About me
I’m Risheekkumar Baskaran, a Senior Data Scientist at American Express, based in Chennai, India. My background spans data analytics and machine learning, with a focus on finding useful, actionable signals in complex data.
I build production machine learning systems across fraud, risk, marketing, and recommendations. My recent work includes pre-training a foundation transformer on 60 billion transactions, large-scale retrieval and ranking systems, and GenAI applications built with DSPy and GEPA. I care about models that work beyond a notebook: sound evaluation, efficient training, reliable inference, and measurable outcomes.
I also build open-source tools. fastpragma turns the PRAGMA architecture into a practical workflow for tokenizing event data, packing variable-length sequences, pre-training, and fine-tuning. Outside work, I enjoy playing chess and writing about systems I want to understand from first principles.
Experience
Senior Data Scientist
American Express
- Pre-trained and deployed a custom foundation transformer on 60 billion transactions, powering production use cases across fraud, risk, and marketing with $10M in projected uplift.
- Published fastpragma, a Python package for PRAGMA data processing and pre-training that reduced ablation cycles from one week to one day using Polars-based tokenization, sequence packing, and variable-length attention.
- Built GenAI and recommendation systems using DSPy, GEPA, two-tower retrieval, and XGBoost reranking. Models trained on 2 billion transactions across 8 H100 GPUs achieved 95% F1, 3x Recall@K, and 30% higher merchant adoption.
- Received the Amex Most Innovative Project award in 2024 and 2025.
Data Scientist
American Express
- Built sequence and NLP models for customer behavior, including Transformer4Rec for next-action prediction, driving a 2% CTR uplift.
- Derived behavioral features from petabyte-scale clickstreams and validated them across more than 4 million users, contributing $3M in incremental revenue.
- Deployed a 64-class multi-label intent classifier with asymmetric loss and BERT-family encoders. ONNX quantization reduced inference cost by 75% while keeping p99 latency below 30 ms.
Management Trainee, Data Science
IDFC FIRST Bank
- Reframed website cursor movement as spatial heatmaps for ResNet-based fraud detection, improving AUPRC by 20% over existing baselines.
- Designed an NLP entity-extraction and OCR-verification pipeline for B2B onboarding, reducing document turnaround time from 14 days to 3 days.
ULIP Intern
Unilever
Selected for Hindustan Unilever’s flagship internship program.
Education
Indian Institute of Management, Indore
Master of Business Administration, 2019 - 2021
Madras Institute of Technology Campus
Bachelor of Engineering in Aeronautical and Aerospace Engineering, 2015 - 2019
Led innovation for Flight, a national-level technical festival, and coordinated theatrical performances and industrial visits for the Department of Aerospace.
About this blog
This blog is my working notebook for machine learning systems that I want to understand deeply. I write implementation-driven guides on foundation models, vector search, prompt optimization, recommender systems, cryptography, and the engineering choices that turn an idea into a working system.
The goal is simple: explain what I learned clearly enough that someone else can build on it. The Archive contains all long-form articles, while Projects collects the open-source work behind them.