How Modern AI Systems Work
A complete guide to building your own intelligent automation
Start Here
The big picture — before any tools or jargon
The Old Way vs The New Way
The Old Way
- You do everything manually
- You forget things constantly
- You run out of time
- Information lives in scattered places
- Nothing talks to anything else
The New Way
- Your system works while you sleep
- Nothing is ever forgotten
- Tasks happen automatically
- All information is connected
- Everything integrates seamlessly
What This Guide Covers
By the end of this guide, you'll understand every piece that makes up a modern AI system. You'll know:
- What each tool does — and why it exists
- How they work together — the connections between components
- How to build your own — step-by-step from zero to fully operational
This is not theory. Every example in this guide comes from a real, working system managing 11 clients, $657K in revenue, 32 contacts, and thousands of automated tasks per month.
What's Possible
Here's what you can actually build with this knowledge:
- Email that triages itself — categorized, summarized, and responded to automatically
- A database that never forgets — every person you meet, every conversation, instantly recalled
- Reports that write themselves — financial summaries, client updates, research briefs
- Intelligence on demand — send a voice note, get a 12-minute research report
- Work that happens while you sleep — invoices sent, reminders triggered, data synced
Real Example: Morning Intelligence
At 6:00 AM, the system automatically:
- Pulls financial data from accounting software
- Checks project status across all clients
- Scans email for urgent messages
- Generates a formatted intelligence report
- Delivers it via email and phone
Total time: 5 minutes. Human involvement: zero.
Ready? Let's start with the brain — the intelligence that makes it all work.
The Brain: Claude & Claude Code
The intelligence at the center of everything
What is Claude?
Claude is an AI assistant made by Anthropic. Think of it like ChatGPT, but with some key differences:
- It reasons through complex problems — not just pattern matching
- It writes and understands code — can build entire applications
- It researches deeply — can analyze documents, cross-reference sources
- It makes decisions — can evaluate options and recommend actions
Claude vs Claude Code: The Critical Difference
Regular Claude
Can do:
- Chat and answer questions
- Write content and code
- Analyze text and images
Cannot do:
- Access your files
- Run commands on your computer
- Connect to external services
- Actually execute tasks
Claude Code
Can do everything Claude can, PLUS:
- Read files on your computer
- Run terminal commands and scripts
- Connect to Gmail, Calendar, databases
- Build and deploy websites
- Execute multi-step workflows
- Run for hours without supervision
The difference matters. Regular Claude is like having a brilliant consultant who can only talk. Claude Code is like having a brilliant consultant who can actually DO the work.
Skills: Saved Instructions Claude Follows
Instead of typing the same long prompt every time, you save it as a "skill" — like a recipe Claude follows.
Real Skills in Action
/morning — Runs a complete morning briefing:
- Reads STATE.md for current system status
- Checks queue depth in Redis
- Scans recent activity logs
- Formats a 60-second briefing
- Delivers via Telegram
/research [topic] — Deep research workflow:
- Searches web with Perplexity
- Scrapes relevant pages with Firecrawl
- Cross-references multiple sources
- Writes structured report
- Saves to Obsidian vault
/client-report [name] — Generates client summary:
- Queries people database for contact
- Pulls recent interactions from Gmail
- Checks project status from Airtable
- Formats professional report
- Deploys to Cloudflare Pages
How to Create a Skill
Skills are just markdown files with instructions. Here's the structure:
That's it. Save the file, and typing /morning runs the entire workflow.
Real Scenario: Voice Note to Research Report
11:30 AM — User sends voice note while walking: "Research Newstead Aesthetics, need to know their tech stack, team size, and growth trajectory"
11:31 AM — Whisper transcribes the audio
11:31 AM — Claude Code executes:
- Searches web for "Newstead Aesthetics"
- Scrapes their website for tech indicators
- Searches LinkedIn for team profiles
- Analyzes Instagram engagement patterns
- Cross-references business registry data
- Writes 1,200-word intelligence report
- Saves to Obsidian vault with tags
- Sends summary via Telegram
11:42 AM — User receives complete report. Total time: 12 minutes.
Next, we'll look at how the system remembers everything — forever.
The Memory: How It Remembers Everything
The most important concept — permanent memory that never forgets
The Memory Problem
Normal AI forgets everything when you close the chat. That's completely useless for a real system.
Imagine if every morning you forgot:
- Everyone you've ever met
- Every project you're working on
- Every decision you've made
- Everything you learned yesterday
That's how regular AI works. To build something useful, we need permanent memory.
The 4-Layer Memory System
Think of memory like a building with 4 floors. Each floor serves a different purpose:
Floor 4: Right Now (Redis)
What it is: Redis is an in-memory database — extremely fast storage for temporary data.
What it stores:
- Current task queue (urgent, normal, background priorities)
- Active jobs being processed right now
- Session data for ongoing conversations
- Cache for frequently accessed information
How long it lasts: Minutes to hours. Resets when the system restarts.
Real example: When you send a message via Telegram, it hits Redis first. The queue manager pulls jobs from Redis and assigns them to workers.
Floor 3: This Week (STATE.md + Obsidian)
What it is: Plain text files synced across all your devices.
What it stores:
- Current project status (what's active, what's blocked)
- Recent deployments and changes
- This week's focus and priorities
- Quick reference for "what was I working on?"
How long it lasts: Weeks to months. Updated every session.
Real example: STATE.md contains lines like "LOVR Collections Plan DEPLOYED ✅" — updated 20 minutes ago, synced to phone, accessible everywhere.
Floor 2: Long Term (PostgreSQL + People Database)
What it is: PostgreSQL is a powerful, permanent database. Rock solid, industry standard.
What it stores:
- 32 contacts with full profiles (role, company, last contact, relationship strength)
- Every interaction logged (emails, meetings, calls)
- Financial records ($657K revenue tracked)
- Project history across 11 clients
How long it lasts: Forever. This is permanent storage.
Real example: Query "Who haven't I spoken to in 30 days?" returns a list with exact dates of last contact.
Floor 1: Deep Memory (Semantic Search)
What it is: AI-powered memory that understands meaning, not just keywords.
The difference:
- Keyword search: "Find emails containing 'invoice'" — misses "payment request", "bill", "statement"
- Semantic search: "Find emails about money owed" — finds ALL variations, understands context
Real example: Ask "What do I know about Jack?" and it searches across all 4 memory layers, understanding that "Jack Brinsmead", "BP Wharton owner", and "the guy from the construction job" all refer to the same person.
How a Query Moves Through Memory
Question: "What's the status on the LOVR project?"
- Floor 4 (Redis): Check if there's an active task running right now → Yes, email workflow scheduled for tomorrow 6am
- Floor 3 (STATE.md): Check recent updates → "LOVR BI DEPLOYED ✅ — Workflow active, first sync tomorrow 6am"
- Floor 2 (PostgreSQL): Query client record → 11 active clients, $657K total value, last update 2 hours ago
- Floor 1 (Semantic): Pull related context → Financial crisis, collections plan, Xero integration, Metabase dashboard
Result: Complete answer in 2 seconds, pulling from all 4 layers automatically.
This is why the system never forgets. Every piece of information lives somewhere permanent. Even if Redis clears, even if you close the app, the knowledge persists.
Next: The 12 tools that give the AI actual superpowers.
The Tools It Connects To
12 tools. One brain controlling all of them.
What Are MCP Servers?
MCP stands for Model Context Protocol. Think of them as plugins that give Claude the ability to USE external tools.
- Without MCP: Claude can only talk to you
- With MCP: Claude can read your email, search the web, update databases, manage your calendar
MCP is like USB-C for AI — a standard way to connect anything to everything.
The magic: Once an MCP server is connected, Claude can use that tool as naturally as you do. It doesn't just suggest what to do — it actually does it.
The 12 MCP Servers
📧 Gmail
What it is: Your email, automated.
What Claude can do:
- Read every email in your inbox
- Categorize messages automatically
- Draft professional replies
- Send emails on your behalf
- Search for specific conversations
Why it matters: "Triage my inbox" becomes a 60-second job instead of an hour.
📅 Google Calendar
What it is: Your schedule, managed.
What Claude can do:
- See what's coming up today
- Create events automatically
- Check availability before booking
- Block focus time
- Get meeting prep briefs
Why it matters: Ask "Am I free Thursday at 2pm?" and get an instant answer with context.
📁 Google Drive
What it is: Your files, accessible.
What Claude can do:
- Read documents and extract data
- Create spreadsheets with formulas
- Manage folders and sharing
- Work with Docs, Sheets, and Slides
Why it matters: "Summarize the Q4 report" works on 50-page Google Docs.
📊 Airtable
What it is: A spreadsheet that thinks.
What Claude can do:
- Read and write to your CRM
- Update project trackers
- Manage content pipelines
- Query databases like asking questions
Why it matters: Your database becomes conversational. "Add this person to the prospects list" just works.
📝 Notion
What it is: Your knowledge base.
What Claude can do:
- Read and write pages
- Search across entire workspace
- Create databases and update records
- Manage project documentation
Why it matters: "Find that note about the client meeting" searches thousands of pages instantly.
💬 Slack
What it is: Your team, reachable.
What Claude can do:
- Read channel history
- Search conversations
- Find information buried in threads
- Summarize discussions
Why it matters: "What did the team decide about pricing?" pulls answers from 6 months of messages.
🗒️ Obsidian
What it is: Your personal notes vault.
What Claude can do:
- Read and write your notes
- Search years of documentation
- Create linked knowledge graphs
- Update daily logs automatically
Why it matters: 597MB of notes instantly searchable and updateable.
🔍 Perplexity
What it is: AI-powered web search.
What Claude can do:
- Search the web with understanding
- Get citations for every fact
- Find recent news and updates
- Research topics deeply
Why it matters: Like Google, but it actually understands what you're asking.
🌐 Firecrawl
What it is: Read any website.
What Claude can do:
- Turn any webpage into clean text
- Scrape entire websites
- Extract specific data from URLs
- Monitor competitor sites
Why it matters: "Analyze their pricing page" works on any competitor's site.
🌍 Bright Data
What it is: Access anywhere.
What Claude can do:
- Access geo-restricted content
- Scrape sites that block normal tools
- Enterprise-grade data collection
Why it matters: Reaches websites that block standard scrapers.
📱 Apify
What it is: Social media intelligence.
What Claude can do:
- Scrape Instagram profiles
- Monitor TikTok trends
- Extract YouTube data
- Access 3,000+ pre-built scrapers
Why it matters: "Check their Instagram engagement" returns actual metrics.
⚡ Obsidian Local API
What it is: Live vault access.
What Claude can do:
- Real-time read/write while vault is open
- Instant updates to notes
- Live search without delays
Why it matters: Changes appear in Obsidian immediately as they're made.
Real Example: Client Research Workflow
Command: "Research Newstead Aesthetics and create a client profile"
What happens:
- Perplexity: Searches web for company info
- Firecrawl: Scrapes their website for tech stack clues
- Apify: Pulls Instagram engagement metrics
- Gmail: Searches for any past emails
- Airtable: Creates new client record with all data
- Obsidian: Saves full intelligence report to vault
Result: Complete client profile in 12 minutes. All tools working together.
Next: The automation engine that runs while you sleep.
The Automation Engine: n8n
The invisible workforce that runs while you sleep
What is n8n?
n8n (pronounced "n-eight-n") is workflow automation software. Think of it like hiring workers who do the same tasks perfectly, every time, on schedule.
Each "workflow" is a sequence of steps that runs automatically:
n8n vs Zapier/Make
Zapier / Make
How it works:
- Cloud-based (runs on their servers)
- Pay per task executed
- Limited control over data
- Easy to start, expensive to scale
Cost example: 10,000 tasks/month = $150-300/month
n8n
How it works:
- Runs on YOUR computer (self-hosted)
- Unlimited tasks
- You own all the data
- Free forever (or $30/month cloud option)
Cost example: 10,000 tasks/month = $0 (self-hosted)
Real Workflows Running Right Now
Morning Intelligence — runs 6:00 AM daily
Trigger: Every day at 6:00 AM AEST
Steps:
- Fetch financial data from Xero (revenue, expenses, bank balance)
- Pull project status from Airtable (active clients, overdue tasks)
- Check Gmail for urgent overnight messages
- Query PostgreSQL for yesterday's metrics
- Format HTML intelligence report
- Email summary to primary inbox
- Send brief to Telegram
Output: Complete daily briefing delivered before waking up
Value: Saves 20 minutes every morning. Never miss critical updates.
Invoice Chase — checks daily
Trigger: Every day at 9:00 AM
Steps:
- Query Xero for all outstanding invoices
- Filter by days overdue (7, 14, 30, 60+ categories)
- Check if reminder already sent this week (avoid spam)
- Generate personalized reminder email for each
- Send via Gmail with appropriate tone (friendly → firm)
- Log activity in Airtable
- Notify via Telegram if any 60+ day invoices
Output: Automated collections system
Value: Improved cash flow from 4.7% to projected 85% collection rate. $310K annual impact.
Lead Alert — instant notification
Trigger: New lead form submission (webhook)
Steps:
- Receive lead data via webhook
- Create record in Airtable CRM
- Check if lead already exists (dedupe)
- Enrich with Perplexity research (company size, tech stack)
- Calculate lead score based on criteria
- Send formatted alert to Telegram
- Add to follow-up sequence
Output: Instant notification with context
Value: Never miss a hot lead. Response within 5 minutes.
Content Calendar — daily sync
Trigger: Every day at 8:00 AM
Steps:
- Fetch content schedule from Airtable
- Check which posts are due today
- Verify assets are ready (images, captions)
- Alert if anything is overdue
- Update Google Calendar with posting schedule
- Send daily plan to Telegram
Output: Daily content schedule synchronized
Value: Team always knows what's due. Nothing falls through cracks.
Weekly Report — every Monday 8:00 AM
Trigger: Monday at 8:00 AM
Steps:
- Pull metrics from last 7 days (revenue, client activity, tasks completed)
- Compare to previous week (percentage change)
- Generate charts with Google Sheets
- Format professional HTML report
- Email to stakeholders
- Save PDF to Google Drive
Output: Weekly performance summary
Value: Stakeholders stay informed without manual reporting.
Building a Workflow
n8n has a visual interface — you drag boxes and connect them with arrows. No coding required for basic workflows.
The key insight: Once you build a workflow, it runs forever. Zero additional cost. Zero additional effort. The system just works.
Next: How you control all of this from your phone.
The Messenger: Telegram Bot
Control your entire system from one messaging app
What is a Telegram Bot?
Telegram is a messaging app (like WhatsApp). You can create "bots" — automated accounts that respond to messages.
A Telegram bot is your personal AI assistant that lives in your messaging app.
Why Telegram Instead of a Web App?
- Always in your pocket — Phone, tablet, desktop, all synced
- Voice notes work perfectly — Speak your request while driving, walking, at a BBQ
- Push notifications — System can alert you proactively
- No login required — Already authenticated, instant access
- Media support — Send screenshots, PDFs, images for analysis
How It Works: Phone to System
The 6 Intent Types
When a message arrives, Claude Haiku (fast, cheap) classifies it into one of 6 categories:
QUICK_LOOKUP
Simple questions with fast answers.
Examples:
- "What's my calendar look like today?"
- "How many clients are active?"
- "Check the queue depth"
Priority: Normal (30-60 seconds)
SKILL_EXECUTE
Running a saved skill/workflow.
Examples:
- "/morning" — Daily briefing
- "/status" — System health
- "/sync" — Sync to Mini 2
Priority: Urgent (15-30 seconds)
DEEP_RESEARCH
Multi-step research workflows.
Examples:
- "Research Newstead Aesthetics"
- "Find competitors in Brisbane"
- "Analyze their tech stack"
Priority: Background (5-15 minutes)
MEMORY_QUERY
Searching across all memory layers.
Examples:
- "Who haven't I spoken to in 30 days?"
- "What's the status on LOVR?"
- "Find that email about invoices"
Priority: Normal (1-2 minutes)
SYSTEM_CMD
System diagnostics and control.
Examples:
- "/status" — Full health check
- "Check Mini 2 connectivity"
- "What's running on port 5678?"
Priority: Urgent (immediate)
CHAT
General conversation and questions.
Examples:
- "Explain how Redis works"
- "Draft an email to Jack"
- "What should I prioritize today?"
Priority: Normal (30-90 seconds)
Voice Notes: The Killer Feature
Send a voice note while:
- Driving to a meeting
- Walking the dog
- At a BBQ talking to a prospect
- Lying in bed before sleep
What happens:
- Whisper (OpenAI) transcribes audio to text
- Text processed exactly like a typed message
- Full Claude Code execution with all tools
- Formatted response delivered back
Real Scenario: Voice Note Research
11:30 AM — Walking between meetings, send voice note:
"Hey, I just met a guy named Nils Barry at E&S Residential, he does property stuff in Brisbane, need to know if he's worth following up with, check his company size, recent projects, and what their tech situation looks like"
11:31 AM — Transcribed and classified as DEEP_RESEARCH
11:31-11:42 AM — System works:
- Searches "E&S Residential Brisbane" with Perplexity
- Scrapes company website with Firecrawl
- Searches LinkedIn for team profiles
- Checks business registry for company details
- Analyzes Instagram for brand presence
- Writes 1,200-word intelligence report
- Saves to Obsidian vault with tags
- Creates contact record in PostgreSQL
11:42 AM — Notification arrives:
"✅ Nils Barry research complete. E&S Residential: 12-person team, $4.2M recent project pipeline, currently using basic Wix site + manual workflows. High-value target. Full report saved to Obsidian → Clients/Nils-Barry/"
Total time: 12 minutes. You were walking the entire time.
This changes everything. Your most valuable ideas come when you're away from the computer. Now you can capture and act on them immediately.
Next: The infrastructure that keeps everything running.
The Infrastructure: Docker + Servers
The engine room — what keeps everything running
What is Docker?
Docker packages software into "containers" — like shipping containers for code. Each container is:
- Isolated — Runs independently without affecting other containers
- Consistent — Works the same on any computer
- Easy to restart — If something breaks, restart the container, not the whole computer
The analogy: Instead of installing PostgreSQL directly on your Mac (messy, hard to uninstall), you run it in a Docker container. To restart PostgreSQL, you just restart the container. Clean and simple.
The Container Stack
These containers run 24/7 on the two Mac Minis:
Mini 1 Containers (8 running)
- postgres — The main database (people, interactions, financial data)
- redis — The queue manager (task priorities, session data)
- n8n — The automation engine (workflows, schedules)
- n8n_worker — Background worker for long-running workflows
- mem0-api — AI memory service (semantic search)
- obsidian-couchdb — Vault sync server (syncs notes to all devices)
- grafana — Trading dashboard (market monitoring)
- qdrant — Vector database (semantic memory search)
Mini 2 Containers (3 running)
- lovr-n8n — LOVR's automation engine
- lovr-postgres — LOVR's database
- lovr-metabase — LOVR's business intelligence dashboard
Why separate Minis? Mini 1 is Nathan's personal system. Mini 2 is dedicated to LOVR client work. Clean separation — client data never mixes with personal data.
LaunchAgents: Background Services
LaunchAgents are like apps that run invisibly in the background. They:
- Start automatically when the Mac boots
- Restart themselves if they crash
- Run without requiring someone to be logged in
Mini 1 Services
- com.jarvis.agent2 — Telegram bot (PID 61195, running)
- com.jarvis.cloudflare-tunnel — Expose services to internet
- com.jarvis.ollama — Local AI models
- com.jarvis.orchestrator — Main AI loop (port 3100)
- com.jarvis.hourly-sync — Auto-sync to Mini 2
Mini 2 Services
- com.lovr.cloudflare-tunnel — LOVR's public URLs
- com.lovr.metabase — Dashboard service
- com.lovr.ollama — LOVR's local AI
- com.lovr.fleet-monitor — System monitoring
- com.lovr.control-dashboard — Admin interface
- com.lovr.licence-server — Software licensing
What Happens When Something Breaks?
Real example: PostgreSQL container stops responding.
- Detection: Next query fails, returns connection error
- Alert: System sends Telegram notification "PostgreSQL down"
- Auto-restart: Docker automatically restarts the container
- Recovery: Container comes back up in 5 seconds
- Resume: Queued jobs continue automatically
Total downtime: 5-10 seconds. No data lost. No manual intervention.
Real Scenario: Docker Restart After Mac Update
Problem: Mac restarts after system update. All Docker containers stop.
What happens:
- Mac boots up
- Docker Desktop starts automatically
- All containers with
restart: alwayspolicy start - LaunchAgents detect services are up
- Cloudflare tunnels reconnect
- Agent 2 connects to Telegram
- System operational in 60 seconds
Manual intervention required: Zero. System recovers on its own.
Next: How all these pieces stay connected across the internet.
The Network: Tailscale + Cloudflare
How everything stays connected and accessible
The Problem Networks Solve
- Mini 1 is in one location
- Mini 2 is in another location
- MacBook Pro travels everywhere
- iPhone is always with you
They all need to talk to each other. Securely. From anywhere.
Tailscale: Your Private Network
What it is: Creates a private network between all your devices.
The magic: Your devices see each other as if they're on the same desk — even if one is in Brisbane and another is in Melbourne.
Benefits of Tailscale
- Zero configuration — Install app, log in, done. No router config, no port forwarding
- Encrypted — Uses WireGuard protocol. Military-grade encryption
- Fast — Direct peer-to-peer connections. No central bottleneck
- Works anywhere — Hotel wifi, mobile data, coffee shop — always connected
Real example: From iPhone, can SSH directly to Mini 2: ssh jarvis@100.107.30.117 — works from anywhere in the world.
Cloudflare Tunnels: Making Services Public
What they are: Secure tunnels that let you share services from your home computer to the public internet.
Why you need them: Some things need to be accessible to clients, not just you.
Live Tunnels Running Now
- jarvis.nathan-jarvis.com → Mini 1 port 3100 (JARVIS orchestrator)
- n8n.nathan-jarvis.com → Mini 1 port 5678 (Nathan's n8n)
- lovr-n8n.nathan-jarvis.com → Mini 2 port 5678 (LOVR's n8n)
- lovr-bi.nathan-jarvis.com → Mini 2 port 3100 (LOVR's Metabase dashboard)
- obsidian.nathan-jarvis.com → Mini 1 port 5984 (Obsidian vault sync)
Cloudflare Pages: Instant Website Hosting
What it is: Free hosting for static websites. Upload an HTML file, get a public URL instantly.
How it's used: All client deliverables deploy here:
- Intelligence reports
- Financial audits
- Strategy documents
- Research briefings
- System documentation
Example workflow:
- Claude Code generates client report (HTML file)
- Runs
npx wrangler pages deploy - Gets public URL:
https://lovr-audit.pages.dev - Sends URL to client via email
- Total time: 30 seconds
The power of this combo: Tailscale for private access (your devices), Cloudflare Tunnels for semi-private services (password-protected client tools), Cloudflare Pages for public sharing (client deliverables).
Real Scenario: Client Dashboard Access
Client request: "Can I see our business metrics in real time?"
Solution:
- Build Metabase dashboard on Mini 2 (pulls from Xero + Airtable)
- Create Cloudflare Tunnel:
lovr-bi.nathan-jarvis.com - Generate public dashboard link with Metabase
- Send link to client:
https://lovr-bi.nathan-jarvis.com/public/dashboard/...
Result: Client can view live dashboard from anywhere. No VPN. No login. Just works.
Next: Where all your knowledge lives and how it's organized.
The Knowledge Base: Obsidian + GitHub
Where everything is stored and remembered
What is Obsidian?
Obsidian is a note-taking app that stores everything as plain text files on your computer.
Why plain text matters:
- You own it forever — No vendor lock-in, no subscriptions required to access your notes
- Readable by anything — Any app can open .md files. Future-proof.
- Fast search — 597MB of notes searched in milliseconds
- Version control ready — Works with Git for full history
The Vault Structure
Everything lives in ~/Desktop/Nath/:
Folder Organization
- Clients/ — All client work
- LOVR-Agency/ (deliverables, research, brand kit)
- Newstead-Aesthetics/
- DC-Services/
- Nils-Barry/
- ...11 total clients
- Jarvis-Systems/ — The system itself
- Architecture/ (system specs, brand guidelines)
- Skills/ (40 saved skills)
- System-Reports/ (health checks, audits)
- MCP-Servers/ (configs and credentials)
- Projects/ — Personal ventures
- Trading-System/
- Calibre-Wellness/
- Ultratune/
- Smart-Home/
- General/ — Notes, screenshots, daily briefs, miscellaneous
- STATE.md — Current system status (updated every session)
How Obsidian Thinks
Obsidian isn't just a folder of text files. It's a connected knowledge graph.
Links create connections:
- Link to a person:
[[Nils Barry]] - Link to a project:
[[LOVR Collections Plan]] - Link to a concept:
[[Zero Trust Security]]
Click a link and Obsidian shows:
- The note itself
- All notes that link TO it (backlinks)
- All notes that IT links to
- Visual graph of connections
This is how human memory works. Ideas aren't stored in isolation — they're connected to other ideas. Obsidian mirrors that structure.
LiveSync: Real-Time Everywhere
The setup: CouchDB running in Docker on Mini 1, accessible at obsidian.nathan-jarvis.com
What it does:
- Change a note on MacBook → syncs to iPhone in 2 seconds
- Add a note on iPhone → appears on Mini 1 instantly
- All devices always have the latest version
Real example: At a BBQ, meet someone interesting. Pull out iPhone, create note in Obsidian with their details. By the time you get home, the note is already on your MacBook.
GitHub: Version History Forever
What GitHub does: Stores every version of every file, with full history.
Two repositories:
- jarvis-systems — All infrastructure code, configs, scripts
- jarvis-vault — The complete Obsidian vault (all notes, all clients)
Why this matters:
- Disaster recovery — If both Minis die, everything is on GitHub
- Version history — Go back to any point in time: "What was the LOVR plan on March 15?"
- Blame tracking — See exactly when and why each change was made
Real Scenario: Accidental Delete Recovery
Problem: Accidentally delete an important client note.
Recovery:
- Open terminal:
cd ~/Desktop/Nath - Check git history:
git log -- Clients/LOVR-Agency/client-notes.md - Find the commit before deletion
- Restore:
git checkout [commit-hash] -- Clients/LOVR-Agency/client-notes.md - File restored in 10 seconds
Data lost: Zero. Git saved everything.
Next: The database layer — where structured data lives permanently.
The Database: PostgreSQL + Airtable
Where your data actually lives
Two Types of Databases
The system uses two different database types for different purposes:
PostgreSQL — The Engine Room
What it is: Industrial-strength database. Powers half the internet.
Designed for:
- The system to use (not humans)
- Fast complex queries
- Absolute reliability
- Massive scale
Interface: SQL commands, APIs
Airtable — The Human Interface
What it is: Spreadsheet that thinks like a database.
Designed for:
- Humans AND automation to use
- Visual editing and relationships
- Collaboration and sharing
- Flexible schema changes
Interface: Beautiful web UI, mobile apps
What Lives in PostgreSQL
People Table (32 contacts)
Columns:
- id, name, email, phone
- company, role, industry
- relationship_type (prospect, client, partner)
- relationship_strength (1-5)
- first_contact_date, last_contact_date
- tags (array), notes (text)
- created_at, updated_at
Real example: "Nils Barry, E&S Residential, Property Development, Brisbane, Prospect, Strength 3"
Interactions Table (logged activities)
Columns:
- id, person_id (foreign key)
- interaction_type (email, meeting, call, message)
- date, subject, notes
- outcome (positive, neutral, negative, pending)
Real example: "Email to Jack Brinsmead, April 5, Subject: Invoice reminder, Outcome: Pending"
LOVR Financials (Xero sync)
Data:
- Revenue: $657,383 (11.5 months)
- Expenses: $171,957
- Net profit: $118,121 (18% margin)
- Outstanding: $101,611.30
- Bank balance: $4,432.54
Updated: Daily via n8n workflow
LOVR Clients (11 active)
Tracked:
- Client name, status, value
- Retainer amount, billing frequency
- Project count, active projects
- Last invoice date, payment status
Powerful Queries
With SQL, you can ask sophisticated questions:
Query: Who haven't I contacted in 30+ days?
SELECT name, email, company, last_contact_date
FROM people
WHERE last_contact_date < NOW() - INTERVAL '30 days'
AND relationship_type IN ('client', 'prospect')
ORDER BY last_contact_date ASC;
Result: List of 7 people who need follow-up, sorted by urgency.
Query: Top 5 clients by project value
SELECT client_name, SUM(project_value) as total_value FROM lovr_clients WHERE status = 'active' GROUP BY client_name ORDER BY total_value DESC LIMIT 5;
Result: BP Wharton $39,819, ALT Agency $35,200, Hipwood $24,570...
What Lives in Airtable
When Airtable is set up (ready to build):
- Clients base — Visual CRM with status, notes, linked contacts
- Content pipeline — Editorial calendar, asset tracking, approval workflows
- Project tracker — Tasks, deadlines, dependencies, team assignments
- Lead management — Capture, scoring, nurture sequences
Why Airtable for these? They need human collaboration. Team members can edit records, comment, attach files — without learning SQL.
The Sync Flow
PostgreSQL and Airtable work together:
- Automation writes to PostgreSQL — Fast, reliable, permanent
- n8n syncs key data to Airtable — Human-friendly interface
- Humans edit in Airtable — Visual, collaborative
- n8n syncs changes back — PostgreSQL stays authoritative
Real Example: Contact Profile Query
Request: "Tell me about Jack Brinsmead"
What happens:
- Query PostgreSQL people table → Find record
- Query interactions table → Get last 5 emails, calls, meetings
- Query LOVR financials → Check outstanding invoices ($27,885)
- Search Obsidian → Find meeting notes
- Semantic search → Pull related context
- Format complete profile
Response delivered in 2 seconds:
"Jack Brinsmead, Owner of BP Wharton (construction). Last contact: March 15 (phone call, positive). Outstanding: $27,885 across 3 invoices (45+ days overdue). Relationship strength: 4/5. Recent project: Gabba redevelopment. Recommended action: Personal call this week with payment plan."
The database is the source of truth. Redis might clear, Obsidian might have old data, but PostgreSQL is always accurate. Everything important gets written there.
Next: How to build your own system from scratch.
Building Your Own
Step-by-step from zero to fully operational
The 4-Week Build Plan
You don't build everything at once. Start simple, add layers week by week.
Week 1 — The Foundation
Goal: Get Claude Code working with your first skill and one MCP server.
Steps:
- Buy a Mac Mini — M4, 16GB RAM ($800-1,200)
- Why Mac Mini? Low power, always-on, macOS works great with Claude Code
- Install Claude Code — Free to start, upgrade to Pro later ($20/month)
- Download from
claude.ai/code
- Download from
- Create your first skill —
/morningbriefing- Read STATE.md, check calendar, send summary
- Connect Gmail MCP — Takes 20 minutes
- OAuth setup, test reading inbox
By end of week: Say /morning and get your daily briefing. First automation working.
Cost so far: $800-1,200 (Mac Mini) + $20/month (Claude Pro)
Week 2 — Memory and Knowledge
Goal: Set up permanent memory so nothing is forgotten.
Steps:
- Install Obsidian — Free
- Create vault at
~/Documents/Vault/
- Create vault at
- Install Docker Desktop — Free
- Download from docker.com
- Run PostgreSQL in Docker — Free
- Command:
docker run -d --name postgres -e POSTGRES_PASSWORD=yourpassword -p 5432:5432 postgres
- Command:
- Create people table — Use SQL or Claude Code
- Columns: id, name, email, company, last_contact
- Install n8n via Docker — Free
- Command:
docker run -d --name n8n -p 5678:5678 n8nio/n8n
- Command:
- Build first workflow — Daily email summary
- Trigger: Schedule 8am daily
- Action: Gmail → Get unread → Send digest
By end of week: Database storing contacts. Obsidian capturing notes. First n8n workflow running daily.
Cost so far: Still just Mac Mini + Claude Pro
Week 3 — Automation
Goal: Work happens automatically without you.
Steps:
- Connect Google Calendar MCP — See your schedule
- OAuth setup similar to Gmail
- Set up Airtable — Free tier is plenty
- Create "Clients" base
- Add Airtable MCP to Claude Code
- Build lead capture workflow — n8n
- Webhook receives lead → Create Airtable record → Notify you
- Add Perplexity MCP — $20/month for API
- Enables web search in Claude Code
- Add Firecrawl MCP — $20/month
- Enables website scraping
- Create research skill —
/research [topic]- Search web → Scrape pages → Write report → Save to Obsidian
By end of week: Lead capture working. Research on demand. Calendar integrated.
Cost so far: $800-1,200 + $60/month (Claude + Perplexity + Firecrawl)
Week 4 — Intelligence
Goal: Full system intelligence and remote access.
Steps:
- Install Tailscale — Free for personal use
- Install on Mac Mini, MacBook, iPhone
- Instant secure access from anywhere
- Set up Cloudflare account — Free
- Create tunnel to expose n8n externally
- Create Telegram bot — Free
- Talk to @BotFather on Telegram
- Get bot token
- Connect Telegram to Claude Code — Custom integration
- Use grammY library (Node.js)
- Route messages to Claude Code
- Add Redis for queues — Free (Docker)
docker run -d --name redis -p 6379:6379 redis
- Build priority queue system — Urgent/Normal/Background
- Intent classification with Claude
- Queue management with BullMQ
By end of week: Full system accessible via Telegram. Voice notes working. Priority queues managing workload.
Cost so far: $800-1,200 + $60/month + $0 (all new services free)
Total Cost Breakdown
One-Time Costs
- Mac Mini M4 16GB: $800-1,200
- Everything else: $0
Total: $800-1,200
Monthly Costs
- Claude Pro: $20
- Perplexity API: $20
- Firecrawl: $20
- Airtable: Free tier (or $20)
- Everything else: $0
Total: $60-80/month
Compare to hiring someone: An assistant costs $3,000-5,000/month. This system costs $60-80/month and works 24/7 without vacation or sick days.
What You Can Skip
Not everyone needs everything. Here's what's optional:
- Second Mac Mini — Only needed if you're managing client work separately
- Paid MCP servers — Start with free ones (Gmail, Calendar, Drive)
- Telegram bot — Can use Claude Code desktop app instead
- Tailscale — Only needed for remote access to local services
- Redis — Only needed for queue management at scale
Minimal viable system: Mac Mini + Claude Code + Gmail MCP + Obsidian + n8n = $820 + $20/month
Common Mistakes to Avoid
- Building too much at once — Start with one workflow, get it perfect, then add more
- Not documenting — Write down what you build in Obsidian as you go
- Ignoring backups — Set up GitHub from day 1
- Over-engineering — Simple workflows often beat complex ones
- Not testing failures — Restart your Mac, make sure everything comes back up
Next: See how all the pieces work together in real scenarios.
The Full Picture
How it all works together — start to finish
Morning Scenario: Automated Intelligence
6:00 AM — System Wakes Up
What happens:
- n8n triggers — "Morning Intelligence" workflow starts
- Fetches data:
- Xero API → Financial data (revenue, expenses, bank balance)
- Airtable → Project status (11 clients, active tasks)
- Gmail → Unread urgent messages
- PostgreSQL → Yesterday's metrics
- Claude processes:
- Analyzes trends (revenue up/down, cash flow issues)
- Identifies priorities (overdue invoices, urgent emails)
- Formats HTML report
- Delivers:
- Email to primary inbox
- Brief to Telegram
- Saves to Obsidian daily log
6:05 AM — Complete. Before you're awake, you're already briefed.
Human involvement: Zero.
Midday Scenario: Voice Note Research
12:30 PM — Walking Between Meetings
User action: Send Telegram voice note while walking:
"Research this company I just met with, Newstead Aesthetics, need to know their tech stack, team size, whether they're worth following up with"
12:31 PM — Transcription:
- Audio hits Telegram bot
- Whisper transcribes to text
- Text: "Research this company I just met with..."
12:31 PM — Classification:
- Claude Haiku analyzes intent
- Classification: DEEP_RESEARCH
- Priority: Background queue
12:31-12:42 PM — Execution (Claude Code):
- Perplexity search: "Newstead Aesthetics Brisbane"
- Finds website, social media, business listings
- Firecrawl scrape: newsteadaesthetics.com
- Identifies Wix platform (basic tech stack)
- Finds team page (6 staff members)
- Extracts service offerings
- Apify Instagram: @newsteadaesthetics
- 8,400 followers
- Engagement rate 3.2%
- Posting frequency 4x/week
- Cross-reference: ABN lookup, Google Maps reviews
- Est. 2019 (5 years operation)
- 4.8/5 stars (142 reviews)
- Analysis: Compare to typical aesthetics clinics
- Tech stack: Basic (high upgrade potential)
- Size: Small-medium (room to grow)
- Digital presence: Strong organic, weak automation
- Write report: 1,200-word intelligence brief
- Company overview
- Tech assessment
- Growth opportunity analysis
- Recommended approach
- Save and notify:
- Full report → Obsidian vault (Clients/Newstead-Aesthetics/)
- Contact record → PostgreSQL people table
- Summary → Telegram notification
12:42 PM — Notification arrives:
"✅ Newstead Aesthetics research complete. 6-person team, Wix site, strong Instagram (8.4K followers), basic automation. High-value target for systems upgrade. Est. potential: $4-8K/month retainer. Full report: Clients/Newstead-Aesthetics/intelligence-report.md"
Total time: 12 minutes. User was walking the entire time.
Evening Scenario: System Works While You're at the BBQ
6:00 PM — You're Off the Clock
What the system does without you:
6:07 PM — Hourly sync:
- Mini 1 syncs ~/Desktop/Nath/ to Mini 2
- All client work, system updates, new notes transferred
- 622 files checked, 3 updated, 30 seconds total
7:00 PM — Queue processing:
- Background research jobs continue
- 5 pending deep research tasks complete
- Reports saved to Obsidian
- Notifications queued for tomorrow morning
9:00 PM — LOVR invoice workflow:
- n8n checks Xero for overdue invoices
- Finds 3 invoices 14+ days overdue
- Generates polite reminder emails
- Sends via Gmail
- Logs activity in Airtable
- Updates PostgreSQL
11:59 PM — End of day:
- STATE.md updated with today's activity
- Daily backup pushed to GitHub
- Obsidian vault synced across all devices
- Redis queue cleared
- System ready for tomorrow
Your involvement: Zero. You were at a BBQ the entire time.
The Key Insight
This is not the future. This is running right now.
Every example in this guide comes from a real, working system:
- Managing 11 clients across multiple industries
- Tracking $657K in revenue
- 32 contacts with complete interaction history
- Thousands of automated tasks per month
- Zero human intervention for 95% of operations
The tools exist. The knowledge is here. The question is: What will you build?
Where to Start
- Pick one pain point — What's the most annoying manual task in your day?
- Build one workflow — Make that task automatic. Get it working perfectly.
- Add one tool at a time — Don't try to integrate everything at once
- Document as you build — Future you will thank present you
- Iterate and improve — Each week, add one new capability
In 4 weeks, you'll have a working AI system that remembers everything, automates your routine work, and gives you superpowers you didn't know were possible.
Final Thought
The most valuable companies in the world aren't built by people who work hardest. They're built by people who work smartest — who build systems that amplify their effort.
This guide gave you the blueprint. Now it's your turn to build.