Experience

• Built a real-time session-based recommendation model using HuggingFace T5 to predict next asset interactions from user event sequences, achieving a 20% offline recall gain versus rules-based baselines
• Deployed a PyTorch ensemble fraud-detection system on AWS SageMaker, improving precision by 8.2% and reducing analyst review time by 20 hours per week
• Led the implementation of an RNN-based fraud detection system for P2P withdrawals using Airflow, Kafka, and Hive that achieved a 5x lift in fraud detection
• Launched an LLM monitoring agent (Dify + GPT-4 API) to summarize dashboards and logs, cutting manual review by 60%
• Partnered with Compliance teams to design SQL-driven fraud investigation workflows and train analysts on their adoption

• Executed comprehensive market analysis for weighted blankets in Europe, developing a data-driven entry strategy targeting an 8% market share in the DTC and e-commerce sectors
• Conducted A/B testing to refine product attachment rates, significantly boosting cross-sell and up-sell strategies
• Analyzed customer purchase behaviors using BigQuery, achieving a 16% boost in targeted marketing effectiveness
• Engineered data pipelines and dashboards to define and track KPIs, driving business growth and improving key metrics by 15%

• Enhanced in-app user experience by designing React components, boosting user engagement by 7%
• Streamlined the GraphQL layer, improving retrieval and update times for PostgreSQL attributes by 30%
• Collaborated with the Product Team on the engineering design doc to implement three new service-driven backend features
• Maintained and updated unit tests and end-to-end tests using Minitest and Cypress to ensure code coverage higher than 85%

PI: Prof. Daniel Aliaga

• Collaborating on uTree, an AI-powered platform that combines remote sensing and high-performance computing to map and monitor urban trees, helping cities plan for a more sustainable and climate-resilient future
• Designing and implementing the backend infrastructure and simulation pipeline

Supervision of Prof. Yasir Zaki

• Engineered and deployed a neural network model using Kotlin on a mobile browser to classify and block JavaScript code, resulting in a 27% improvement in web page load time for underserved countries
• Compiled a comprehensive dataset of 127,000 JavaScript elements labeled with categories by extracting 500,000 elements from 20,000 popular web pages

Supervision of Prof. David Holland and Dr. Daiane Gracieli Faller

• Developed and assessed a Bayesian probabilistic model to predict long-term shoreline changes based on geomorphological and geological conditions under various sea level rise scenarios

Supervision of Prof. Hisham Sati

• Exploring how the dependently typed programming language Agda can be used to formalize braid groups as automorphisms of free groups