UC Berkeley · Open to Summer & Fall 2026 internships

Noel Sason — Software & AI Developer. Useful things,
made well.

I'm Noel Sason — a UC Berkeley undergrad studying Molecular Biology & Data Science. Founder of Canvascope, research affiliate at Lawrence Berkeley National Laboratory, and a builder working at the intersection of software, data, and medicine.

See selected work Get in touch Résumé ↗
Based in
Berkeley, CA
Studying
MCB · Data Science
Building
Canvascope
Researching
LBNL · Arkin Lab
46active Canvascope users
10+university domains
$200K+student funding administered
80+disease–microbiome studies
<100mssemantic search
4products shipped end-to-end
3.8GPA at UC Berkeley
200+member org served
46active Canvascope users
10+university domains
$200K+student funding administered
80+disease–microbiome studies
<100mssemantic search
4products shipped end-to-end
3.8GPA at UC Berkeley
200+member org served
01 — Selected Work

Things I've shipped

Products, research systems, and tools — each one closing the gap between a system that technically works and the people who have to use it.

012024 — Present·canvascope.org

Canvascope ↗

Founder · CEO · Lead Scientist

A privacy-first AI study assistant for students — a Chrome extension (Manifest V3) with GCP OAuth, local-first indexing, and sub-100ms semantic search across every course PDF, page, and assignment. Includes a citation-grounded RAG pipeline, on-device Gemini Nano inference, the Lectra iPad companion (Apple Pencil + DropBridge transfer) — now a full sandboxed coding workspace with a from-scratch POSIX shell, real git (isomorphic-git bridged to native Swift over a WKWebView), and an embedded Python runtime — a multi-step, tool-using AI agent that runs a client-driven tool-use loop inside the MV3 service worker, planning and executing integrity-gated read/write tools (corpus search, deadlines, grades, calendar, todo/event/study-plan creation) with a daily briefing, prompt-cached system context, an append-only audit log, and a kill switch — and PersonalGraph, a TypeScript MCP server.

46
active users
10+
university domains
<100ms
semantic search
Chrome MV3TypeScriptRAG · embeddingsTool-use agentGemini NanoSupabaseSwift / iPadOSisomorphic-gitPyodideDropBridgeMCP server
canvascope — semantic search
query
→ BIO 1A · Lab 4 — Enzyme Kinetics.pdf98%
→ Module: Catalysis & Rates91%
→ Assignment: Problem Set 487%
dropbridgepair → device-token upload → sync
sourceindexed locally · 0 data left device
02Dec 2025 — Present·rxbrief.org

RxBrief ↗

Co-Founder · CTO

A full-stack healthcare platform that collapses two daily clinician headaches into one tool. Instead of digging through PDF inserts, clinicians pull instant FDA-sourced label data — indications, dosing, interactions; instead of checking a dozen boards, they see openings from Greenhouse, Lever, and RSS feeds normalized into a single, server-filtered feed, with real-time Supabase chat keeping the conversation in-app. Every feature was scoped directly from interviews with healthcare professionals, so it solves the problems they actually named — turning scattered, manual lookups into seconds of work.

3
job sources unified
FDA
live label lookup
RT
supabase chat
Next.jsReactTypeScriptSupabaseFDA APIVercel
rxbrief — drug lookup
GET
drugNesina (alogliptin)DPP-4
indication · type 2 diabetes, adjunct to diet
dosing · 25 mg PO once daily
interactions · insulin · sulfonylureasrisk ↑
sourceopenFDA · structured label · cached
03Jan 2026 — Present·Arkin Lab

Microbial Trait Pipelines

Research Affiliate · Lawrence Berkeley National Laboratory

A DuckDB-backed pipeline over the KG-Microbe knowledge graph enabling SQL-based traversal of taxa, chemicals (CHEBI), environments (ENVO), and phenotypic traits — powering cross-cohort meta-analysis across 80+ disease–microbiome studies. A YAML-driven generator emits reproducible long/wide TSV tables, standardized with uv and Justfile task runners.

80+
studies meta-analyzed
4
disease cohorts
0
env-setup issues
PythonDuckDBKG-MicrobeGTDB / NCBIuvJustfile
pipeline — kg-microbe.duckdb
SELECTtaxon, trait FROM kg
JOIN chebi USING (chem_id)
WHERE study IN (80+ cohorts)
→ scanning 2.4M edges …duckdb
→ 14,820 rows0.31s
outtraits_long.tsv · traits_wide.tsv
042024 — Present·UC Berkeley

SIS Community App

Co-President · South Indian Society

A native SwiftUI iOS app replacing 4+ fragmented tools for a 200+ member organization — Stripe-integrated event management, a Cloudflare R2-backed shared photo album, real-time messaging, and a Supabase backend with role-based access control, committee-scoped archives, and push notifications.

40
beta testers
200+
member org
4+
tools replaced
Swift / SwiftUISupabaseStripeCloudflare R2RBAC
sis — role-aware club os
authsession · supabase RLS
✓ events · create / edit / delete
✓ stripe payouts · refunds
✓ roles · manage 200+ members
✓ photos · files · chat
05April 2026·DataHacks · Scripps

CLENS

Frontend & ML — Team of 4 · Scripps Challenge Track

An iOS app that turns grocery shopping into an ocean-health feedback loop. A binary CoreML classifier trained on 400+ images auto-routes the camera between receipts and barcodes at ~6 fps; Apple Vision OCR parses line items; and a zero-dependency ridge regression — codegenned into Swift coefficients — predicts CO₂, runoff, water, and plastic footprints on-device. A live "marine stress" multiplier is derived from Scripps CCE2 mooring data via STL decomposition and an Isolation Forest.

400+
images trained
~6 fps
on-device scan
0
runtime ML deps
SwiftUICoreML / CreateMLApple Vision OCRFlaskscikit-learnSupabase
clens — oceanscore
scanreceipt ⇄ barcode · hysteresis fsm
→ Vision OCR · 18 line items~6 fps
→ categorize · produce · dairy · plastic
→ ridge · CO₂ · runoff · water · plastic
stressCCE2 mooring · STL · isolation forest
OceanScore72/100
062024 — Present·Raspberry Pi 5

Self-Hosted Homelab

Systems & Infrastructure

A private cloud on a Raspberry Pi 5 (8GB, Debian arm64) reachable anywhere over Tailscale. OpenMediaVault serves a 1.8TB SSD as a Time Machine–backed NAS over SMB; a Dockerized Pi-hole handles DNS adblock. On top sit homegrown services: a telemetry dashboard (Express + cron + Upstash Redis), an edge-TTS voice-alert server casting to Google Home, a 7am OpenClaw daily-brief generator, and a Zelle/Venmo repayment tracker over SQLite.

1.8TB
NAS storage
5+
self-hosted services
24/7
uptime
Linux / DebianDockerOpenMediaVaultTailscaleNode.jsPython
pi — telemetry
hostrpi5 · debian arm64 · tailscale
nas: 1.8TB ssd · smb · time machine
pi-hole · edge-tts · daily brief
cpu21%
temp51°C
disk62%
live
02 — Experience

Where I've worked

Research, healthcare, campus finance, and student leadership — the operational ground beneath the products.

Jan 2026 — Present
Research Affiliate
Lawrence Berkeley National Laboratory · Arkin Lab
Building reproducible Python/DuckDB pipelines that normalize heterogeneous microbial trait and signature datasets into analysis-ready long/wide tables for cross-cohort meta-analysis.
Sep 2024 — Present
Finance Director / ASUC Roles
Associated Students of the University of California
Administered ~$200K in student-organization funding across CASSA allocations, grants, reimbursements, and waivers — with on-time disbursements and zero compliance violations.
Jun 2025 — Present
COPE Health Scholar
Kaiser Permanente
Direct patient support across 5+ clinical departments — vital signs, mobility, and discharge coordination — keeping my healthcare-software instincts grounded in real bedside workflow.
2024 — Feb 2026
Community Service Officer
UC Berkeley · UCPD
Supported campus safety operations through late-night student escorts and student-facing service — reliability under shift work and high-stakes contexts.
2023 — 2024
State Secretary
California HOSA
Led healthcare-leadership programming at statewide scale — conference logistics, fundraising systems, and chapter operations across California.
03 — About

The throughline

B.A. Molecular Biology · B.A. Data Science · UC Berkeley · GPA 3.8 · Class of 2027.

I keep ending up inside broken workflows — Canvas tabs, reimbursement systems, research data formats, clinical handoffs — and building the missing interface.

The underlying systems usually already work. What's missing is the usable layer: the part a student, a clinician, or a researcher actually touches. That's the gap I build into.

I founded Canvascope, a privacy-first academic platform used across 10+ universities, and I work at Lawrence Berkeley National Laboratory on reproducible data pipelines for microbiome research. In between, I've run $200K+ in campus finance, supported patients at Kaiser, and shipped iOS, web, and Chrome-extension products end to end.

I care about clean code, intuitive products, and shipping things that work. If you build at the intersection of student tools, data, or medicine, I'd love to talk.

NowCanvascope · LBNL pipelines
StudyingMCB + Data Science
SchoolUC Berkeley '27
GPA3.8 / 4.0
FocusSoftware × data × medicine
Open toInternships · research
04 — Toolkit

What I reach for

Grouped by where it shows up in the work.

Languages
PythonTypeScriptJavaSwift / SwiftUIJavaScriptSQLBashHTML / CSS
ML & Data
PyTorchscikit-learnPandasNumPyDuckDBMongoDBEmbeddings & vector searchSTL · anomaly detection
Web & Mobile
Node.jsNext.jsReactiOS / SwiftUIChrome Extensions (MV3)FlaskFastifyONNX Runtime WebREST / API design
Infra & Tools
GitGitHub ActionsDockerLinuxSupabaseVercelStripeCloudflare R2TailscaleOpenMediaVaultuvJustfile
05 — Direction

What drives the work

The strengths I lean on, where I'm headed, and the problems I keep coming back to.

Strengths

I ship end-to-end and finish what I start — from a Chrome extension to an iOS app to a research pipeline, I own the whole stack and the last mile that usually gets dropped. I learn fast under ambiguity, reason from first principles, and I am relentless about the details that make a tool feel reliable instead of merely functional.

Full-stack ownershipFast learnerDetail-drivenPragmatic
Aspirations

I want to build AI and data products that real people depend on — tools that earn trust because they work every time, not demos. Long-term I am aiming at the intersection of software, data, and medicine: systems that make expert knowledge usable for the students, clinicians, and researchers on the front line.

Trustworthy AIProducts at scaleHealthcare impact
Areas of interest

Applied machine learning and retrieval, on-device and privacy-first AI, and the data infrastructure that makes research reproducible. I am drawn to problems where a clean interface and a rigorous pipeline turn messy, real-world data into a decision someone can actually act on.

Applied ML & RAGOn-device AIReproducible dataHuman-centered tools
06 — Contact

Let's build something that works.

noel_sason@berkeley.edu
GitHub ↗ LinkedIn ↗ Résumé ↗