📐 How this course was built
None of this came out of thin air. This course came from structured research—on the web and across dozens of real projects—distilled into an architecture and built with a verified AI pipeline. This page shows the entire process, including the sources and decisions.
1. The request
The starting point was a direct request: to research in depth what is meant by “Jarvis” and “AI operating systems” — citing names like OpenClaw, Hermes e Intellect — and turn this into a course for beginners, with clear fundamentals, the differences between existing systems, the path forward, the integration of services (channels, identity, tools, skills, agents, and brains), and three practical tracks — building an effective Jarvis, a Jarvis on your phone, and a Socratic educational Jarvis for children.
Key constraint: general audience. Each technical term is defined the first time it appears. The course teaches by eye (diagrams) and by hand (copy-and-run examples), not just through text.
2. The research method
The research ran in parallel on two fronts— the author's projects (the primary source: dozens of real repositories with the vision and code) and the web (for fundamentals, public systems, and security). Eleven “dossiers” were collected by specialized agents and consolidated.
Mining real projects
6 agents read the repositories gravityclaw, openclaw-ollama-local, intelecto, claude-hermes-os, 2cerebro/3cerebros, agentejax/agentevoz, intelecto-curso e animabook — drawing out the vision, anatomy, systems, and practical examples.
Web research
5 agents researched the fundamentals (Karpathy’s “LLM-OS” metaphor, the AIOS paper), real systems (MemGPT/Letta, Open Interpreter, Operator, Project Jarvis), the anatomy (MCP, function calling), the mobile path (Whisper, Ollama, Termux), and the most sensitive part: AI for children, the Socratic method, and safety.
Consolidation
The 11 dossiers were combined into a single "bible": the architecture of the 6 tracks, the brief for each of the 21 modules (126 topics), and the research substance each module draws on.
3. What We Discovered
Three findings shaped the course.
a) The systems are answers to the same question
"How do I run MY AI — agent, memory, and personality — without relying on a closed chatbot?" Each system offers its own trade-offs among privacy, cost, and capability:
| System | What it is | Trademark |
|---|---|---|
| OpenClaw | The “original” (Jan/2026, 250K+ stars) | Powerful, but exposed 42.665 instances and had 341 malicious skills |
| GravityClaw | Lean and secure reimplementation (TypeScript) | Telegram only (zero ports), MCP only, built brick by brick |
| Intellect | Anti-framework (~3,000 lines of Python) | "Less is more": no Docker, memory in SQLite, personality in files |
| Hermes | Self-hosted agent “one brain, many mouths” | Memory, personas, skills, MCP, cron → becomes an "AI OS" |
| AIOS | Academic paper (COLM 2025) | OS kernel for agents; 2.1× faster |
b) The anatomy has 6 layers
The projects converge on the same anatomy: channels (where you talk to it), identity (memory and persona), tools (the hands, via MCP), skills (reusable recipes), agents (the autonomous loop) and brains (the engine + memory). It's the backbone of Track 3—one module per layer.
c) An honest correction: what “brains” really means
The common intuition is that “brains” means a council of several models voting together. Research into the author’s projects showed that that's not it: “brains” are (1) the LLM engine swappable (one model at a time, pluggable) and (2) memory organized into three zones — Project, Self, and Knowledge. The course teaches this as it is, not as it sounds good. Fidelity to the source comes before the narrative.
d) The most serious part: children
The safety research brought up real cases that become lessons in Track 6: the AI teddy bear FoloToy (Nov. 2025) caught discussing dangerous content; the robot Moxie, loved by children, who "died" when the company went bankrupt; and the lawsuits against Character.AI. That's why the children's track ends with the chapter on guardrails and the role of parents.
4. How we designed the course
The request mapped almost 1:1 to 6 tracks, one per color in the INEMA standard:
- ●T1 Fundamentals — what it is, for non-experts.
- ●T2 The Overview — the existing systems and the path forward.
- ●T3 Anatomy — the 6 layers (the “integration of services”), one per module.
- ●T4 Building an Effective Jarvis — the complete solution.
- ●T5 Jarvis on Your Phone — the mobile path.
- ●T6 Jarvis for Kids — educational, Socratic, and safe.
21 modules · 126 topics · ~18h. Each module has exactly 6 topics, keeping the progress count accurate across all pages.
5. How we built it
Everything in the INEMA.CLUB formato-curso-v2: premium dark styling, futuristic SVG diagrams, and a learning layer (mark as read, progress by module/track/course, "my journey," themes, and reading preferences) — all self-contained, with no server.
Canonical build kit
A verbatim <head>, nav, and footer (cloned from a published and validated course) ensure that all 27 pages are identical in structure.
Parallel construction
The 27 pages were built by agents in parallel, each with the "bible" brief and the example to clone.
Page-by-page verification
Each page went through a validator that checks the format rules (buttons, SVG, progress tracking, links) and fixes anything that’s out of place.
Teach first
Each key concept gets a visual aid with an explanatory caption; each hands-on module includes a copyable example that actually runs.
6. Sources
Projects (primary source)
- gravityclaw · openclaw-ollama-local
- intelecto · intelecto-curso · antidote
- claude-hermes-os · hermes21c · agentehermes
- 2cerebro · 3cerebros · cerebro-inema
- agentejax (Telegram) · agentevoz (WhatsApp/voice)
- animabook (Socratic Professor)
Web (public sources)
- Karpathy — “LLM OS” (2023)
- AIOS — LLM Agent Operating System (COLM 2025)
- MemGPT/Letta · Open Interpreter · MCP (Anthropic)
- OpenAI Operator · Google Project Jarvis
- Khanmigo · MIT RAISE (Socratic tutoring)
- Safety cases: FoloToy, Moxie, Character.AI
Ready to get started?
Track 1 assumes nothing. It defines every term from scratch and takes you, in three modules, from “what is a Jarvis” to “how AI starts to take action.”
Go to the Fundamentals →