• Designed and executed product experiments in collaboration with external behavioral economists and social policy experts• Wrote backend, front-end, and analytical code to shepherd a hypothesis from ideas to concrete, causal results

• Made experimentation a first-class citizen of our data models, and built tools to make product experimentation and safe rollouts of features as frictionless as possible• Migrated old batch-daily ETL infrastructure to a new, streaming ETL by writing infrastructure-as-code that reproducibly configured AWS Database Migration Service instances, Lambdas, Redshift clusters, Glue data catalogs, Relational Database Service instances, and Cloudwatch telemetry upon deploy• Augmented the data analysis ecosystem by introducing and operationalizing the Data Build Tool to allow analysts and engineers an intuitive way to create denormalized views and materialized tables in our data warehouse

• Built in-house machine learning model versioning system to sync source code with binaries in production• Created a new model and training system for merchant classification, based on transaction name strings from banking data. This fed downstream transaction enrichment, to identify payday loans or particular target merchants• Built statistical models to analyze member support tickets and suggest stereotyped responses, for use in our internal chat support service• Identified key design decisions that cut time to support ticket closure to a fraction of what they were, with exploratory data analysis

• Built a framework to instrument physical & behavioral simulations (in Java), automatically generate initial configurations and experimental conditions (in Scala), and store and organize collected data in a PostgreSQL database. In addition, served up modeling results to a UI using a Scalatra server, and visualized summary statistics from the SQL tables with R.• Built a memory network to create context-based entity embeddings in a new and unfamiliar language (Lua + Torch), and created an interactive UI to showcase the model's question-answering abilities within a three week deadline (with Javascript).• Boosted performance of a large, massively parallelized embedding architecture (in Scala + Java). This involved work with concurrency, over-the-wire communication between compute nodes, ad-hoc improvements to the TransE algorithm for embedding triples, as well as integration of gradients posted from other embedding algorithms such as Skipgram (for embedding free text words) and a vanilla neural network (for embedding numeric literals).

• Programmed psychophysical experiments, employing Matlab toolboxes such as Snow-Dots and Tower of Psych• Assembled full experimental pipelines, from physiological signal inputs, stimulus output, and robust synchronization of all data streams along a single timeline• Built associated analytical tools for data from such experiments, including feature extractors, multivariate regressions, etc.