J.A.R.V.I.S.

Just A Rather Very Intelligent System

A production personal AI assistant modeled after the MCU's JARVIS. 120,000+ lines of code across 5 platforms.

Built for one user. Personalized, autonomous, always running.

Python FastAPI Swift / SwiftUI React TypeScript Redis PostgreSQL Qdrant Claude API Gemini API Cerebras ElevenLabs

View Architecture Try the Demo

Overview

The Vision

JARVIS is my attempt at building the real thing: a personal AI that knows me deeply, anticipates what I need, and gets smarter every day. It's not a chatbot wrapper. It's a continuously-running intelligence with its own goals, time awareness, and the ability to act on my behalf across every device I own.

System Metrics

By the Numbers

120K+

Lines of Code

Active AI Tools

Input Channels

Architecture Layers

LLM Providers

Platforms

DB Tables (AES-256)

Autonomous Goals

Technical Design

Architecture Overview

JARVIS System Architecture — Online

Clients

Web Interface iOS App Apple Watch CLI iMessage Phone Calls Wake Word

FastAPI Backend (Railway)

68 AI Tools Event Bus Goal System Deferred Actions Owner Intelligence Self-Healing

AI Layer

Claude Sonnet 4.6 — primary reasoning Gemini 2.5 Flash — backup Cerebras — fast routing, background tasks Local LLM on Mac Mini — offline fallback

Data Layer

PostgreSQL — encrypted messages, user data Redis — real-time state, caching, queues Qdrant — vector search, knowledge base

Hardware Layer

Mac Mini M4 — TTS, local LLM, camera, wake word MacBook — Apple account data bridge

Capabilities

Feature Deep Dives

Module 01

Conversational Intelligence

Multi-Provider AI with Failover

JARVIS routes across Claude, Gemini, Cerebras, and a local Mac Mini LLM. Cerebras handles fast routing and background tasks where latency matters. If any provider goes down, automatic failover kicks in. If all providers fail, it returns a graceful response without crashing.

68 AI Tools

Weather, calendar, email, reminders, navigation, web search, health data, music control, smart home, timers, contacts, file access, and more. JARVIS selects and chains tools automatically based on the request.

Dynamic Context Injection

JARVIS only includes relevant context in each request. Health query? It injects health data. Schedule question? It injects calendar. This keeps responses fast and focused without unnecessary token overhead.

Module 02

7 Ways to Talk to J.A.R.V.I.S.

Web Interface

MCU-inspired HUD with real-time streaming, tool execution visualization, and inline widgets.

iOS App

Native SwiftUI with SSE streaming, HealthKit integration, MapKit navigation, and passkey auth.

Apple Watch

Energy orb interface, quick voice commands, and live health data push.

Command Line

Installable CLI package with a rich terminal UI for developers.

iMessage

Text JARVIS like any other contact. Responses arrive in under 5 seconds.

Phone Calls

Call JARVIS's dedicated number. Voice synthesis powered by ElevenLabs.

Wake Word

"Hey JARVIS" on the Mac Mini. Always listening, always ready.

Module 03

Autonomous Intelligence

Event Bus

Real-time event processing. When health data arrives, location changes, or a tool fails, JARVIS processes it immediately rather than on a scheduled poll.

Autonomous Goals

Three persistent objectives running 24/7: keep me healthy (monitors heart rate, sleep, activity), keep me on schedule (watches calendar, alerts for conflicts), and maintain system integrity (tracks tool reliability, auto-heals).

Time Awareness

Timers, scheduled actions, recurring checks, follow-ups. "Start a 5-minute timer" just works. "Check the stock price at 4 PM and tell me" just works.

Self-Healing

JARVIS monitors his own deployment. When the backend crashes, he detects the error, dispatches an AI coding agent to fix it, commits the fix, and redeploys. Then sends a notification via iMessage. A circuit breaker prevents infinite loops.

Module 04

Deep Personalization

Owner Intelligence

JARVIS builds a comprehensive profile from health data, calendar patterns, location history, communication patterns, music taste, active projects, and conversation themes. Rebuilt daily, learned from continuously.

Cross-Domain Reasoning

"What do you know about me?" triggers 6 parallel queries across health, calendar, tasks, knowledge base, location, and music, synthesized into one comprehensive answer.

Capability Gap Detection

When JARVIS can't fulfill a request, it logs the gap. Over time, it builds a backlog of capabilities it knows it needs, forming the foundation for runtime tool creation.

Module 05

Security & Reliability

AES-256 Encryption

Every message and contact is encrypted at rest with per-user PBKDF2-derived keys.

Passkey Authentication

WebAuthn/FIDO2 with iCloud Keychain sync. No passwords.

Graceful Degradation

Every external service has timeout protection and fallback behavior. If Redis goes down, JARVIS still responds. If the Mac Mini is offline, affected tools return clear messages instead of crashing.

Module 06

Infrastructure

Mac Mini M4 Server

Runs voice synthesis, local LLM, security camera interface, wake word listener, and iMessage relay.

MacBook Bridge

Serves Apple account data via a dedicated agent with a Cloudflare tunnel.

Automated Testing

12-scenario smoke test suite runs after every deployment, verifying tools, latency, and graceful degradation across all channels.