All projects

Case study

ZOYA OS

Product vision and development · 2026

A personal AI operating system built for one owner: it remembers approved context, answers from my own notes with sources, and acts only with permission. A product vision in active development.

Active Development
Planned Skills
12
Working Locally
3
Voice Round Trip
Not yet

Project overview

ZOYA OS

ZOYA OS is my personal AI operating system, in the spirit of a JARVIS-style assistant. The aim is an assistant I own and control, one that connects my knowledge, projects, tasks and tools instead of starting every conversation from zero. It is not the chatbot on this site, which is a separate, simple project named Zoya. ZOYA OS is private, in active development, and only part of the plan works today.

Project type
Personal AI operating system (product vision)
Status
Active development, private
My role
Product vision and development
Timeline
2026, ongoing
Access
Owner only. No public demo.
Engineering underneath
Brain V2

Tools Used

Technologies, platforms, and tools used to design, develop, and deliver this project.

  • React, a JavaScript library for user interfaces
  • TypeScript, JavaScript with types
  • Vite, a front-end build tool
  • Three.js, a 3D graphics library for the web
  • React Three Fiber, Three.js for React
  • Express, a Node.js web server framework
  • Tailwind CSS, a utility-first CSS framework
  • Shadcn UI, a React component library
  • Auth0, sign-in and identity
  • Obsidian, a note-taking and knowledge app
  • LiveKit, real-time audio and video infrastructure
  • Faster-Whisper, speech-to-text
  • Kokoro, text-to-speech
  • ONNX Runtime, a machine-learning inference engine
  • Silero VAD, voice-activity detectionOptional
  • RAG, retrieval-augmented generation
  • Agent Skills, reviewed capabilities an AI agent may use
  • GitHub Actions, automated workflows
  • Gitleaks, a secret scanner for code

The challenge

Most AI assistants are a chat window. They forget what you told them last week, they cannot show where an answer came from, and the moment they are given tools, it is unclear what they are allowed to touch. ZOYA OS starts from those three gaps. It is also not the Zoya chatbot on this site: that one is a small, rule-based guide to my portfolio, and this is a separate, private system.

  • The problem

    • Projects, notes and decisions are spread across many places, and every new chat starts from nothing.
    • An answer with no source cannot be trusted for real work.
    • An assistant that can act needs clear limits on what it may read, change and send.
    • Most assistants are a chat window, and they forget what you told them last week.
  • What ZOYA is not

    • It is not the Zoya chatbot on this site, which is a small, rule-based guide to my portfolio.
    • It is a separate, private system.
    • It is not just a chat window.
    • Abilities that are not built are said plainly, not implied.
  • What ZOYA is for

    • One assistant that knows my projects and my context, and says where each answer came from.
    • Memory that I approve, so it stays accurate and I stay in control.
    • Tools that run only with permission, one skill at a time.
    • One assistant across all my projects, not a new chat for each.
  • What I wanted to improve

    • Answer from my own notes, with the source shown, instead of guessing.
    • Keep everything private and owner-controlled by default.
    • Add abilities slowly, each behind its own approval.
    • Remember what I told it last week, with my approval.

How the system is meant to fit together

ZOYA has five layers. Open a layer to see what it holds and where it stands today. The notes say whether a part works on my machine, is planned, or is waiting on something else.

ZOYA OS, the layers

How one question moves through it

  1. I ask a question
  2. It checks that the request comes from me
  3. It finds the matching approved notes
  4. It answers and names its sources
  5. Anything beyond reading waits for my approval

Requirements and technical plan

The order matters more than the speed. Trust comes first, then reading, then memory, then voice, then any action in the outside world.

  • Now: a trusted core

    Private dashboard, answers grounded in my notes, approved memory and three skills, all working on my own machine.

  • Next: from my machine to production

    Set up the hosted database, choose a model provider I am willing to send private notes to, and deploy only after the security checklist passes.

  • Then: voice

    Build and measure a worker that listens and speaks. Voice counts as done only after a real spoken question gets a spoken answer in a measured time.

  • Later: reach into my other tools

    Task automation, links to approved apps, proactive help, access from more than one device, and specialist workflows that always ask a human first. None of these is built.

The order is a plan. Dates are not promised, and the plan changes when something fails a security check.

How it's built

Where it stands today, from the project's own status tables. Working means it runs and passes its tests on my machine, not in public.

  • 12

    Skills planned

  • 3

    Working locally

  • 9

    Waiting on connections

  • Not yet

    Voice round trip

  • Working locally

    Dashboard, grounded answers with citations, approved memory, and the three skills.

  • Checked on real accounts, not in production

    Owner-only sign-in and reading approved notes from the private repository have been tested against real accounts, but only from my machine.

  • Planned or blocked

    A hosted memory database, a production deployment, the speech worker and the nine outside-facing skills.

Nothing here is deployed. ZOYA OS has no public site, and its repository and dashboard are private.

Key features

  • Grounded answers

    Replies come from approved notes and carry the note and version they used. If the notes do not cover a question, ZOYA is meant to say so rather than fill the gap.

  • Memory with approval

    New memories start as proposals. Only the ones I approve, and that have not expired, can shape an answer. Working locally; a hosted database is still to come.

  • Skills with permissions

    Twelve skills are planned. Three work today: search my notes, plan a task and manage approved memory. The other nine need safe connections first and are marked blocked, not done.

  • A private dashboard

    An owner-only view with a glowing 3D visual, charts, chat and skill panels. Its overview is built from a counts-only snapshot, so it shows totals and charts, not note text.

  • Voice, planned

    The browser side and short-lived access passes are in place. The speech engine that listens and talks back is not working yet, so voice is not claimed as a feature.

  • Project intelligence, planned

    Reading repositories and tracking projects is a future skill. It is blocked until it can run in a safe, limited environment.

My part

  • What I own

    • The product idea and what it should and should not do.
    • The trust rules: owner-only access, approval for memory, one permission per skill.
    • The order in which abilities are added.
    • The development of the working parts, using AI coding assistants under my direction.
  • Not claimed here

    • Voice, outside-app connections, proactive help and multi-device access are plans.
    • No public users, no benchmarks and no production traffic exist.
    • The Zoya chatbot on this site is a separate project.

On everyscreen

The portfolio home page on an iPhone

iPhone

The portfolio home page on a tablet

Tablet

The portfolio home page on an Android phone

Android

You may also like

Brain V2 screenshot

Brain V2

Active Development

The knowledge and engineering base behind ZOYA OS: a structured note vault, a private dashboard, reviewed AI skills and security rules in one repository. In active development.

Brain Agent

In Development

A private voice assistant that answers questions from my own notes, out loud, and only from what is actually written there. It can also save new notes.

The chat assistant on this site. It answers questions about my projects, skills and experience using only what is on the site, and alerts me on Telegram when a visitor wants to talk.