Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Windows AI Assistant

JARVIS is the internal project codename. This independent project is not affiliated with or endorsed by Marvel or Disney.

A desktop assistant that connects natural-language interaction to typed, locally authorized task operations. Built with Rust, Tauri 2, React, TypeScript and SQLite, with Gemini used for interpretation rather than operational authority. The interface and sample prompts are in Brazilian Portuguese.

Local portfolio candidate based on M3.1 (902a76c). Publication is pending. Project-owned code and materials are licensed under MIT; third-party material retains its own terms.

Interface

Native Windows WebView captures from an isolated test build; system figures are simulated. No AI provider was called. A synthetic task was created using the Central UI; separate backend tests confirmed persistence after restart. No provider-generated action is shown; see validation.

Central task panel showing a successful synthetic task creation

Floating assistant window

Engineering focus

  • Separate interpretation from execution. Structured model responses are validated locally. The model cannot approve writes, run a shell, or confirm effects on its own.
  • Keep conversation coherent across windows. The Central and floating assistant share a canonical Rust session with revisions, turn ownership and bounded temporary context.
  • Make task execution inspectable. Conversational writes require review and scoped approval. Delegated complete_task operations use durable plans, attempts and receipts; restart requires explicit review and consent.
  • Evaluate boundaries, not only happy paths. Tests cover stale revisions, ambiguity, authorization, interruptions and recovery. Real-provider evidence is reported separately from fixtures.

Demonstrated capabilities

Capability Boundary
General conversation and corrections across windows Bounded temporary context; no persistent personal memory
Create, query, update, complete and delete local tasks Conversational writes pass local authorization; manual controls also exist
Sequential completion of selected tasks Only complete_task is delegable; no automatic restart or rollback of confirmed effects
Weather lookup and opening a Google search Opening results does not mean reading or researching them
Offline task management Model interpretation and weather require network access

Architecture · Validation · Synthetic demo script · Rights and attribution

Run the frontend preview

Requires Node.js 22.12+ and npm. From this directory:

npm ci
npm test
npm run build
npm run dev

Open http://127.0.0.1:1420, or http://127.0.0.1:1420/?window=assistant for the floating layout. This preview does not invoke desktop commands, call Gemini or persist tasks. It is not a native integration test.

Run the desktop app on Windows 11

Use a native Windows checkout with Node.js 22.12+, stable Rust MSVC, Microsoft C++ Build Tools with Desktop development with C++, Windows SDK and WebView2. See the Tauri prerequisites. Keep Windows and Linux dependencies/build directories separate.

npm ci
npm test
npm run build
cargo test --manifest-path src-tauri/Cargo.toml --locked --lib
npm run tauri dev

For native compilation without an installer: npm run tauri build -- --no-bundle. Installer packaging is disabled. Do not point demonstration runs at an existing personal installation: the unchanged application identifier is com.jarvis.desktop.poc, and its real database is %LOCALAPPDATA%\com.jarvis.desktop.poc\tasks.sqlite3. Use a separate Windows user/profile with synthetic data. This preparation does not launch the application.

Gemini interpretation requires GEMINI_API_KEY supplied privately to the Rust process environment. Never put it in frontend variables, source, screenshots or logs. Without it, use manual task controls; remote interpretation reports an error. No credential is needed for unit tests or the web preview. Supplying a key and sending a message can incur provider charges; this repository does not grant a free-tier or availability guarantee.

Validation and limitations

Historical M3.1 evidence records a 403-test TypeScript baseline, later targeted revalidation (112 executions, not additive unique coverage), 114 native Windows Rust tests, and 21 aggregate real Gemini calls. These are different boundaries and configurations, not a single end-to-end score. See validation and acceptance provenance for caveats and this preparation's checks.

The product uses a 25-second remote timeout, at least six seconds between calls and a limit of 20 calls per process. Evaluation-only builds used a 60-second timeout; that change is not in this candidate. Latency and semantic reliability remain variable. No voice, arbitrary computer automation, source-reading research, persistent personal memory, or operational OpenAI integration is claimed. There is no production-readiness claim.

M3.2 is excluded from this source selection. Its separate memory experiments documented 72.7% capture coverage on an adjusted development corpus versus 8.3% on a blind holdout; automatic capture stayed blocked. This is evidence of a rejected generalization hypothesis, not a delivered capability.

Project-owned code and materials are distributed under the MIT license. Third-party components and data remain subject to their own licenses and terms. See rights and attribution.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages