# From HTML to 15-Wing AI OS: One Year of Technology
July 13, 2025. YouTube Clone. HTML, CSS, and the particular thrill of pressing F5 and seeing something appear in a browser that I typed.
June 2026. JarvisOS. 15 intelligence wings. 1,194 RAG knowledge chunks. Redis inter-wing signal protocol. Claude Sonnet and Haiku routing based on task complexity. PWA offline-first. Still running on the same curiosity as the YouTube Clone, now with 12 months of accumulated technical debt that I've mostly paid off.
This is the arc.
Month 1-2: Foundations (Jul-Aug 2025)
What I built: YouTube Clone, MoodCast Weather App, Netflix Landing.
What I learned: HTML is forgiving in ways that will hurt you later. CSS is unforgiving in ways that build the discipline you need for everything else. The first time something works in your browser because you made it work, not because the tutorial made it work, is the only inflection point that matters.
What I'd tell month-1 me: The box-model is not obvious. Neither is z-index. Learn them properly now. You'll debug them for a year if you don't.
Month 3-4: Production Reality (Sep-Oct 2025)
What I built: Cortex Hub Booking, GreenVault eCommerce, True Access v1, PoetryTube.
Tech added: Next.js 14, TypeScript, Supabase, PayFast ITN, Mapbox, PostGIS.
The moment: First PayFast ITN webhook in production. A real customer paid real money for a real product. The R99 hit the Supabase orders table and I stared at it for longer than I should have.
What I'd tell month-3 me: Set up TypeScript strict mode from the first commit of every project. Not because you understand why yet. Because the 6-month version of you will thank you when you're maintaining 8 codebases simultaneously.
Month 5-7: AI Engineering Begins (Nov 2025 - Jan 2026)
What I built: VarsityOS, StokvelOS, K53 Drill Master.
Tech added: Claude API, prompt engineering, two-block prompt caching, SM-2 spaced repetition, routerless state machines.
The architecture shift: Realizing that AI isn't a feature you add, it's a layer that changes the product's relationship to the user. Nova on VarsityOS isn't a chatbot in a student app. Nova is the primary interface. Everything else is scaffolding.
What I'd tell month-5 me: Cache your system prompts before the API bill arrives. The bill will teach you this anyway, but learning it earlier is cheaper.
Month 8-10: Enterprise Complexity (Feb-Apr 2026)
What I built: AdminOS (multi-tenant, WhatsApp-native, 5 specialist agents).
Tech added: Inngest async queues, Upstash Redis, Meta WhatsApp Cloud API v19.0, multi-tenant RLS at privilege level, immutable audit logs.
The hardest thing: Two Supabase clients in one codebase, one for platform operations, one for tenant-scoped data. The architectural discipline required to never mix them. The database-level enforcement that backs up the application-level discipline.
What I'd tell month-8 me: The audit log should be immutable before you write any other admin feature. Not after. Before.
Month 11-12: Personal Operating System (May-Jun 2026)
What I built: JarvisOS (15 wings), Sanyu Botanicals, True Access v2 (Expo SDK 52).
Tech added: Upstash Vector RAG, Redis inter-wing signal protocol, Expo SDK 52, NativeWind, Paystack for mobile, offline-first PWA with IndexedDB.
The realization: JarvisOS is not more complex than AdminOS. It's differently complex. Multi-agent orchestration for a business is about service reliability and tenant isolation. Multi-wing orchestration for a personal OS is about context depth and signal fidelity. Different problems, same underlying patterns.
What Didn't Change
TypeScript strict mode. Supabase RLS. Soft deletes only. UTC timestamps. PayFast in ZAR. Building for users I actually understand from the inside.
The fundamentals established in month 1 are the same fundamentals running in month 12. The stack evolved. The principles didn't.
What would your year look like if you started building today?
Reader Insights
0 responses
No insights yet. Be the first!