Matthew Schoenbauer is a founding engineer at Traversal, building AI systems that investigate production failures across sprawling software infrastructure. His approach combines experience in quantitative trading with a precise distinction between work requiring adaptable judgment and work better handled by conventional software.
Schoenbauer graduated from the University of Notre Dame in 2020 with majors in honors mathematics and philosophy. A Goldwater Scholar and Glynn Award recipient, he applied machine learning to food insecurity and children’s healthcare and wrote a senior thesis on neural-network complexity. His undergraduate research and academic background anticipated a career spanning mathematical theory and practical AI.
He subsequently worked at Proof Trading and Citadel Securities, where production incidents repeatedly interrupted the work of operating high-frequency trading systems. A causal-AI class introduced him to Anish Agarwal and the possibility of automating incident remediation; Schoenbauer became Traversal’s first engineering hire.
In 2024, he also co-authored research on quantization-aware training, establishing conditions under which seemingly different gradient estimators behave equivalently to straight-through estimators.
At Traversal, Schoenbauer helped develop autonomous incident investigation using coordinated agents that search operational telemetry and deliver findings directly to engineers’ incident channels. In a DigitalOcean production case study, he reported an approximately 40 percent reduction in incident-resolution time. The system surfaces supporting telemetry, confidence assessments, explanatory reasoning, and interactive impact maps, giving engineers evidence they can interrogate before acting.
- Cognitive judgment versus procedural execution. Schoenbauer argues that enterprise AI’s central challenge is unpredictability in unfamiliar situations, not whether a model reproduces identical answers. His framework for dividing cognitive and procedural work assigns interpretation and adaptation to models while entrusting repeatable operations to software. He identifies persistence, user experience, and observability as prerequisites for making that division dependable.
- Flexible agent harnesses. Schoenbauer rejects rigid routers, predefined specialists, isolated summarizers, and fixed investigative sequences when they force a system to inherit its designers’ assumptions. His approach to agent architecture instead favors general-purpose subagents, runtime tool discovery, and interfaces that let models adapt investigations as evidence emerges.
- Evidence-backed incident diagnosis at scale. Production failures require locating consequential clues across enormous volumes of logs, metrics, and service activity. Schoenbauer treats parallel search, responsive infrastructure, and inspectable root-cause explanations as equally necessary: an investigation must finish quickly enough to help during an outage and expose enough evidence for engineers to verify its conclusions.
