Kozak Technologies Inc

Kozak Technologies Inc/Los Angeles/Est. 2024

Speech AI that never leaves the device.

An independent engineering practice building on-device machine learningand mobile systems — inference that runs on the phone’s Neural Engine, not in the cloud. Founded by Nazar Kozak, IEEE Senior Member.

input: microphone↳ stays on-device
◇ On-device inference◇ Swift 6 · CoreML · ANE◇ IEEE Senior Member◇11 years · mobile & ML
01Product

One product carries the research.

Disfluo— speech AI for stuttering and fluency practice that runs entirely on the user’s iPhone. Built for the problem the founder has lived with since age seven.

Disfluo

On-device speech AI
In development · applied R&D

Fully on-device speech AI for stuttering and fluency practice. All inference runs on the user's iPhone — audio never leaves the device. Built as the applied-research arm of Kozak Technologies, and in active development.

  • Privacy by design
    Speech inference runs entirely on-device via CoreML and the Apple Neural Engine. Audio never leaves the iPhone — there is no server to send it to.
  • Open models
    Fine-tuned Whisper adapters released openly on the Hugging Face Hub, so the recognition models can be inspected and reproduced.
  • Research-led
    Developed alongside preprints on disfluency-aware speech recognition; one manuscript is under peer review at Speech Communication (Elsevier).
  • Why it exists
    It grew directly out of the founder's lifelong stutter — the private, on-device practice tool he wanted to exist.
whisper-small-disfluent-smoothed-lora↗whisper-small-disfluent-verbatim-lora↗Dashboard — work in progress→
02Engineering services

Selective consulting and senior-IC engagements.

Capability areas, scoped tightly and chosen for technical fit.

  1. 01iOS & macOS native developmentSwift 6, SwiftUI
  2. 02On-device ML deploymentCoreML, the Apple Neural Engine, model quantization & LoRA adaptation
  3. 03Cross-platform mobileReact Native, Flutter
  4. 04Mobile SDK architecture & distribution at scaleversioning, rollout, multidex / dynamic loading
  5. 05Speech & audio signal processingon-device ASR, fluency & prosody analysis
03Selected engagements

Eleven years, named roles.

A decade across founding-engineer and senior-IC mobile roles.

PeriodOrganizationScope
2025 — present

Yahoo Inc.

Staff Software Engineer · Orion Browser (engaged via Toptal)
Real-time tab clustering and AI-assisted browser features; contributor to the OrionClusteringSuite testing framework.
2025

American Honda Motors

Mobile Tech Lead
Vehicle companion app architecture and release management.
2022 — 2024

Datasya Inc.

Lead Mobile Engineer → Head of Mobile
Built and led the mobile engineering team; cross-platform delivery and on-device machine learning.
2023 — 2024

Shader Inc.

Senior iOS Engineer · part-time
iOS performance work and GPU optimization.
2014 — 2018

Appodeal Inc.

Founding Engineer
Built the Unity SDK and a dynamic multidex loader handling the Android 65K-method limit across 20+ integrated ad frameworks.
04Founder

Nazar Kozak.

Nazar Kozak has stuttered since the age of seven. Speech as something effortful — and the tools for it being scarce, clinical, and rarely private — is the direct motivation for Disfluo and for the on-device speech work at the center of Kozak Technologies.

waveform · disfluent speech

He is a software engineer with eleven years across mobile platforms, cross-platform SDKs, and on-device machine learning. Currently Staff Software Engineer at Yahoo on the Orion Browser team; previously founding engineer at Appodeal, Lead Mobile / Head of Mobile at Datasya, and Mobile Tech Lead at American Honda.

He publishes preprints on disfluency-aware speech AI and releases his models openly so the work can be checked.

  • IEEE Senior Member
  • FIET — application under assessment

Publications and full research detail at nazarkozak.github.io

05Research & open source

Open, and checkable.

Everything here is public and verifiable — preprints, open-source models, and academic detail. One manuscript is under peer review at Speech Communication (Elsevier); the rest are preprints.

PreprintsDisfluency-aware speech recognitionOne manuscript under peer review at Speech Communication (Elsevier); others available as preprints.nazarkozak.github.io ↗Open modelsFine-tuned Whisper disfluency adaptersReleased openly on the Hugging Face Hub.huggingface.co/nazarkozak ↗Academic siteResearch profile & publication listFull detail, kept current.nazarkozak.github.io ↗