Skip to content
Systems & CorePrototype

Project J.A.R.V.I.S. — Desktop AI Orchestrator & Delegation Engine

Role
Lead Architect & Systems Engineer
Dates
November 2026
Stack
Tauri v2 · Rust · Next.js 15 · TypeScript · Python 3.12 · FastAPI · LangGraph · Docker

01.Executive Summary

Project J.A.R.V.I.S. is a zero-latency, secure desktop AI orchestration hub engineered for personal executive operations, smart home automation, and autonomous multi-agent tool delegation. Operating via a native Windows 11 Tauri v2 Rust container and a local FastAPI/LangGraph daemon, it coordinates 26 MCP server tools behind a strict 4-tier security authorization gate.

02.Problem & Constraints

Cloud-based AI assistants lack native OS integration, introduce unacceptable latency for local system management, pose severe security risks when granted unrestricted tool execution, and risk leaking sensitive API credentials and private emails during autonomous multi-agent workflows.

03.System Architecture

Architected with a dual-runtime separation of concerns. A native Tauri v2 desktop application written in Rust provides hardware access and a Next.js 15 HUD dashboard. This shell communicates over authenticated local IPC and port 8420 with a background Python 3.12 daemon hosting FastAPI and LangGraph multi-agent graphs. A deterministic Hydrogen Router triages intent in <80ms, enforcing an L0–L3 security hierarchy with regex secret redaction before delegating tasks across 26 MCP servers.

04.Engineering Decisions & Tradeoffs

Prioritized a local daemon with explicit IPC gating over single-binary bundling to allow independent hot-reloading of agent skill graphs and isolated sandboxing of external tool credentials.

Decision 1: Local FastAPI & LangGraph Daemon vs Cloud Agent Host

Option chosen

Local background Python 3.12 daemon on port 8420 with authenticated localhost tokens.

Alternative rejected

Cloud-hosted LangGraph Cloud or AWS Lambda container.

Cost accepted

Requires local Python 3.12 environment and ~300MB idle RAM overhead.

Measured impact

Zero cloud latency for local OS actions, complete data sovereignty, and offline operational resiliency.

Decision 2: 4-Tier Security Matrix (L0-L3) vs Unrestricted Tool Execution

Option chosen

Deterministic permission policy gate: L0 (read) runs automatically, L1/L2 (workspace edits) log audit trails, L3 (system/destructive/financial) requires explicit user approval.

Alternative rejected

Autonomous agent with blanket execution privileges.

Cost accepted

User confirmation prompt required for destructive filesystem or outbound communication tasks.

Measured impact

Zero unauthorized file deletions or unintended actions across 1,000+ benchmarked operational delegation cycles.

Decision 3: Deterministic Secret Redaction Filter vs Post-Hoc LLM Censorship

Option chosen

Pre-flight and post-flight regex pattern scanner intercepting API keys, bearer tokens, and private credentials.

Alternative rejected

Prompting the LLM to redact sensitive credentials in its responses.

Cost accepted

Sub-millisecond regex processing overhead on all tool invocation inputs and outputs.

Measured impact

100% interception recall on test credential injection attacks with zero credential leakage to external LLM providers.

05.Evaluation & Verified Metrics

Key performance indicators and latency profiles benchmarked under explicit workload and hardware constraints.

Intent Triage Latency

< 80 ms

Local Hydrogen Router classification

Measured on: Windows 11 desktop, Python 3.12, FastAPI, Tauri v2 Rust core, AMD Ryzen 9 7950X.

Secret Redaction

100.0%

Deterministic regex credential interception

Measured on: Windows 11 desktop, Python 3.12, FastAPI, Tauri v2 Rust core, AMD Ryzen 9 7950X.

MCP Tool Connectors

26 Servers

Local and cloud protocol integrations

Security Policy Tiers

4 Tiers

L0 (auto) to L3 (explicit authorization)

06.Personal Contribution

Sole creator and lead architect. Developed the Tauri v2 Rust desktop bridge, the Next.js 15 HUD interface, the FastAPI/LangGraph daemon, the 4-tier security gate, and all 26 MCP tool integration connectors.

07.Limitations & Next Steps

Known limitations:

Requires local background daemon execution on the host machine; multi-device synchronization is currently constrained to local network LAN.

Next engineering steps:

Adding on-device local SLM (Phi-3 / Gemma 2) fallback inference for full offline operation when internet connectivity drops.