EECS · UC Berkeley · Class of 2028
Seth Raphael
About
Who I am.
I started teaching myself to code six years ago out of boredom during class, and when COVID pushed everything online that hobby became a full-time fixation: first one Python course, then another, and another after that.
I built LinkJoin from scratch: 2,000+ registered users, 75+ countries, a US trademark, a patent pending, and a message from Zoom's CEO calling it "a breath of fresh air." I spoke at the EDUCAUSE Annual Conference as the youngest presenter in the event's history.
Now I'm studying Electrical Engineering and Computer Science at UC Berkeley, focused on machine learning, distributed systems, and AI interpretability. I'm a research assistant in Berkeley's Adesnik Lab, decoding neural responses in the mouse visual cortex, after earlier ML research at UCSF on songbird language acquisition. I'm also interning at Wander, building infrastructure for autonomous AI deployment, and serve as CTO of the Concord Table Tennis Club.
Experience
Where I've worked.
I build the guardrails that keep AI agents on task: execution environments where an agent works under enforceable contracts, sealed off from the network by egress blocking so it can be tested safely before deployment. I also built the plumbing for canary deploys, routing HTTP traffic by header and creating and tearing down an isolated PostgreSQL database on Kubernetes for each one, and automated our testing and deployment with GitHub Actions.
I analyze in vivo imaging data from mouse visual cortex, recorded while the mouse watches visual stimuli, and fit linear and non-linear decoders to find patterns in how different regions respond. The question I keep coming back to is surround suppression: why a neuron can fire less when its stimulus gets bigger.
I built and run the website for our 501(c)(3) club. It turns 240+ nights of round-robin results into ratings, trends, and head-to-head records, and I keep adding what members ask for, like coaching signups and email updates. Once a quarter I report to the board on how the club is doing.
In eighth grade I kept showing up late to Zoom school, so I built an app that opens your meetings for you. It now has 2,000+ users in 75+ countries, a registered trademark, and a pending patent. The latest feature is an AI calendar import that finds your meetings so you can add them in one click. Along the way I've been on 9 EdTech podcasts and was the youngest presenter in EDUCAUSE history.
Before Berkeley I spent six months labeling songbird vocalizations in DeepSqueak, then compiling the data to see whether young birds learn their songs from their tutors or their parents.
Work
Things I've built.
LinkJoin
Full-stack virtual meeting aggregator with 2,000+ registered users and 4,000+ monthly visitors across 75+ countries. Real-time pub/sub over WebSockets, Argon2 password hashing, one-time token auth, and AI-powered calendar import. US trademark registered; patent pending.
ExplorePython
Python learning platform where students write real code from the first lesson rather than reading about it. Reached 3,000 monthly visitors and placed as an EdTech Digest Cool Tool Award finalist.
Recipe Transformer
Character-level transformer built from scratch in PyTorch, trained on millions of recipe examples. Uses a custom tokenizer and sliding window dataset pipeline for batched next-token prediction, with multi-head self-attention via learned Q, K, and V projections.
MNIST Polysemanticity
Tested the relationship between hidden layer dimensionality and polysemanticity in a 2-layer MNIST classifier, finding results consistent with superposition theory as described in Anthropic's "Towards Monosemanticity." (April 2026)
Concord Table Tennis Club
Public website for a 501(c)(3) nonprofit club. Computes player ratings, trends, and head-to-head records from 240+ round robin sessions, with a filterable results archive spanning 2+ years of play.
Recognition
Awards & press.
In the press
Stack
What I work with.
Contact
Let's build
something.
Likely hiking, running, lifting, coding, eating, or sleeping. Best reached by email.