Anna Marie Benzon is the founder and chief executive of Stock Knowledge, a Philippine education-technology company, and an AI researcher developing interpretable reinforcement learning for operational decisions. Her projects address problems ranging from access to science education to equipment maintenance and automated responses to data-pipeline failures.
Benzon taught college physics before founding Stock Knowledge in 2016. She built the company around adaptive coursework, assessments, performance tracking, and game-like feedback, with UNICEF Venture Fund support advancing open-source development and offline access. Stock Knowledge subsequently expanded into augmented- and virtual-reality STEM education, including pilots and research with education authorities in Manila and Makati.
Her open-source projects extend that work into reinforcement learning, predictive maintenance, and edge computing. A Jetson-based point-of-sale prototype combines object detection, hand-gesture recognition, optimized inference, and a kiosk interface.
- Bounded ETL remediation. Her rl-etl-remediation-agent connects AWS Glue, EventBridge, Lambda, CloudWatch, and the Glue Data Catalog to investigate schema drift, incompatible data types, and abnormal null rates. Deterministic checks identify failures; inspectable tabular Q-learning selects restricted responses including retries, schema coercion, rollback, quarantine, or escalation.
- Independent safety guardrails. Benzon separates the authority to approve an action from the policy proposing it. Critical or unfamiliar incidents can be escalated, and the system distinguishes an action that appears safe from one that is actually executable. Her AI Engineer World’s Fair presentation identifies shadow-mode testing against representative incident traces as the next step before granting execution authority.
- Transparent evaluation against simpler baselines. In a synthetic ETL benchmark, anomaly detection achieved precision of 1.0, recall of 0.8, and an F1 score of 0.889. Successfully resolved incidents averaged approximately 5.24 minutes, against a modeled manual baseline of 2.5 working days. Crucially, the reinforcement-learning policy matched an equivalent deterministic policy; these controlled results establish feasibility, not superior learning performance or production readiness.
- Interpretable predictive maintenance. Her Dyna-Q maintenance benchmark compares Dyna-Q, Q-learning, SARSA, and rule-based decisions using NASA’s CMAPSS FD001 turbofan dataset. Additional planning accelerated convergence without improving the final learned policy; the project remains an educational and research benchmark, not an aircraft-maintenance system.
