Ben Glickenhaus

Leading ML initiatives from ideation to deployment, focusing on personalization and scalable optimization systems that impact millions of users.

• Built Patreon's Learning-to-Rank system powering personalized content discovery, significantly improving user engagement metrics.

• Led development leveraging email as a discovery channel for Patreon creators - scaling to 10s of millions of emails weekly

• Developed and deployed a large-scale contextual multi-armed bandit framework integrated into Patreon's experimentation platform, enabling automated optimization across multiple product areas.

• Used contextual bandits to personalize email copy. Integreated with Patreon experimentation framework and wrapped in extensible API, allowing product teams to quickly spin up new bandits

• Optimized email marketing campaigns for Patreon's creators by building a personalized value model - Resulted in a significant reduction in total email volume and an increase in clickthrough and engagement rates

• Worked with email team to build promotional digest emails using the value model

• Optimized audio transcription ML pipelines, achieving ~90% reduction in operational costs.

Led frontend development of the Digital Commerce/One-time payment functionality for Patreon creators. Worked closely with product, leadership, and engineering to conceptualize and execute on 0 to 1 product development.

• Primarily led frontend product development in React, leading a team of 3-4 engineers to build high craft product UI/UX. Designed and implemented complex state management systems and reusable hooks for commerce product with features like media editing and discount creation.

• Worked in frontend, backend, and data repos to build end to end product features

• Contributed to overall product craft by adding reusable loading component API to component library - used over 100 times across the Patreon frontend

• Helped lead engineering effort to build Digital Commerce product 0 to 1. Worked closely with engineering, product, and design to scope and delegate work, provide timelines and updates to leadership, and build a durable engineering solution.

• Built semantic search for commerce products using multi-modal embeddings, leading to significant increase in checkout lands. Ran series of A/B experiments to measure incremental revenue from the feature

• Engineered embedding-driven recommendation systems, significantly boosting creator discoverability.

• Led engineering effort to build discount functionality for Digital Commerce product

Participate in numerous software engineering and research projects at the UMass Rescue Lab, including mobile app development, cloud infrastructure (AWS and Azure), machine learning, backend/API development, and dev ops infrastructure. Formulate model card for ML models presented to FBI ethics board. Support security analysis of blockchain protocols through utilization of reinforcement learning. Teach CS596E (Machine Learning Applied to Child Rescue) course to enhance leadership and project management skills.

• Implemented Test-Time Training methods on MAE-VIT architectures, improving low-sample class accuracy by 400%.

• Designed cloud-based async worker framework for investigator data triage tool, running in K8s with Argo

• Applied Reinforcement Learning for security analysis of blockchain systems, identifying state-of-the-art vulnerabilities.

• Created automated ML model card generation tools for FBI ethics compliance.

• Supported security analysis of blockchain protocols through utilization of reinforcement learning

• Created media review tool for investigators by using CLIP-powered semantic search engine and integrated into full stack Desktop application AKS

• Taught CS596E (Machine Learning Applied to Child Rescue) course to enhance leadership and project management skills

Developed new techniques to explain black box AI systems. Created evaluation framework for efficacy of explanations. Presented findings to PIs and DARPA program manager at DARPA progress meetings. Directed panel discussion at progress meetings about explanation for RL.

• Coordinated with psychologists team to comprehend explanation and evaluation process

• Utilized approaches, including causal modeling, gradient based, bayesian teaching, and learned explanations that were applied to both data analytics (predictive) and autonomy (RL) responsibilities