Skip to content
YA
All work
In developmentAndroidAIDesktop

NOVA

A voice-first personal AI operating system — where a permission layer, not the model, decides what may run.

Status
0.1 verified on a device; 0.4 Tools shipped; 0.5 Desktop agent in progress.
Period
2026
Platforms
Android · Windows
Repository
Private — available on request
Live
—
NOVA — personal AI OS campaign art

01Overview

An Android assistant designed to grow into a Windows desktop agent and a cloud backend. The language model is one component inside a deterministic system: NOVA Core owns context, memory and routing, every tool call passes one validated, permission-checked, logged gateway, and the phone can pair with its own PC agent over TLS.

By the numbers

unit tests (0.1)
152
unit tests (0.1)
README
architecture decisions
40+
architecture decisions
docs/DECISIONS.md
Kotlin files
312
Kotlin files
source tree
commits
77
commits
git history
  1. 02The problem

    Assistants that can act on your devices are only as safe as the code around the model. Prompted 'guardrails' are suggestions; a model that can call tools must not be the one deciding which calls are allowed.

  2. 03The vision

    One assistant across phone and PC that talks naturally in Levantine Arabic, remembers only what you allow, and leaves a tamper-evident record of everything it did.

  3. 04The solution

    Every tool call passes a single choke point — lookup → exposure → validate → lock screen → authorize → confirm → execute → audit. Medium- and high-risk actions are confirmed on screen. The model only receives the tools that can actually run right now, so it never pretends to act.

05Key features

  1. 01

    Streaming voice with barge-in

    Answers are spoken sentence by sentence while they stream. Interrupt by voice (with echo cancellation) or by tap; NOVA stores only what you actually heard and tells the model where you cut in.

  2. 02

    Memory you control

    'Remember that…' saves a memory (secrets refused). NOVA may suggest a memory but never saves without your tap. Search, edit, export or clear.

  3. 03

    Agent Trace

    A tamper-evident activity log: each step is hash-chained with SHA-256 and redacted before storage. The verifier reports the first edited, missing or moved step.

  4. 04

    Vision and screen awareness

    Camera or gallery photo → quick actions (what is this, read text, explain error, summarize), and screen-aware answers when you summon NOVA over another app.

  5. 05

    Desktop agent

    Pair by QR with NOVA Agent on Windows over LAN (TLS 1.3, pinned key). The PC keeps its own permissions, a kill switch and an always-on-top indicator.

  6. 06

    Model routing

    Gemini in the cloud or any OpenAI-compatible server on your network (LM Studio, Ollama), with ONLINE / LOCAL / HYBRID modes and light models for plain questions.

06Screens

07System architecture

  1. L1Surfaces
    • Home orb
    • Chat
    • Quick Summon
    • Tasks
    • Memory
    • Settings
  2. L2Core
    • ConversationEngine
    • LlmRouter
    • VoiceTurnController
    • ToolGateway
  3. L3Safety
    • Permission layer
    • On-screen confirm
    • Redactor
    • Agent Trace (hash chain)
  4. L4Providers
    • Gemini
    • OpenAI-compatible
    • Android STT / TTS
    • CameraX
  5. L5Data
    • Room
    • DataStore
    • Keystore-encrypted keys
  6. L6Desktop
    • NOVA Agent (Windows)
    • TLS 1.3 pinned
    • Kill switch
UI → ViewModel → Engine / use case → repository or provider interface → data source. Domain logic (ConversationEngine, LlmRouter, VoiceTurnController) is pure Kotlin.Diagram — derived from the source code

08Engineering decisions

  1. Keys are write-only

    API keys are encrypted with the Android Keystore, never shown again, sent only in headers, never logged; cleartext is allowed only to private-network hosts, enforced in code.

  2. Dialect accuracy measured, not promised

    Rather than switching the recognizer to a regional tag that might break what works, an evaluation harness (WER/CER over recorded sentences) decides.

  3. Honesty over appearance

    An ADR documents every place where 0.1 deviated from the designs instead of faking a feature that was not ready.

09Technical challenges

  1. Interrupting a speaking assistant

    The platform recognizer is endpointed, not streaming. Barge-in runs a separate voice-activity detector with echo cancellation; playback is cancelled first, then the model call, and the stored message keeps only the heard text — cut at the word being spoken.

10Lessons learned

The safest agent is a deterministic system with a model inside, not a model with tools attached.

My contribution

Architecture, 40+ architecture decision records, Android app and desktop agent — 312 Kotlin files, 77 commits.

Technology

  • Kotlin
  • Jetpack Compose
  • Navigation 3
  • Room
  • DataStore
  • Android Keystore
  • Gemini
  • OpenAI-compatible
  • Ktor
  • CameraX
  • WorkManager
  • Coroutines
  • Flow
  • Robolectric
  • TLS 1.3
  • AI Agents

How this was verified

Source, README and ADRs reviewed. Images are design previews kept in the project's docs/design folder, not captures of a running build.

Related projects

Next project

Vellora