➤ Leading the reinforcement learning environment engineering team

➤ Led data quality tooling, building from scratch data quality tools that are used to monitor and improve quality for every data point the company delivers as it grows to >$1.2B in revenue and beyond

➤ Built a system to match new workers on our platform with work, helping workers become active 5-10x faster

➤ Trained models to understand the impact of data quality on model performance

➤ Conducted research that doubled a language model’s robustness to attacks through adversarial training, resulting in a NeurIPS 2022 publication.

➤ Fine tuned large language models based on deBERTa with custom EC2 experimentation platform and built language model-assisted adversarial attack tools with React, Flask, tailwind.css, dvc, Lambda Labs, and HuggingFace.

➤ Built a PyTorch-based framework for rapidly prototyping adversarial attacks and training for transformer language models and used it to discover “relaxed” adversarial attacks that made toy models robust to all known adversaries without degrading performance.

➤ Used cutting-edge interpretability tools to find new circuits in GPT-2 that extrapolate patterns and explained more than 90% of that behavior.

➤ Mentored two teams of researchers who studied a language model and identified new compositions of attention heads, akin to induction heads.

➤ Led a team of 5 to build data infrastructure with Celery and Redshift to submit millions of insurance claims worth hundreds of millions of dollars.

➤ Developed and maintained laboratory operations software in Django, TypeScript, React, GraphQL, and Postgres.

➤ Increased COVID-19 test plating throughput by 40% by building and deploying a plating tool in React, Django, GraphQL, and SQLAlchemy on AWS.

➤ Led research, planning, and rollout across product, engineering, and lab operations to deploy the plating tool across all labs in weeks.

➤ Led 6 engineers to develop a tool in Typescript to automate optimization of 80% of Flexport’s ocean shipment routing, saving core operations team members 4-8 hours a day and decreasing response times to customer-facing teams from 12-24 hours to 2-3 hours.

➤ Migrated to a Flask service to enable data scientists to use Coin-OR and pyomo to optimize pricing. Built data pipelines using dbt and Snowflake.

➤ Scaled the Ocean team’s processes to 12 people by mentoring senior and junior engineers and junior product managers and instituting postmortems, quarterly planning, software design reviews, improved oncall process, and regular metric reviews.

➤ With partner teams, planned a two-year migration to a service-oriented architecture (SOA) in Flow, Ruby on Rails, Java, PostgreSQL.