Data Science student at UC Berkeley. I work in machine learning, software engineering, and data infrastructure. From San Francisco.
alextoohey.dev · LinkedIn · alex_toohey@berkeley.edu
Data Science Intern at Visa in Foster City, building an LLM-based merchant recommendation model fine-tuned on dataset pulled, processed, and transformed from 76+ billion payment transactions.
| Visa · Data Science Intern | Recommending merchants to cardholders is a language problem disguised as a payments problem. I extended Llama with 10,000+ new merchant tokens, domain-specific embedding initialization, and LoRA fine-tuning. 84% Recall@10, 88% NDCG@10, 210+ ablation studies, 6x above the strongest heuristic baseline on unseen merchants. Getting it there required a Spark ETL pipeline on Kubernetes that turns raw transactions into training sequences. |
| Databricks · Software Engineer (Contract) | Databricks' Partner Demo Catalog ships 169 enterprise AI/ML partner solutions, and every one used to be verified by hand. I built the testing infrastructure that automates it: 350+ Pytest and Playwright tests covering 12+ resource types, with results streaming to a Delta table after every run. Verification time dropped 94% and the system caught 36+ production issues before partners ever saw them. |
| BART · Software Engineer Intern | BART runs on production codebases that outlive their documentation. I built a fully local code intelligence platform that generates architecture docs and diagrams for 20,000+ line codebases in under two minutes, using Tree-Sitter/AST parsing across 8+ languages and a multi-stage validation pipeline that keeps the LLM grounded. 97.5% summary accuracy. Nothing leaves BART's own machines. |
| SubscriptionFlow · AI/ML Engineer Intern | A summer in London at a YC (W22) startup. Support and success teams needed answers that lived in SQL they couldn't write, so I built a RAG-powered Text-to-SQL model with VannaAI, GPT-4o, and ChromaDB, trained on 500+ Q&A pairs, 140+ schemas, and 55+ docs. 75% fewer hallucinations, 92% faster retrieval. Also designed a 4-class churn prediction pipeline on AWS SageMaker so retention work could be targeted instead of guessed. |
AskOski · [private repo] · AskOski helps 45,000+ Berkeley students plan their academic paths. I built the ML and data pipeline behind it: Apache Airflow workflows that retrain models automatically from live student-information-system data, BERT-based course modeling, and sub-250ms response times.
EquiPath · repo · Most college search tools rank schools by prestige. For a first-generation student or a student-parent, that's the wrong question. EquiPath matches students to schools based on the constraints they actually face: financial aid, transfer rates, campus childcare, first-gen support. Built with scikit-learn and the Claude API. Won first place in the Educational Equity Track at Berkeley's Datathon for Social Good.
ProbatePilot · repo · When someone dies, the executor is usually a grieving family member, not a lawyer, and they're personally liable if a deadline slips or creditors get paid out of order. ProbatePilot reads uploaded estate documents into a live estate record, then a deterministic California probate rule engine paired with a Claude tool-use agent flags the next deadline or liability risk before it becomes expensive. Built with the Claude API, FastAPI, and Next.js, deployed on Google Cloud Run and Vercel with path-scoped auto-deploy CI/CD. Built with a 4-person team in 24 hours at UC Berkeley AI Hackathon 2026. I led document intelligence during the hackathon, then refactored and deployed the system solo afterward.
- Aurora anomaly detection · Space Sciences Laboratory @ UC Berkeley (Fall 2025) · Computer vision models for identifying rare auroral structures in nightly sky data.
- Sheriff elections & incarceration · Stanford Medicine, Epidemiology & Population Health (Spring 2025) · Co-engineered the first national county sheriff election dataset, then built inferential models linking sheriff demographics to local incarceration rates.
- Diffusion steering via RL · Data Science Society @ Berkeley (Spring 2026) · Using reinforcement learning to steer diffusion models toward legible text in generated images.
- The Art of the Underdog · Sports Analytics Group at Berkeley · What March Madness Cinderellas have in common, and which traits actually predict upset runs.
- The Battle of the Surfaces · Sports Analytics Group at Berkeley · A lifelong tennis player asks how much of the clay/grass/hard-court folklore holds up in the data.
- Data C8 (Foundations of Data Science) · Tutor on content team for Berkeley's intro data science class with 1500+ students
- Data Science Society at Berkeley · Project Manager, Technical Consultant, and Teaching Assistant
- Sports Analytics Group at Berkeley · Data Journalist, Editor, and Advisor
- Code for Good · Software Developer and Recruitment Chair