Matt Okner

EECS @ UC Berkeley · Machine learning.

Berkeley, CA

I'm an EECS student at Berkeley working toward ML engineering. I like training models end to end and understanding the math underneath them. I recently trained a sparse mixture-of-experts language model from scratch, and I've built a real-time voice agent on top of LLMs.

GitHub X LinkedIn Email CV

Experience

Wrodium

Software Engineering Intern

InternshipJan 2026 – Mar 2026Berkeley, CA

Shipped a production content-synchronization pipeline that parses external URLs, diffs them against internal mirrors, and surfaces approve/reject proposals in an inline review UI
Designed a content refresh system on the Perplexity Deep Research API and PostgreSQL that flags stale pages and drafts revisions; refreshed pages were cited ~60% more often by AI search engines
Built scraping and backfill pipelines in Python and TypeScript with scheduling, retry logic, and deduplication across 500+ documents

Python · TypeScript · PostgreSQL · Perplexity API · n8n

Blockchain at Berkeley

Developer

Dev / ConsultingAug 2025 – PresentBerkeley, CA

Building Solidity smart contracts and on-chain development projects

Solidity · Smart Contracts · On-chain Systems

San José State University

Machine Learning Research Intern

ResearchMay 2023 – Aug 2023San José, CA

Trained a multilayer perceptron in PyTorch for 10-way malware family classification on 15,000+ labeled samples, reaching 91% accuracy on held-out test data
Engineered feature extraction pipelines converting opcode sequences, byte sequences, and byte histograms into structured model inputs

Python · PyTorch · Malware Analysis · MLP

Projects

GitHub →

miniMoE

Built · 2026

280M-param MoE LLM trained from scratch; beats GPT-2 on HellaSwag.

Python · PyTorch · DDP · tiktoken · Hugging Face

GitHub Weights

Details

A sparse mixture-of-experts language model pretrained on 10B tokens and instruction-tuned for chat. Each token runs 2 of 8 experts, so the model carries 280M parameters but spends 110M.

Model: 6 transformer blocks, 8 experts each, top-2 routing
Pretraining: 10B FineWeb-Edu tokens over 19k steps on 8 GPUs with DDP
Chat: Instruction-tuned on smol-smoltalk with prompt-masked loss, cutting SFT validation loss from 1.89 to 1.54
Result: 31.2% HellaSwag acc_norm vs 29.6% for GPT-2 124M, same harness
Analysis: Routing ablations show the learned router beats random routing by 6 points

Warden

Built · 2026

Policy layer for agents before tools run.

TypeScript · OpenAI Agents SDK · Node · pnpm · YAML policies

GitHub

Details

Runtime guardrail for OpenAI Agents SDK tools. It checks tool names, schemas, and arguments, then allows, blocks, or asks for approval before side effects happen.

Tool boundary: guardTools() wraps an agent's tool array and classifies, policy-checks, and audits every call before execution
Policy engine: YAML risk labels with tiered argument rules: refunds under $50 auto-run, $50 to $500 need approval, over $500 are denied
Approvals: Fail-closed approvals via terminal, Telegram, or callback; denied on timeout
Audit trail: JSONL decision logs with secret redaction
Setup: CLI scaffolds a fully guarded starter agent

Procrastinot

ETH · 2026

USDC commitment stakes with oracle-judged evidence.

Next.js · TypeScript · Solidity · Foundry · Supabase · USDC

GitHub

Details

Commitment app where users stake USDC, submit proof before a deadline, and receive a pass, retry, or forfeit outcome from an oracle.

Contracts: Foundry contracts for commitments, attempts, verdicts, forfeits
App: Next.js wallet flow for staking and evidence upload
Oracle: Evaluates evidence and writes pass/fail outcomes on-chain
Indexer: Syncs contract events into Supabase for fast reads
Settlement: USDC escrow with retry and forfeit paths

Sage.ai

Built · 2025

Real-time phone agent for booking, intake, and triage.

Python · Twilio · Deepgram · Anthropic API · MongoDB · WebSockets · Railway

GitHub

Details

Voice AI for inbound and outbound calls where response time and reliable handoff matter.

Latency: Audio-in to spoken response in under 1 second
Transport: Twilio Media Streams over WebSockets
Speech: Streaming Deepgram ASR piped into the Anthropic API
Turn-taking: Tuned end-of-speech detection to tell a caller pausing from finishing a turn
Booking: Google Calendar tool calls book appointments mid-conversation without handoff

Show 4 more projects

CharCounting

Evaluations · 2026

Eval for when LLMs should stop guessing and use tools.

Python · OpenAI API · Prime Lab · Verifiers

GitHub

Details

Compared GPT-4o character counting with direct reasoning versus a Python interpreter. Tool use removed the measured counting errors.

Baseline: 77.8% accuracy without tools on repeated or long strings
Tool run: 100% accuracy with a Python interpreter
Finding: Token boundaries cause avoidable counting mistakes
Harness: Reproducible eval prompts and result checks

Zetamac, but Better

Built · 2026

SwiftUI mental-math trainer with an AI coach.

Swift · SwiftUI · Anthropic API · UserDefaults

GitHub

Details

Native iOS arithmetic trainer with zetamac-style drills, local history, and an AI coach that turns recent solve times into practice plans.

Platform: SwiftUI app for iOS 17+
Profiles: Local users and score history in UserDefaults
Coach: Anthropic-powered feedback from recent solve times
Drills: Zetamac-style presets by operation and difficulty
Stats: Tracks streaks, speed, and accuracy over time

PE Malware MLP

Presented · 2023

91% accuracy malware family classifier on 15k PE-file samples.

Python · PyTorch · NumPy · Matplotlib

Details

Static Windows PE-file classifier trained on a curated VirusTotal slice and presented at the SJSU Research Symposium 2023.

Model: Multilayer perceptron in PyTorch for 10-way family classification
Data: 15,000+ labeled VirusTotal samples
Features: Opcode sequences, byte sequences, and byte histograms
Result: 91% held-out test accuracy vs 10% chance

Smart Irrigation System

Deployed · 2022

Raspberry Pi + Flask automated garden controller.

Python · Flask · Raspberry Pi · SQLite · HTML/CSS

Read on arXiv

Details

Raspberry Pi controller that reads garden sensors, drives solenoid valves, and logs moisture data through a small Flask API.

Sensors: Capacitive soil probes plus temperature and humidity
Control: Schedule and threshold watering modes
Valves: Solenoid control with manual override
Logs: SQLite moisture and run history
Result: About 30% water reduction in a 200 sq ft test garden

Education

UC Berkeley

Aug 2025 – May 2028

B.S. Electrical Engineering & Computer Science

GPA 4.0 · Dean's List, College of Engineering

Coursework

CS 182Deep Neural NetworksFall 2026
CS 162Operating Systems and System ProgrammingFall 2026
IEOR 221Introduction to Financial EngineeringFall 2026
CS 189Introduction to Machine LearningA
EECS 126Probability and Random ProcessesA
CS 61CGreat Ideas of Computer Architecture (Machine Structures)A
CS 61BData StructuresA+
CS 61AStructure and Interpretation of Computer ProgramsA
CS 70Discrete Mathematics and Probability TheoryA+
EECS 16AFoundations of Signals, Dynamical Systems, and Information ProcessingA

West Valley College

Aug 2023 – May 2025

A.S. Mathematics / A.S. Data Science / A.S. Physics

GPA 4.0 · Summa Cum Laude

Coursework

MATH 3ACalculus IA
MATH 3BCalculus IIA
MATH 4ACalculus IIIA
MATH 4CLinear AlgebraA
MATH 4BDifferential EquationsA
PHYS 4APhysics: MechanicsA
PHYS 4BPhysics: Electricity & MagnetismA
PHYS 4CPhysics: Light, Heat & WavesA
CHEM 1AChemistry HonorsA

Contact

Email is the fastest way to reach me.

mokner@berkeley.edu

GitHub X LinkedIn CV