Sandlines
In-app polling data pipelines delivering voter insights up to 14x faster than traditional methods, with modular Python transforms and time-series visualizations for campaigns.
Alex Toohey · Data Science @ UC Berkeley
The trail starts here.
I enjoy building software systems that solve interesting problems using AI, machine learning, and data. Inspired by my love of hiking and exploring National Parks, I built this interactive park as a way to share the projects I've worked on, the experiences that have shaped me, and a little about who I am. Follow the trail at your own pace, or open the map to jump anywhere.
Walk by scrolling, clicking ahead on the trail, or with ← → · A/D · space
Or skip the walk — résumé (PDF).
02 · Visitor Center — About
Orientation
I'm Alex, a UC Berkeley student studying Data Science with a minor in Computer Science. Most of my experience has been building software, AI, and machine learning systems through internships, research, and consulting projects, while also helping teach multiple data science courses at Berkeley.
Outside of school and work, I enjoy hiking, traveling, skiing, trying new restaurants, and playing intramural sports like flag football, soccer, and dodgeball. One of my long-term goals is to visit all 63 U.S. National Parks, and I've visited 15 so far, which inspired this site.
The giants past this porch are borrowed from Sequoia. Mind the creek.
03 · The Overlook — Foundations
Reading the layers
Every trail stands on older rock. The strata in this canyon are the earlier stints the rest of the climb was built on:
PlayVision
Basketball computer vision at a YC (F25) startup: film labeling for model fine-tuning, analytics outputs for clients, and market analysis for the business.
Layered deep time, the way the Grand Canyon tells it.
04 · The Trail — Experience
The climb
4 stretches of trail, each one a little higher than the last. The waysides along the climb tell each story — walk on, or take the map to skip ahead.
This stretch takes its warm rock from Zion and Bryce Canyon.
Milepost 1 of 4 · June 2025 – August 2025
SubscriptionFlow
AI/ML Engineer Intern · London, UK
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, embedding 500+ SQL Q&A pairs, 140+ schemas, and 55+ docs so the model answered from the company’s real data instead of hallucinating (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.
Milepost 2 of 4 · January 2026 – May 2026
BART
Software Engineer Intern · Berkeley, CA
Bay Area Rapid Transit runs on production codebases that outlive their documentation. I built a secure, fully local code intelligence platform that generates architecture documentation and diagrams for 20,000+ line codebases in under two minutes. It combines static analysis, code graphs, and grounded LLMs. A modular Tree-Sitter/AST parsing framework supports 8+ languages, and a multi-stage validation pipeline keeps the LLM honest: 97.5% summary accuracy, with everything staying on BART’s own machines.
Milepost 3 of 4 · March 2026 – June 2026
Databricks
Software Engineer (Contract) · Berkeley, CA
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. It runs Pytest, Playwright, and the Databricks SDK in a modular two-layer architecture: browser automation for user flows, extensible backend validation for 12+ resource types, 350+ tests in all. Results stream to a Delta table after every run, which turned verification into CI/CD and caught 36+ production issues before partners ever saw them. Verification time dropped 94%.
Milepost 4 of 4 · May 2026 – Present
Visa
Data Science Intern · Foster City, CA
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 to build a merchant recommendation model reaching 84% Recall@10 and 88% NDCG@10. Feeding it meant engineering a distributed Spark ETL pipeline on Kubernetes that turns 76+ billion raw transactions into natural-language training data, then 210+ ablation studies to find what actually mattered, ending 6x above the strongest heuristic baseline on unseen merchants.
05 · The Workshop — Projects
Sawdust and experiments
The bench inside is where things get built for the joy of it. What's on the shelves so far:
AskOski ML & Data Pipeline
ML and data infrastructure for UC Berkeley's personalized academic planning platform, serving 45,000+ students. Read the build notes →
EquiPath
Equity-centered college matching for first-generation, low-income, and student-parent applicants. Won first place in the Educational Equity Track at Berkeley's Datathon for Social Good. Read the build notes →
The desert flat outside is Joshua Tree's — quiet, wide, good light for building.
06 · The Campfire — Teaching & Leadership
Stories around the fire
Where the knowledge gets passed on. A few told most often:
The steaming spring beyond the fire is a nod to Yellowstone.
07 · The Observatory — Research & Writing
After dark
On the last shoulder before the top, the sky takes over. This is where the questions live — research and writing done for the curiosity of it:
Aurora anomaly detection
Supervised and unsupervised computer vision models that sift rare auroral structures out of thousands of nights of ordinary sky.
County sheriff elections & incarceration
Helped engineer the first national county sheriff election dataset, then built inferential models revealing associations between sheriff demographics and incarceration rates.
Diffusion steering via reinforcement learning
Steering diffusion models with reinforcement learning to improve text-in-image generation.
The Battle of the Surfaces
Which playstyle wins on clay, grass, and hard courts? Data journalism from a lifelong tennis player.
The Art of the Underdog
Analyzing what March Madness Cinderellas have in common, and which traits actually predict upset runs.
The summit is just ahead — one last push.
Haleakalā keeps telescopes on its high shoulder. So does this park — and the aurora overhead belongs to the research below it.
08 · The Summit — Résumé & Contact
The view from the top
The whole trail behind you, stars out overhead. If you'd like the formal version of everything you just walked past:
View résumé (PDF)Get in touch
- Email[email protected]
- LinkedInin/alextoohey1
- GitHubalextoohey
The trail keeps going — new stretches are cut every season.
A summit under full stars — Haleakalā is famous for exactly this.