CTO at Agents of Chaos
I am building Agents of Chaos, Inc. as CTO, focused on AI evaluation, red teaming, and realistic agentic workflow testing.
Machine learning researcher-engineer
Generative AI, interpretability, and adaptive systems.
CTO at Agents of Chaos, EPFL MSc Data Science, OpenAI contractor, ETH research work, NeurIPS/TMLR publications, and Oracle Labs ML systems experience.
I like work where careful modeling, experiments, and engineering all matter: building AI evaluation and red-teaming systems, understanding model behavior, and turning research ideas into working artifacts.
Generative AIMechanistic InterpretabilityProbabilistic ModelsReinforcement LearningAI EvaluationML Systems
Building now
Agents of Chaos
I am CTO at Agents of Chaos, Inc., building evaluation and red-teaming systems for realistic agentic workflows.
Background
My background
I move between academic ML research and practical systems work: understanding how models behave, then building tools and experiments that make that understanding useful.
Contractor at OpenAI
Contract work connected to frontier AI systems, adding production-facing model experience to my research background.
EPFL MSc Data Science
My MSc at EPFL gave me a broad ML foundation: NLP, diffusion models, RL, probability, and data visualization. GPA 5.83/6; Teaching Assistant for Modern NLP.
ETH Research Assistant
At ETH, I work on adaptive learning systems, especially test-time meta reinforcement learning for policies that adapt during deployment.
NeurIPS and TMLR publications
My publications span diffusion interpretability, sparse autoencoders, probabilistic circuits, and tensor factorization theory.
Oracle Labs Intern
At Oracle Labs, I worked on ML systems for cloud-security defense inside Spark-scale pipelines processing billions of events.
Questions
Research
I am interested in machine learning problems where empirical behavior, theory, and engineering constraints all matter.
How can we understand and control what neural networks compute?
How can systems learn to learn?
How can reinforcement learning become more sample-efficient and reliable?
Conference
One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models
Applies sparse autoencoders to analyze text-to-image diffusion models, contributing to interpretability methods for modern generative systems.
Co-author Diffusion ModelsSparse AutoencodersInterpretability
Journal
What is the Relationship between Tensor Factorizations and Circuits (and How Can We Exploit it)?
Establishes a rigorous connection between tensor factorizations and probabilistic circuits, unifying model families and exposing new architecture-search opportunities.
Co-first author Probabilistic CircuitsTensor FactorizationsTractable Inference
Workshop
Unifying and Understanding Overparameterized Circuit Representations via Low-Rank Tensor Decompositions
Unifies overparameterized probabilistic circuit architectures and studies low-rank decompositions as a way to understand and compress expressive layers.
First author Probabilistic CircuitsLow-Rank ModelsTensor Decompositions
Liminal Training: Characterizing and Mitigating Subliminal Learning in Large Language Models
Supervised student research on characterizing and mitigating subliminal learning behavior in large language models.
Supervisor LLMsResponsible AISupervision
Diffusion Transformers use Sink Registers
Supervised student research investigating sink-register behavior in diffusion transformers.
Supervisor Diffusion TransformersXAIMechanistic Interpretability
Experience
Selected roles
Agents of Chaos, Inc.
Chief Technology Officer
Building AI evaluation and red-teaming systems for realistic agentic workflows.
- Leading technical work on evaluation methodology, adversarial assessment, and evidence-rich AI audits.
- Developing synthetic multi-agent ranges for realistic workflow testing.
- Focused on tool use, permissions, durable state, and failure modes in deployed AI systems.
Oracle Labs
Research Assistant
Developed machine learning systems for cloud security defense in a large-scale Spark pipeline.
- Enhanced a pipeline processing billions of events daily.
- Implemented anomaly detection models using PyOD and custom approaches.
- Engineered data-loading mechanisms for large-scale model training.
ETH Zurich
MSc Thesis
Investigating test-time meta reinforcement learning for policies that adapt during deployment.
- Designing a meta-learning framework for rapid policy adaptation.
- Studying how to reduce gradient steps needed at test time.
- Bridges reinforcement learning, meta-learning, and adaptive systems.
Algoverse LLC
Research Mentor
Mentoring students on mechanistic interpretability research for LLMs and diffusion transformers.
- Mentored 12 students on frontier model-interpretability projects.
- Supervised workshop publications on LLM and diffusion-transformer behavior.
- Helped students turn early research ideas into concrete experiments and submissions.
EPFL
MSc Data Science
Graduate training in machine learning, NLP, diffusion models, reinforcement learning, and probability.
- GPA: 5.83/6.
- Teaching Assistant for CS-552 Modern Natural Language Processing.
- Coursework includes ML, NLP, RL, visual intelligence, probability, and data visualization.
Writing
Technical notes
Wrapped up my final semester of courses and launched this website as a research-engineering profile hub.
Our work on tensor factorizations and probabilistic circuits was accepted to TMLR and received Featured Certification.
Contact
Open to research and ML engineering conversations.
Best way to reach me: antonio.mari02 [at] outlook [dot] com