Worked on large language modeling and general AI applications at Normal Computing, a deep-tech AI startup focused on production-ready, reliable, and generative AI workflows.

Led the research, development, and initial deployment of multiple, distinct viewership/impressions models. All models included fully Bayesian high-dimensional shrinkage, so that estimation could be successfully applied to a much wider array of datasets without direct intervention. Our work involved data and feature sets of non-trivial sizes (e.g. on the order of 10k observations and 1k features on average).

• Managed a small team of data scientists, supported a 24-hour real-time computer vision service based on in-house hardware and simple inference-based systems for internal monitoring, automated client data retrieval, etc.• Performed numerous analyses and optimizations using detailed city-wide data in order to determine billing rates, pick-up schedules, subtle user/product interactions, fraud, new locations for services that balance constituent travel costs and service demand.• Built and managed our AWS-backed Kubernetes data science infrastructure, which included Airflow and Dask clusters, an S3 data lake and AWS Aurora warehouse backed by replicated Postgres production databases and external Zendesk and Jira feeds.

I taught and helped design the curriculum for an introductory data science class.

Energy usage, efficiency and savings modeling for project evaluation and prediction.

Custom Bayesian retail demand modeling and estimation of price elasticity for high-dimensional sparse data in Stan. Models were designed to address full estimation (e.g. hyper-parameters) with modern, robust parameter shrinkage via Horseshoe priors.

Designed and developed models for the prediction of demand and sales of used automobiles from wholesale auction and listings data. The models addressed various truncated and missing data scenarios, large feature sets, and possible causal factors.• Development for–and integration with–Mongo, AWS hosted NodeJS services. • Set up and maintained PostgreSQL RDS and EMR services–hosting Hue, Hive and Spark.• Wrote custom R-driven, standalone JSON RPC (via Rook) and Shiny services for predictions and interactive graphics.

In-house, full-time consultant to faculty on matters of research involving statistics, optimization, experimental design, machine learning, as well as data manipulation and management of large datasets. • Managed the setup, maintenance and use of Cloudera Hadoop and Spark clusters for terabyte-sized Nielsen datasets. • Hadoop deployments on virtualized hardware (e.g. Isilon) using the VMware stack and Serengeti. • Extensive Hive, traditional SQL query development and tuning, SAS, Stata and Matlab scripting and cluster integration. • Aided maintenance and use of in-house high performance grid (Sun Grid Engine) computations in OpenMPI.

Inaugural mentor for DSSG on the CTA-sim and Energywise projects.

Developed the location tracking and operational-state prediction model behind MTA Bus Time, as well as traffic and road-network inference for the Cebu Traffic project.

Worked on sequential Bayesian models of market micro-structure for high-frequency market making.

Designed, developed, and operated high-frequency trading algorithms.

Responsible for developing, implementing and optimizing new code for trading, risk management, and analytics.

Worked on a hybrid geometric Brownian motion switching model modulated by afinite-state Markov chain for option pricing.

Wrote a critical analysis of particle physics data using a novel Goodness ofFit theory. Used Monte Carlo simulations to demonstrate efficacy of the theory,which showed a significant increase in bias rejection for certain cases.