Coach Artie

Coach Artie

AI assistant that meets you where you are - Discord, SMS, or web - with full context awareness.

Philosophy

AI assistants work better when they have context. Different situations call for different interfaces. Same brain, multiple access points.

Core principles:

• Not trying to replace human interaction
• No illusion of personality - it's a tool
• Privacy-first: your data, your server, your control
• Built for longevity, not novelty

The Three Interfaces

Interface	Context	Why
Discord	Already hanging out with friends	Natural async conversation, group context, bot sees history
SMS	Away from computer, quick questions	Works everywhere, weather + health integration, mobile-first
Web UI	Deep work, configuration	Visual interface, history browsing, settings management

Discord (Most Used)

Where conversations already happen. Can @mention the bot when needed. Group conversations where bot participates. Natural async flow.

Trade-off: Not private, Discord-dependent.

SMS (Emergency/Mobile)

Quick questions while traveling. Weather advice considers if you've been sedentary. "It's gonna be a nice day and you're due for a walk."

Trade-off: Character limits, costs money.

Web UI (Configuration)

Visual interface for complex interactions. History browsing and search. Less used but important for setup.

Trade-off: Least used, most complex.

Why Context Matters

The bot knows:

• GitHub activity patterns
• Health data from Apple Watch
• Current location and weather
• Calendar and schedule

Makes connections humans would miss: "You haven't committed in 3 days, everything ok?" Weather advice considers sedentary patterns.

Architecture

Built on TypeScript monorepo:

• pnpm + turborepo for workspace management
• Capabilities system - modular features that can be enabled/disabled
• Redis for queueing (prevents duplicate work)
• SQLite for persistent state
• OpenRouter for LLM access (model flexibility)

Production Deployment

• VPS hosting (not cloud - total control)
• PM2 process management
• Cloudflare Tunnel for HTTPS
• Health monitoring at dedicated port
• 512MB Node heap, 512MB Redis

Current Capabilities

Capability	Description
Chat	Conversational AI (obviously)
Memory	Semantic recall of past conversations
Calculator	Wolfram Alpha integration
Email	Send emails on request
Scheduling	Meeting coordination
GitHub	Webhook integration for activity

What I Learned Building This

Interface > Model

The right interface matters more than model quality. SMS at the right moment beats GPT-4 in a browser tab you forgot about.

Context > Intelligence

A dumber model with your context beats a smarter model without it. Personal data makes generic AI feel personal.

Modular > Monolithic

Capabilities system means features can be added/removed without touching core. Redis queues prevent duplicate work. Health checks catch issues before they matter.

The Multi-Interface Pattern

This is the same pattern as Scrapbook-core (Alfred, web, CLI) and Connectology (web, desktop). Same brain, different access patterns:

• Desktop ≠ mobile ≠ terminal
• Browsing ≠ searching ≠ querying
• Each interface optimizes for its context

Related

• Personal APIs - data sources that feed context
• Quantified Self - tracking that enables context
• Same Data Many Interfaces - the design pattern
• Building Tools for Yourself - why personal tools work