The Press
Field NotesIssue 002 · Eight Apps, One Year

How I Built a 15-Wing Personal AI Operating System in 6 Months

JarvisOS isn't a productivity app. It's a full intelligence layer for my life: CEO decisions, finance, cycle tracking, RAG knowledge base, crisis routing. Here's the architecture.

Nandawula Regine10 June 202610 min read

# How I Built a 15-Wing Personal AI Operating System in 6 Months

Six months ago I had a problem every solo founder who also writes, who also runs a business, who also manages their health knows: the cognitive overhead of switching between 15 different tools was destroying my execution capacity.

I didn't download another app. I built JarvisOS.

What JarvisOS Actually Is

JarvisOS is a personal AI operating system with 15 interconnected intelligence wings. Each wing handles a domain of my life. Together, they form a unified intelligence layer that knows context across domains.

The 15 wings:

  • CEO: Decision-making engine, strategic clarity
  • Finance: Cash flow, invoice tracking, ZAR projections
  • Engineering: Build logs, architecture decisions, commit histories
  • Marketing: Campaign planning, content calendar, audience strategy
  • Cycle: Menstrual cycle intelligence, energy optimization per phase
  • Scholar: Learning library, course notes, research synthesis
  • Corpus: RAG-powered personal knowledge base (1,194 chunks)
  • Body: Health tracking, sleep, nutrition intelligence
  • Sanyu: Wellness rituals, mental health check-ins
  • Client Portal: Client management, project status, billing
  • UX Intelligence: Usability research, design decisions
  • Docs: Documentation hub, living specs
  • Consulting: Proposal generator, scope estimator
  • Autobiography: Living biography, memory system
  • Crisis/Sankofa: Crisis detection, SA support routing (SADAG, Lifeline)

The Architecture Decisions That Actually Matter

Model Routing: Sonnet + Haiku

Not every wing needs the same model. The CEO wing makes high-stakes decisions: it gets Claude Sonnet. The Sanyu wing sends a wellness check-in notification, that gets Haiku.

This single decision cut my AI costs by 40% without reducing quality anywhere it mattered.

function getModelForWing(wing: WingId): string {
  const heavyWings = ['ceo', 'corpus', 'engineering', 'consulting'];
  return heavyWings.includes(wing)
    ? 'claude-sonnet-4-6'
    : 'claude-haiku-4-5-20251001';
}

Prompt Caching on Every System Block

Every wing has a system prompt. The system prompt is static: it describes the wing's role, my context, the rules of engagement. In Claude API, any static block ≥4096 tokens can be cached. I made sure every system prompt hit that threshold.

Result: cache read tokens cost 10% of standard tokens. On 15 wings used daily, this compounds fast.

Redis Signal Protocol: Wings Talk to Each Other

This is the architectural decision I'm most proud of. Wings don't just respond to me: they signal each other.

When the Cycle wing logs a new phase, it writes a signal to Upstash Redis:

cycle:phase_changed → {phase: 'follicular', energy: 'rising'}

The Marketing wing reads this signal when planning content. The Body wing adjusts workout recommendations. The CEO wing knows to schedule high-stakes calls in the high-energy follicular phase.

This is not just an AI chatbot. This is an orchestrated intelligence system.

The Corpus: 1,194 Knowledge Chunks via Upstash Vector

The Corpus wing ingests everything: my book, my project documentation, my journal entries, my research notes. 1,194 chunks, vectorized and stored in Upstash Vector.

When I ask any wing a question, the Corpus retrieval runs first. The AI answers with the context of everything I've ever thought or documented. It's like having a second brain that never forgets.

Inngest for Async Everything

Some things can't happen synchronously. The daily brief compiles insights from 8 wings. The production log update pings finance. The crisis check-in waits for a response before routing.

Inngest handles all of this. Every long-running task is an Inngest function. No timeouts, no blocking, full visibility into every step.

The Technical Baseline: TypeScript Strict Mode

Every single wing. Strict mode. tsc --noEmit exits 0. This wasn't a nice-to-have: it was the rule I set before writing the first line of wing-specific code.

When you're building 15 interconnected systems solo, TypeScript is not overhead. It's the only thing standing between you and a 3am debugging session caused by a type that crept from the Cycle wing into the Finance wing's calculations.

What 6 Months Taught Me About AI Systems

Context is everything. An AI that knows I'm in luteal phase, have a product launch in 3 days, and slept 5 hours will give fundamentally different advice than one that just sees the question. The investment in multi-wing context is the investment in actually useful AI.

Orchestration beats raw capability. A brilliant isolated AI agent is less useful than a well-orchestrated network of specialized agents that share state. This is why JarvisOS exists.

Your personal operating system is a competitive advantage. The compounded intelligence of a system that has 6 months of your decisions, your patterns, your knowledge, that's not replaceable with a new app.

JarvisOS is live, actively used, and still evolving. It is the most technically sophisticated thing I've built. It is also, in its own way, the most personal.

What would a personal AI operating system designed for your actual life look like?

Colophon

Imprint
Field Notes
Issue
Issue 002, June 2026
Published
10 June 2026
Length
793 words
Drafts on file
None yet
Set in
Cormorant Garamond & DM Sans

Published by The House of Roses Press, KuGompo City.

2 views

Reader Insights

0 responses

No insights yet. Be the first!