whoami — rajarshi

rajarshi@portfolio:~$ whoami

Rajarshi Dutta

Data Scientist — Commercial Marketing Decision Sciences, American Express

ようこそ" (Yōkoso)

Gurugram, India
01

About

cat about.md

I am currently working as a Data Scientist in the Commercial Marketing Decision Sciences team here at American Express. focusing on ranking and propensity models for commercial and consumer card offers. I have also worked on an end-to-end project involving the creation and validation of a multi-offer ranking framework for Email channel which replaced 3 XGBoost models with 2 segmentations solutions.

I had graduated from IIT Kanpur with a B.Tech in Materials Science and Engineering, where I also explored courses ranging across domains like Control Systems, Machine Learning, NLP, and Reinforcement Learning. My academic journey involved extensive involvement with Aerial Robotics as a Team Head and other associations like Astronomy, Quiz and Programming club. My interest in Astronomy led me to participate in various Inter IIT Tech Meets with ISRO, working on challenges like Lunar Atlas creation from Chandrayaan Orbiter imagery and Web-Based X-Ray Burst Identification.

Outside of work, I'm usually chasing a good badminton or lawn tennis match. I'm also into photography — mostly street and still life — and have been learning my way around a Fujifilm X-T50. The rest of my time goes to watching films (I log them sometimes on Letterboxd — an ardent cinephile), and occasionally reading murder mysteries as well as fumbling through a tune on the keyboard.

role
Data Scientist — Commercial Marketing Decision Sciences, American Express
based in
Gurugram, India

toolbelt

Programming

PythonC/C++RMATLABBashJavaScript

Databases

SQLPostgreSQLMongoDBSQLiteChromaDBPineconeDB

AI / ML

PyTorchScikit-learnKerasMLflowOpenCVGymnasiumLangChainLangGraphOpenAI

Platform & tools

BigQueryPySparkHiveGCPDockerGitPower BITableauClaude Code
02

Work

git log --oneline experience
v3.0.0Feb 2025 — present

Data Scientist

American Express · Commercial Marketing Decision Sciences, Credit & Fraud Risk

  • +Architected a multi-offer ranking framework on response-to-gross scores with a 96% match rate, replacing 3 XGBoost models with 2 segmentations — $60M revenue impact for the Email channel across commercial and consumer card offers.
  • +Built a Spend-Centric propensity model in XGBoost across 90+ features (spend, risk, contact history, real-time triggers), consolidating 8+ channel-specific models into one deployed for daily scoring across the entire US AIF base — $30M incremental GCM.
  • +Developed an aggregated Contact History variable joined across multiple channel tables, shipped to production with full UAT and a daily refresh cadence — a 3% lift in Gini.
  • +Owned the model end-to-end: base creation, risk suppressions, variable attachment to scoring, tracking, and analytical runs, under model governance and validation.
  • +Led the migration of 3 marketing models from PySpark ETL to BigQuery, refactoring feature pipelines and automating model tracking — 30% runtime reduction.
XGBoostMicrosoft OfficeEDADeploymentModel trackingSHAPPySparkBigQueryPythonDaskGoogle Cloud Platform
v2.1.0Jul 2024 — Dec 2024

Data Scientist

Decision Point Analytics · Pricing & Promotion Modelling

  • +Built a Hybrid RAG sales-insights tool with Reciprocal Rank Fusion (LangChain + Pinecone), using chain-of-thought reasoning over metadata-filtered sales reports — cut manual review time by 5%.
  • +Streamlined SKU-level promotion modeling across 500+ SKUs in 12 markets using discount adjustments, VARMAX, and OLS.
  • +Built an MLflow-based automated forecasting pipeline for monthly SKU-level scoring with periodic model refresh.
LangChainPineconeMLflowVARMAXSARIMAOLSPythonPandasNumPyScikit-learn
v1.0.0May 2023 — Aug 2023

MITACS Research Intern

Clean Energy Group, University of Toronto Scarborough · under Prof. Oleksandr Voznyy

  • +Implemented graph message-passing networks (relational ConvNet with bidirectional updates and residual connections) and CGCNN for band-gap and orbital-energy prediction on zincblende and rocksalt crystal structures.
  • +Designed a custom magpie, orbital-aware, and structural embedding scheme — MSE of 0.03 and 0.23 across the two prediction tasks.
  • +Built an XGBoost band-gap predictor from 2048-dimension Morgan fingerprints, reaching an MSE of 0.30.
PyTorchGraph Neural NetsMultilayer perceptronXGBoostSlurmTensorboardBash Scriptsview source ↗
03

Education

cat education.log

2020 — 2024

B.Tech, Materials Science & Engineering

Indian Institute of Technology (IIT) Kanpur

Four years of materials science along with other courses across Control Systems, Machine learning, NLP and RL. Involved with Aerial Robotics as a Team Head, Secretary of the Programming Club.

  • CPI 8.2 / 10.0

Achievements

Recognition along the way

A few of the ones that stuck: an Analyst of the Quarter award at American Express for optimizing usage of the Email Response model, and two Inter IIT Tech Meet results working on ISRO challenges and winning a Shastra Challenge in Orbital Mechanics track.

  • Analyst of the Quarter, Q3'25 — American Express
  • Bronze Medal — Inter IIT Tech Meet 11.0 (ISRO Chandrayaan Moon Mapping Challenge)
  • 2nd Position — Inter IIT Tech Meet 10.0 (ISRO Web-Based X-Ray Burst Identification)
  • 2nd Position - Orbital Mechanics track (Shastra Challenge - Orbital Mechanics)
04

Projects

ls projects/

Safe-RL exploration

archived
Reinforcement learning

DDPG and TD3 agents that explore safely in continuous-action spaces with physical constraints violation.

PyTorchDDPGTD3GymnasiumPolicy-gradient
0 violations / 10k+ episodesview source ↗

E-commerce review score forecasting using MLFlow

archived
MLOps

A 4-stage MLOps pipeline — ingest, clean, train, evaluate — for forecasting e-commerce review scores.

ZenMLMLflowPython
F-score 0.88view source ↗

indic-NMT

archived
NLP

Neural machine translation between English and seven major Indian languages encompassing LSTM, GRU, and Transformer architectures with various encoding strategies.

PyTorchTransformersFastTextGloVe
BLEU 0.101view source ↗
cat projects/*.md (read the full writeups) →
05

Beyond work

cat ~/.plan

Reading

Reading

Read across fiction and non-fiction, with a keen interest in science, murder mysteries and philosophy

Sport / fitness

Badminton, Lawn Tennis, Table Tennis

Am personally excited about racket sports in general. Have been a state level badminton player in the past, and always in search of exciting matches around!

Photography

Photography

Keen interest in photography, especially street and still life. I possess a Fujifilm X-t50 and have been tinkering with Adobe lightroom for post-processing. I have a small collection of my work on my Instagram page named fuji_visions