← Editor view

Noel Sason

UC Berkeley · Class of 2028

I build software for students and study how to evaluate AI systems.

Molecular and cell biology + data science. Research affiliate at Lawrence Berkeley National Laboratory, and in two labs on campus: one designing proteins, one building an adaptive math tutor.

Seeking research opportunities for summer 2027.

Selected work

All projects

Lectra

iPad app

An iPad workspace for PDF annotation, notebooks, and on-device computing.

Scope

Web & browser extension

Student tools for finding course materials and working alongside Canvas.

Overclock

Agent runtime

Browser agents that plan a task once, then replay it as a verified program.

Imagined-speech EEG audit

Independent research

A benchmark that a no-EEG baseline beats, and why its evaluation lets that happen.

Now

Undergraduate Researcher, Lewis LabOct 2026 – present
Undergraduate Researcher, CAHL LabSep 2026 – present
Research Affiliate, Berkeley LabJan 2026 – present

Building

Scope

2026 —

Student software built around existing courses: a browser extension for finding materials, an iPad workspace, and a course-grounded tutor. I am the sole developer across the products.

Lectra

on the app store

An iPad app for reading and marking up PDFs with an Apple Pencil. It also runs Python, git, a shell, and a code editor on the device itself, so the problem set and the code for it sit in the same place.

Polya

live

A tutor that works only from your own course materials. It gives hints and steps rather than answers, and every claim points back to the page, slide, or moment in a lecture it came from. Named after George Pólya.

Scope extension

on the chrome web store

A browser extension that searches across your Canvas and Brightspace courses and answers questions with citations. The index stays on your own machine.

Overclock

cal hacks 13.0

A browser agent normally looks at the page, asks a model what to do, clicks, and starts over. Overclock plans the task once, compiles it into a program, and replays the program, reading the page back to confirm every change it made. On the hackathon's own apps, graded by its checker, it reconciled 500 records in a warm median of 13.8 seconds with all 500 verified, and a second machine that imported the first one's compiled workflow finished without planning anything. What it has not shown is a site it has never seen, anything behind a login, or a number next to a published benchmark.

Blink2Talk

cal hacks 13.0

Someone who is awake but cannot speak or move their hands sees four short replies on a screen, looks at one, and holds their eyes closed to choose it; the reply is then spoken aloud. Four of us built it in a weekend. I wrote the screen the patient reads, the panel the people in the room see, and the layer that turns a held gaze into a choice. A prototype from a hackathon, nothing more: it has never been used with a patient.

Noel MCP

2026 —

One server that every assistant I use connects to, so they reach the same twenty-odd places I work: mail and calendar, my repositories and deployments, the databases, my courses, the literature search, the cluster. Reading is free. Anything that sends, publishes, or changes something stops and waits for me to approve that exact text or action. Running my own tools through an approval queue has taught me more about how to let an agent act safely than reading about it did.

Control

open source

A smart paste for the whole Mac: put the cursor in any field, press a key, and it reads what the field is asking for and fills it from a vault on your own disk. It tries what it has already learned, then a table of rules, and only asks a model about labels the rules cannot settle. How often it is right is measured against about 850 real labelled fields, and the build fails if a case that used to pass stops passing.

PyodideKit

open source

Real Python inside an iPhone, iPad, or Mac app, with no server and no network. NumPy, pandas, and matplotlib work, and so does installing packages offline. I pulled it out of Lectra after running into the same walls every iOS developer who tries this runs into, because there was no Swift package that solved them.

ios-agent-use

open source

Lets a coding agent open a real iOS app, tap through it, and get back a written account of what changed on screen: which buttons appeared, which disappeared, and where. Agents can already write Swift; this is how they check that it works. It installs as a Claude Code plugin.

Lectra Intelligence

2026

A vision-language model running on one Mac, with everything around it written to go with it: the harness that drives it, a chat app in the browser, an evaluation suite, and a probe that measures how often it refuses. Nothing calls a cloud API; the weights sit on a disk and the server answers only to that machine.

PersonalGraph

2026

A memory server for AI assistants, so a conversation doesn't start from nothing every time. It keeps what it knows as a graph, tags each fact with where it came from, and stores everything locally. It's the memory my own coding sessions run on.

SCI Cause-List Alerts

prototype

India's Supreme Court publishes tomorrow's hearing list as a PDF each evening, and if you're representing yourself, checking it every night is the whole job. This watches for your case and emails you when it appears. The court's server returns a 200 for files that don't exist, so the checker reads the body instead of trusting the status code. It's finished. I never deployed it.

Homelab

2024 —

A Raspberry Pi in my room that I can reach from anywhere. It backs up my machines, blocks ads for the house, and runs a few things I wrote for it: a status dashboard, a 7am brief, spoken alerts through a Google Home, a tracker for money people owe me. It's all on my own network, so there's nothing to link to.

Research

A published benchmark reports 13.88% accuracy at telling 39 categories of imagined words apart from EEG. Copying the label of the previous training trial, with no EEG at all, scores 18.8%, because the trials run in category clusters and every test trial sits inside a training run. With whole runs held out, real labels beat shuffled ones by 0.64 points, within noise. One participant, self-published, not peer reviewed.

Scope Tutor-1

in progress

Whether fine-tuning a small open model on tutoring conversations makes it a better tutor than the same model handed a carefully written prompt. The prompt it has to beat is frozen, the problems are split so none of the test problems appear in training, and the thing being measured was written down before any model was trained. Training runs are going now; nothing has been scored on the held-out problems, and the answer may well be that the prompt wins.

Fabric

2026, study completed

Designed and tested an uncertainty-routing method for self-supervised vision. After an exploratory study, a pre-registered experiment across eight seeds found no supported improvement from hierarchical targets under the tested protocol. Closed the study with documented results and reproducible analysis. The experiments used synthetic data; no real-image result is claimed.

AdNeuro

in progress

Predicting, second by second, the cortical response a video ad produces, and scoring that against real engagement before the media money is spent. It predicts cortical response, not thoughts, and the category it sits in is discredited for cause. The encoder beats chance on held-out validation, which is the floor, not a result. The comparison against a strong baseline is the piece that would settle it, and that run is what I am working toward.

Papers

Papers page

Also

Experience

Undergraduate Researcher, Lewis Lab, UC Berkeley

oct 2026 —

Designing peptides meant to hold one of the mitochondrion’s RNA-degradation complexes together, so that less of a particular kind of RNA builds up where it should not. The designing is done by a model and checked on a cluster; I am learning both.

Undergraduate Researcher, Computational Approaches to Human Learning Lab, UC Berkeley

sep 2026 —

Working on OATutor, the lab's open-source adaptive math tutor. It uses Bayesian Knowledge Tracing to estimate what each student has mastered and choose their next problem. My project gives every image in a problem its own description for screen-reader users, carried from the spreadsheets authors write in through to the page.

Research Affiliate, Lawrence Berkeley National Laboratory

jan 2026 —

Built the trait-profile generator for the KG-Microbe search pipeline, the step that turns each study's results into a table you can compare across diseases. Reviewed by the lab and merged into main.

COPE Health Scholar, Kaiser Permanente

jun 2025 —

Patient support on hospital floors: vital signs, mobility, discharge coordination.

Finance Director, ASUC

sep 2024 —

Administered about $200,000 in student-organization funding: allocations, grants, reimbursements, waivers.

Co-President, South Indian Society, UC Berkeley

2024 —

Co-run a student organization of more than 200 members.

Community Service Officer, UC Berkeley Police Department

2024 – feb 2026

Walked students home late at night and worked campus safety shifts.

State Secretary, California HOSA

2023 – 2024

Ran conferences and chapter operations for the state organization.

Education

University of California, Berkeley

Aug 2024 – May 2028

Tools on this site

Search all tools

Files and images

  • PDF notes: Type notes onto any PDF, combine files, and reorder or remove pages. The file never leaves your device.
  • Image tools: Convert HEIC photos and screenshots to JPG, PNG, or WebP, resize them, and remove backgrounds. Your photos never leave your device.
  • QR studio: QR codes made in the browser: presets, palettes, logos, halftone art. Uploads nothing.

Everyday

  • Split a bill: Split a bill evenly or by item, then send each person their share with a Venmo link.
  • Word counter: Count words and characters as you write, and see where you go over a limit.
  • Deadlines: Paste an email, get a countdown and a calendar entry. Your list stays on your device.
  • When I'm free: My open times over the next seven days, shown in your time zone. Tap one to copy a message you can text me.
  • time2meet: A study rebuild of when2meet. One link, everyone paints when they are free.

Reading

  • YouTube transcript: Read any YouTube video’s transcript, jump to a moment, and save it as text or subtitles.

For developers

  • JSON API: Read-only JSON under /api/v1. No key, any origin.

Friends

  • Friends: A living wall of moments with the people I know, always fresh.