Fardin Abdi — Principal Scientist, Amazon AGI Labs

Fardin Abdi

Principal Scientist, Amazon AGI Labs · Seattle, WA

I work on getting frontier models to learn proprietary enterprise knowledge while minimizing regression on what they already know — after a decade building production ML systems at Stripe, Pinterest, and Uber.

About

I am a Principal Scientist at Amazon AGI Labs with over a decade of experience building production AI/ML systems and leading the teams behind them.

I mentor engineering teams and turn applied ML research into shipped systems, with 110 granted US patents and peer-reviewed publications (740+ citations, h-index 16) along the way.

Ph.D., Computer Science University of Illinois Urbana-Champaign · Safety and security of cyber-physical systems

B.Sc., Electrical Engineering University of Tehran · Outstanding Student Award (top 5%)

Qualcomm Innovation Fellowship Winner, 2013 ($100K, 1 of 8 teams from 138 proposals) · Finalist, 2016

Current Focus

Most of the proprietary enterprise data is absent from frontier models’ pretraining corpora. It comes in two shapes: large, highly structured relational sources such as tables and catalogs, and small, niche bodies of internal documentation such as contracts and policies. My current focus is on injecting both types of knowledge into language models while minimizing regression in the skills they already have.

If you want to talk, reach out — the fastest way is email.

Experience

2025 — now

Amazon AGI Labs

Principal Scientist

Working on model customization for frontier models, helping customers adapt foundation models to their domains.
Continued pre-training (CPT) with relational datasets; continual learning without catastrophic forgetting.

2022 — 2025

Stripe

Staff ML Engineer, Risk · Merchant Credit Risk Tech Lead

Reduced high-risk merchant losses while minimizing false positives for legitimate businesses.
Architected DetectiveGPT, Stripe’s LLM agent framework for automated merchant risk investigations, adopted across Sales, Compliance, Crypto, and Strategy teams.
Led ML models for delinquency, loss estimation, user value, and churn — $15M+/year in savings.

2020 — 2022

ML Engineer, Measurement Modeling · Conversion Team Tech Lead

Tech lead for conversion modeling: built evaluation frameworks and operational standards for the team.
Built embedding-based candidate retrieval to keep attribution working as browser privacy changes reduced tracking.

2018 — 2020

Uber AI

ML Engineer, Michelangelo Deep Learning

Founding member of Uber’s ML Application Framework — code/config-driven ML workflows across Keras and PyTorch.
Contributed to Horovod, including its horovodrun CLI and Spark Estimator integration.

2017 — 2018

Capital One

ML Engineer

Built deep learning-powered data profiling and synthetic data generation libraries (GANs/RNNs) — work that produced 110 granted US patents.
Developed Capital One’s first-generation deep learning experimentation infrastructure.

Earlier: internships at Affirm (early engineering hire) and Apple.