Aequitas
Bias auditing and AI fairness toolkit used by practitioners, researchers, and policymakers worldwide. One of the first open-source tools for auditing algorithmic fairness, created during Pedro's postdoc at the University of Chicago.
Pedro Saleiro
Co-founder & Chief AI Officer at Opnova
Building trustworthy AI agents for regulated industries.
Short Bio
Pedro Saleiro is a Portuguese AI researcher and entrepreneur. He is Co-Founder and Chief AI Officer at Opnova, a startup building AI computer-use agents for enterprise operations in regulated industries. He has created several open-source tools, most notably the Aequitas bias auditing toolkit, one of the most widely adopted open-source tools for AI fairness. Before Opnova, Pedro led AI Research at Feedzai, where his teams filed numerous patents, published at top AI venues, and shipped multiple research innovations to product. His career includes a postdoctoral position at the University of Chicago, a Ph.D. internship at Microsoft Research, and a Ph.D. in Artificial Intelligence from the Faculty of Engineering at the University of Porto.
For speaking inquiries, reach out via LinkedIn.
Selected Talks
- Can AI promote gender equality? — TEDxPorto
- Fairness in Machine Learning — Lisbon Machine Learning School
- Fairness in Decision-Making with AI: A Practical Guide & Hands-On Tutorial Using Aequitas — PyData
- Aequitas — DSSG Data Fest, University of Chicago
Open-Source Projects
TimeSHAP
KDD 2021 — Model-agnostic explainability method for recurrent neural networks. Extends KernelSHAP to the sequential domain for explaining predictions of time-series models.
FairGBM
ICLR 2023 — Gradient boosting with fairness constraints. A dual-ascent learning framework for training gradient-boosted decision trees under fairness constraints with minimal impact on predictive performance.
Bank Account Fraud Dataset
NeurIPS 2022 — First large-scale, privacy-preserving, realistic tabular dataset suite for fraud detection research. Designed to provide the research community with a robust benchmark for evaluating novel methods.
Selected Publications
- Aequitas: A Bias and Fairness Audit Toolkit (2018)
- TimeSHAP: Explaining Recurrent Models through Sequence Perturbations, KDD 2021
- How Can I Choose an Explainer? An Application-Grounded Evaluation of Post-Hoc Explanations, FAccT 2021
- Turning the Tables: Biased, Imbalanced, Dynamic Tabular Datasets for ML Evaluation, NeurIPS 2022
- Laundrograph: Self-Supervised Graph Representation Learning for Anti-Money Laundering, ICAIF 2022
- Human-AI Collaboration in Decision-Making: Beyond Learning to Defer, ICML 2022
- FairGBM: Gradient Boosting with Fairness Constraints, ICLR 2023
- On the Importance of Application-Grounded Experimental Design for Evaluating Explainable ML Methods, AAAI 2024
In the Media
- Pedro Saleiro criou um software para combater desigualdades — Público
- O português que combate a discriminação social ao lado do ex-analista de Barack Obama — Observador
- Algoritmos à prova de discriminação — Jornal de Negócios
- World Changing Ideas Awards — Fast Company
- Opnova Emerges from Stealth with $3.75 Million in Funding — BusinessWire
- Interview with Pedro Saleiro — Critical Future Tech
Writing
Opnova Blog — “What is Agentic AI?” Series
- What is Agentic AI? Part 1: Introduction — An introduction to intelligent agents and how agentic AI is transforming enterprise automation.
- What is Agentic AI? Part 2: The Role of LLMs — Demystifying LLMs, their capabilities, and why they are the “brains” driving agentic AI.
- RPA vs. Agentic AI: How to Make Sense of It? — The differences between RPA and agentic AI and why it matters for modern business automation.
Other Posts
- Building Trustworthy Agentic AI for Enterprise Automation — Opnova's vision for reliable, enterprise-grade AI automation in regulated industries.
- The Fallacies of Agentic AI — The pitfalls of legacy RPA and the potential of true agentic AI for complex enterprise processes.
- The Fallacy of the Off-the-Shelf AI Agent — Why off-the-shelf AI agents fail and why customization is essential for enterprise integration.
- Reflexive Memory: When AI Agents Remember How to Work — How reflexive memory enables AI agents to remember actions and adapt reliably.
- Joiner for AI Agents: The Workforce Nobody Hired — Why the Joiner workflow breaks for non-human identities and how to fix it.
- Introducing the OPN Model Family — Computer-use models trained on synthetic IAM trajectories from Opnova Range.
Patents
- Teaching Computer-Use Artificial Intelligence (AI) Agents from Video Demonstration
- Multimodal Reflexive Memory Grounding for Computer Use Agents
- Minimize Rework
- Weakly supervised multi-task learning for concept-based explainability
- Hierarchical machine learning model for performing a decision task and an explanation task
- Human-in-the-loop evaluation for explainable artificial intelligence
- A model-agnostic approach to interpreting sequence predictions
- Surrogate hierarchical machine-learning model to provide concept explanations
- Bandit-based techniques for fairness-aware hyperparameter optimization
- Obtaining a generated dataset with a predetermined bias for evaluating algorithmic fairness
- Method and system for alert review using machine learning
- Alert review using machine learning and interactive visualizations
- Method and system for obtaining a surrogate hierarchical machine-learning model
- Weak supervision framework for learning to label concept explanations
- Constrained optimization for gradient boosting machines (US)
- Constrained optimization for gradient boosting machines (EP)
- Self-supervised framework for graph representation learning
- Method and device for obtaining a generated dataset with a predetermined bias
- Method and system for generating a concept label model to label a dataset
- Method and system for fairness-aware data valuation processing for supervised learning
- Machine-learning method for automatic instance assignment under fairness