the founding data science hire, built Virtuous Insights / Donor360, unifying first-party CRM data with third-party and public data into a donor-intelligence platform for targeting, prioritization, analytics, and fundraising across 1,500+ nonprofits.
- Designed and productionized 4 predictive ML models for donor reactivation, acquisition, retention, and gift-ask optimization, owning feature engineering, automated retraining, validation, batch inference, and CRM delivery. Achieved 5.8×, 2×, and 1.7× lift vs. random targeting and 1.9× over the incumbent gift-ask heuristic in out-of-time validation.
- Built Donor360's identity-resolution system, matching fragmented CRM donor records against 260M US consumer records and stitching in public-record, real-estate, demographic, and wealth data for 100M+ donor households.
- Shipped the customer-facing Insights Research Agent from hackathon to production: a multi-step LLM workflow that combines donor profiles with live web research to produce source-cited briefings in 1–2 minutes, with LLM evaluations for quality and grounding.
- Cut data and ML debugging from ~2 days to 2–4 hours with an agentic engineering workflow using Claude Code and tool access across Snowflake, Azure SQL, dbt/GitHub, and Jira.
- Python
- SQL
- Snowflake ML
- Snowpark
- dbt
- XGBoost
- LLM agents
- LLM evaluation
Nike via Launchpad Inc. & TEKsystems
Mar 2020 – Sep 2023

Data Scientist, Consumer Data Science
- Generated $19M–$48M in annual incremental revenue across 200+ campaigns by replacing propensity-based targeting with causal uplift models (EconML/CausalML), using holdouts to isolate customers whose response was incremental to treatment.
- Expanded the SNKRS sell-through model to 3 additional retail geographies, adding product-image and user-product taste embeddings and improving sell-through by 5–8%.
- Built campaign experimentation, audience, and measurement infrastructure in Databricks/PySpark, growing reusable training data from 300K to 8M+ records and cutting audience sizing, A/B testing, and reporting from ~3 days to ~6 hours.
- Developed a Nike By You trend-discovery system using CLIP embeddings, UMAP, and HDBSCAN to surface emerging custom-design patterns and assortment gaps, reducing merchandiser analysis time by ~80%.
- Built a 90-day launch-demand forecasting pipeline with 10% lower forecast error for regional inventory and investment decisions.
- PySpark
- Databricks
- EconML
- CausalML
- A/B testing
- CLIP
- HDBSCAN
- Forecasting
Fellowship.AI Launchpad Inc.
Sep 2019 – Dec 2019

Machine Learning Fellow
- Led 30 ML fellows developing a mobile computer-vision system with PyTorch and fastai, improving model accuracy from 82% to 90%.
- PyTorch
- fastai
- Computer vision
TrueFort Inc.
Feb 2019 – May 2019

Machine Learning Intern
- Built behavioral threat detection with FP-growth pattern mining on Apache Spark; automated client reporting that saved ~17 hours/week.
- Apache Spark
- Pattern mining
- Security analytics

03 · Selected Work

Case studies

Nike · Causal inference

Causal Uplift Targeting

Propensity models target customers who are likely to buy, including many who would have bought anyway. I replaced them with uplift models that target customers whose response is caused by the campaign, running on experimentation infrastructure I built.

Build audience & experiment data in Databricks, growing training data from 300K to 8M+ records
Train uplift models with EconML and CausalML
Target customers with the highest incremental response
Measure lift against holdouts, with reporting cut from ~3 days to ~6 hours

Causal inference
EconML
CausalML
PySpark
Databricks

$19M–$48Mannual incremental revenue across 200+ campaigns

Virtuous · Predictive modeling

Donor Propensity & Gift-Ask Models

Four production models that tell nonprofits which donors to reactivate, acquire, and retain, and how much to ask for. I owned them end to end, from feature engineering to predictions delivered in the CRM.

Engineer donor features from unified CRM and third-party data
Train reactivation, acquisition, retention & gift-ask models
Retrain automatically and validate out-of-time
Deliver batch predictions into the CRM

Predictive modeling
Feature engineering
Batch inference
Snowflake

5.8×lift vs. random targeting; 1.9× over the prior gift-ask heuristic

Virtuous · Agentic AI

Insights Research Agent

A customer-facing research agent I took from hackathon prototype to production. It combines Donor360's unified donor profiles with live web research to write source-cited briefings for fundraisers.

Load the unified donor profile
Run a multi-step LLM workflow with live web research
Generate a briefing with cited sources
Score quality and grounding with LLM evaluations

LLM agents
Tool calling
LLM evaluation
Python

1–2 minto produce a source-cited fundraiser briefing

Virtuous · Data & ML platform

Donor360 Identity Resolution

The identity layer behind Donor360: matching fragmented CRM donor records to real households, then enriching them with outside data so targeting and prioritization work from a complete picture.

Start from fragmented first-party CRM records
Match against 260M US consumer records
Stitch in public-record, real-estate, demographic & wealth data
Deliver enriched household profiles to Donor360

Identity resolution
Snowflake
dbt
SQL

100M+enriched donor household profiles

Nike · Personalization

SNKRS Sell-Through Model

Nike's SNKRS sell-through model, expanded to three additional retail geographies and improved with embeddings that capture what products look like and what members like.

Embed product images
Learn user–product taste embeddings
Incorporate both into the sell-through model
Roll out to 3 additional retail geographies

Embeddings
Computer vision
Personalization

5–8%improvement in sell-through

Nike · Embeddings & clustering

Nike By You Trend Discovery

A trend-discovery system that surfaces emerging custom-design patterns and assortment gaps for merchandisers, replacing much of their manual analysis.

Embed custom designs with CLIP
Reduce dimensionality with UMAP
Cluster with HDBSCAN to find emerging patterns
Surface trends and assortment gaps to merchandisers

CLIP
UMAP
HDBSCAN
Embeddings

~80%less merchandiser analysis time

These summaries cover work already public on my resume. No proprietary data, code, or screenshots.

04 · Skills

Toolkit

Applied machine learning

Causal inference & uplift modeling
Predictive modeling
Experimentation / A/B testing
XGBoost / gradient-boosted trees
Feature engineering
Identity resolution
Embeddings & clustering
Model explainability (SHAP)
Recommender systems
Forecasting

Production ML & AI systems

Model evaluation & monitoring
Batch inference
LLM agents
LLM evaluation
Tool calling
CI/CD

Languages & frameworks

Python
SQL
PySpark
PyTorch
EconML
CausalML

Data & ML platforms

Snowflake ML
Snowpark Python
Snowflake Model Registry
Databricks
MLflow
dbt
Docker

05 · Projects

Earlier projects

Open-source deep-learning work from earlier in my career.

Food-101 Image Classification

Fine-tuned ResNet-18 and ResNet-50 with fastai using multi-stage training to classify 101 food categories.

PyTorch · fastai · CNNs

Skyline Classifier

Deep-learning web app that identifies whether a skyline photo is New York, Chicago, Dubai, or Hong Kong. Containerized with Docker.

Computer vision · Docker · Web app

Humpback Whale Identification

Kaggle competition entry for CS 583 Deep Learning at Stevens: identifying individual whales from photos of their tails.

Kaggle · Deep learning

06 · Education

Education & recognition

2019 · Hoboken, NJ

Stevens Institute of Technology

M.S. Information Systems

GPA 3.85 / 4.0

2017 · Pune, India

Savitribai Phule Pune University

B.E. Computer Engineering

First Class with Distinction

Speaking · Nike Data Science Labs

Causal inference in practice

Presented uplift-modeling methodology and results to 60–80 senior leaders and non-technical stakeholders.

Publication · JETIR, 2017

Enhanced Security Approach for Online User Authentication

QR-code and AES-based login for banking transactions. View code

07 · Contact

Let's talk

Email is the best way to reach me. I'm always happy to talk applied ML, causal inference, and production AI systems.

Emailtaktaru@gmail.com LinkedInin/taru-tak GitHubtarutak

Based inNew York, NY