Staff Scientist · New York
I'm Armando Ordorica, a staff scientist and PhD researcher with 10+ years building production ML systems at the intersection of reinforcement learning, causal inference, ranking, and recommendation. I lead ranking and recommender systems at Pinterest.
The best recommendation isn't always the next click. Sometimes it's the reason someone comes back next week.
My work focuses on moving machine learning beyond short-term proxies by designing systems that explore responsibly, understand delayed outcomes, and create durable value for people and platforms.
Python · SQL · PyTorch · TensorFlow · AWS · NumPy/SciPy · scikit-learn
English · French · SpanishNative fluency
Research translated into products, policy, and measurable impact.
Leading ranking across HomeFeed, Search, and Related Pins. Pioneered offline replay and contextual-bandit feedback loops; long-term reward work drove +1.5M weekly active users across 8+ experiments.
CurrentJumio
Led aggregate fraud scoring for 10,000+ clients across 40+ countries, unifying contextual, image, clustering, and NLP models while reducing human-labeling costs by roughly 40%.
University of Toronto
Taught approximately 150 students across machine learning in finance, deep learning, databases, cloud computing, and blockchain.
Flexiti Financial
Replaced rule-based fraud decisions with adaptive ML systems, improving F1-score by more than 400% and uncovering fraud through anomaly detection and NLP.
Capital One
Led high-risk authorization strategy, generating $7M annual NIBT; designed policies across $37B in credit exposure and uncovered $20M/month in potential fraud recovery.
Montreal Neurological Institute
Developed a Python framework to simulate electrical brain activity from histological data in support of research into epilepsy and Alzheimer’s disease.
Publications and inventions spanning long-term engagement, responsible ranking, and risk.
A framework for finding and evaluating downstream reward signals that improve long-term outcomes in large-scale recommendation.
A scalable recommender-system framework connecting meaningful user actions to retention.
Methods for measuring borderline content and its effects across online ecosystems.
A review of sequential decision-making methods for optimizing advertising policy within large-scale recommendation ecosystems.
4 United States patent applications listed on Google Scholar.
Recommendation methods that use subscriber and content embeddings to optimize toward durable, long-term outcomes rather than immediate interaction alone.
A system for using large language models to generate semantic representations and identify image-based content at scale.
Machine learning models that combine device, network, geographic, and identity signals into an aggregate score for more robust verification decisions.
Context-aware risk scoring that helps authentication systems select verification requirements and optimize decision points.
Recent public experiments, teaching material, and code from outside the day job.
Private research build. The project name and activity are public; implementation details remain confidential.
Private local-AI systems project. Source code and technical details are not publicly accessible.
Private computer-vision application. Source code and technical details are not publicly accessible.
A living, interactive CV for Armando Ordorica
working version of the paper
Function Approximation and SARSA control
Open to thoughtful conversations