model omar — converged

Portfolio — MMXXVI

Hi, I’m Omar.

A self-taught developer building machine‑learning models, private on‑device AI for iPhone, and web experiences that feel alive.

Scroll
  1. 01Machine Learning
  2. 02iOS Development
  3. 03Web Engineering

01Mission

My mission is to build — or help build — machine learning models that make life better for humanity. I’m self-taught: open-source code, documentation and countless experiments were my classroom. By day I study dentistry, where precision is measured in microns. By night I bring that same precision to models, apps and pixels.

how I learn

  1. read the docs
  2. read open-source code
  3. experiment
  4. repeat ↺
Learned through

Open source, documentation & experiments

Building toward

Machine learning that improves everyday life

Currently studying

Dentistry

02Capabilities

Three disciplines. One standard.

Everything I build sits where intelligence, interface and engineering meet — and every detail is treated like it matters.

fig. 1 — gradient descent on a loss surface

01 / Machine Learning

Models that learn.

Python and the foundations of machine learning — how models learn, how to train them, and how to make them small enough to live on a phone.

  • Python
  • Model training
  • On-device inference
fig. 2 — tokens streaming on-device

02 / iOS Development

Apps that belong.

Swift and Apple’s ecosystem. Native apps that feel fast, private and completely at home on iPhone and iPad.

  • Swift
  • SwiftUI
  • SwiftData
fig. 3 — a flow field, drawn live

03 / Web Engineering

Web that moves.

High-performance, interactive experiences — like this one: a hand-written WebGL engine, live neural networks and motion design.

  • JavaScript
  • WebGL / GLSL
  • Web Audio

03Selected work

Things I’ve built.

From a private AI that lives on your phone to the bots and tools in between.

Featured · iPhone & iPad

Surnia

Offline AI Chat

Private AI chat that runs entirely on your iPhone and iPad. Surnia runs open-source language models on the device itself — no account, no servers, no analytics. Nothing you type ever leaves your phone.

Inference
100% on-device, powered by llama.cpp
Models
Qwen3.5 · Gemma 4 · Phi-4 mini · SmolLM3 · Ministral 3 · LFM2.5 · ALLaM
Vision
Photos are understood locally by vision models
Apple Intelligence
Uses Apple’s on-device model when available
Built with
Swift · SwiftUI · SwiftData
Data collected
None
Surnia’s start screen: an owl mascot, the line “Everything stays on this device”, the Qwen3.5 4B model selected, and suggestions to write a short poem or explain how rain forms.

02 · TypeScript

Telegram Bot Builder

A button-driven Telegram bot for running a store. Customers tap through menus instead of typing commands: catalog, cart and checkout, orders and support tickets. Owners get an in-bot admin panel, broadcasts and analytics, all in Arabic and English.

  • TypeScript
  • Telegraf
  • PostgreSQL
  • Arabic · English
View on GitHub

03 · Python

Telegram Messaging Assistant

A personal relay bot. Anyone can message it, and the owner replies through it. It detects Arabic or English automatically, forwards text, photos, voice notes, stickers and documents, and handles replies and blocking without any database.

  • Python
  • python-telegram-bot
  • Arabic · English
View on GitHub

04Lab

Machine learning, live.

Not videos. Not screenshots. Two real neural networks, computed on your device while you watch.

Exp. 01Learning in real time

Watch a network learn.

A neural network written from scratch — forward pass, backpropagation and the Adam optimizer — training live in your browser. Pick a dataset, shape the network and watch the decision boundary form. Tap the plot to add your own points.

Epoch0
Loss—
Accuracy—
Network— parameters
Dataset
Hidden layers
3
Neurons
8
Activation
Learning rate
Tap to add

Exp. 02Open-source model

Draw a digit.

An open-source convolutional neural network, trained on MNIST’s 60,000 handwritten digits, compressed to 8-bit weights and shipped to your browser — the same on-device idea behind Surnia. Draw any number from 0 to 9 and watch every layer react.

Draw a digit here

Input · 28×28
Conv 1 · 8 filters
Conv 2 · 16 filters
Dense · 64
Prediction – waiting for input

    CNN · 2 conv + 2 dense layers · 54,314 parameters · int8 · 99.33% test accuracy · runs offline

    05Sound

    The listening room.

    Beyond code, I love developing high-quality headphones and advanced audio systems — and the quiet art of choosing and balancing them for true high-fidelity sound. Great sound is physics and taste in equal measure: a decibel at 3 kHz, a gentle shelf below 100 Hz. Press play, then shape the curve.

    5-band parametric EQ · biquad filters (RBJ cookbook) · music synthesized live with the Web Audio API · headphones recommended

    06Subtitles

    A quarter of a million viewers.

    In my free time I translate films and share the subtitles on OpenSubtitles.org under the same handle, 42l6. My subtitles for F1 alone have been downloaded around 250,000 times — not counting the other films I’ve worked on.

    250,000≈

    downloads · subtitles for F1 · published as 42l6

      Subtitling taught me what great software needs too: patience, obsessive attention to detail, and respect for audiences around the world.

      07Story

      Self-taught, on purpose.

      No bootcamp. No computer-science degree. Just open-source projects, documentation, and a long line of experiments that didn’t work — until they did.

      1. 01

        Python

        Programming, and the foundations of machine learning.

      2. 02

        Machine Learning

        How models actually learn — and how to train them.

      3. 03

        Swift

        Native apps for Apple’s ecosystem.

      4. 04

        Shipping

        Surnia — open-source language models running entirely on iPhone and iPad.

      Now

      Studying Dentistry.

      My degree is in dentistry; my obsession is software. Both demand steady hands, deep focus and zero tolerance for “almost right.” One trains my eye for detail — the other gives it somewhere to go.

      08Contact

      Let’s build something that matters.

      Have an idea, a project or a question about machine learning, iOS or sound? I’d love to hear it.

      Email omarluay@42l6.com