Everything is now ONE academy with one path. Five levels, each with its own page, tests, and a gate: pass the level's exam before moving up. The old track names are gone — here's the map so nothing from before is lost.
Green = passed, tap anytime to review. The teal dot = your current step. Locked steps open when the one before is passed. Dashed dots are reference stops — always open.
1. Order is law. Level 0 before 1, 1 before 2. Every "advanced" confusion you've ever had came from a skipped basic.
2. Gates, not vibes. A level counts as passed when its exam score meets the bar (Level 1: 8/10 · Levels 2–3: all module quizzes complete with ≤1 miss each after retake). Report each gate in chat — that's the accountability.
3. Spaced retakes. Any quiz you missed questions on: retake it 1 week later, then 3 weeks later. This is when knowledge actually moves into long-term memory — skipping it means re-learning everything in the interview chair.
4. Doing beats reading. Every level has homework tied to your real projects. The portfolio artifacts ARE the expert level; there is no reading path to expertise.
Fifteen concepts. Assume you know nothing — because pretending is how gaps survive. Tap each card, read all four rows, no exceptions.
Six questions. Pass = 5/6. Misses unlock a simpler lesson; retake in 1 week regardless.
You asked what the best learning design is. Here's the honest answer, and how every piece of it is wired into these pages.
Session A (45–60 min): new module — cards/lesson, then quiz cold. Session B (45–60 min): second module OR homework on a real project. Session C (30 min): retakes of last week's misses + 10 minutes of Tutor drilling. Rest of the week: MyClienta work counts as lab time — after Level 2, every client task doubles as practice. Two focused sessions beat seven distracted ones; put A and B in the calendar like client meetings.
Your four postings were examples of families. Here's the whole map: what each family does, and which academy level makes you dangerous in it. Titles vary wildly ("Digital Lead", "AI Specialist", "Transformation Manager") — match by RESPONSIBILITIES, not titles.
1. Build, advise, govern, or teach? Every AI role is a mix of these four verbs — find the dominant one and you know the family, whatever the title says. 2. Which level does it test? Words like "use cases, governance, stakeholders" = Level 2. "Agents, prompts, MCP, integration" = Level 3. 3. What's the one hard filter? (degree, years, a named tool, a language). Decide honestly if you pass or can argue it. 4. What proof would make them relax? Then bring THAT artifact to the interview. Paste any posting into the Level 2 Tutor and it runs this decoder with you.
The honest ledger. Update it in chat as gates fall.
Every module: a story from your business, plain-language deep dives, a test. Wrong answers unlock a simpler remedial lesson. Your progress saves on this device. The Radar looks two years ahead; the Business tab tells you what to build and who pays.
Nine tools, one job each.
Tap one. The "mix-up" row is where the education is.
"What is agentic AI, do I have it?" Earn the answer.
One question separates all three: who decides the next step?
Trigger fires → fixed steps run. Same thing every time, zero judgment. A light switch with extra steps.
Fixed steps, but a step calls AI — "write this email", "score this ad". The path never changes; only words inside it do.
You give a goal. The AI picks tools, order, checks its own result, loops until done or gives up.
Rule: if you drew the arrows in n8n, it is not an agent — no matter how much AI sits inside the boxes.
Marketing words don't count.
You own zero agents. Everything you've shipped is automation or AI workflow — and that is a feature. Workflows are predictable, cheap, debuggable: what a paying client wants. Agents fail creatively. The only real agent in your business is the one you use: Claude Code.
Sales rule: say "AI-powered automation," not "AI agents," until you've built one. Clients who know the difference multiply monthly.
Try both live on fake RentingPilot data.
SQL is a strict spreadsheet: exact rows, exact matches. Vector is a librarian who finds things by meaning, even when no words match.
Pick a search, then flip the mode and run the same search:
You use SQL only. Supabase runs Postgres — rows and columns. Perfect for bookings, licence keys, client records. Vector memory stores text as lists of numbers ("embeddings") where similar meanings sit near each other — needed for exactly one job: searching documents by meaning, the heart of RAG.
Good news: Supabase has a vector add-on, pgvector. When the day comes, no new tool — just a new switch in one you already own.
Don't add a vector DB because it sounds advanced. No current client project needs one.
You pitched RAG in a 17-slide deck. Own what you pitched.
Claude is a brilliant contractor with total amnesia. RAG = Retrieval-Augmented Generation: look up the relevant notes first, staple them to the question, then let the brain answer.
1 — Tokens. You pay per chunk of text, in and out. Every retrieved document stapled to the prompt costs money on every question. Over-fetching RAG is a quietly bleeding bill.
2 — Retrieval quality. Wrong notes fetched → confident answers from wrong notes. RAG moves the problem from "the brain" to "the librarian."
Your Knowledge Assistant deck — local model, no data leaving the building — is textbook RAG. You can now defend every slide.
Four taps, the tree answers. Run it for several projects.
The seven things a developer sees that you currently don't. Learn them and you stop being hostage to jargon.
The #1 developer skill: reading an error calmly. Three real-style errors from your world — classify each.
"You're using this, but you could use that." For each slot: what you run, the alternatives, and the honest switch verdict.
Notice the verdicts: stay, stay, stay, stay, stay, consider-one. That's not laziness — your stack is genuinely well-chosen for a one-person shop. The trap for self-taught builders isn't picking wrong tools; it's switching tools to feel productive instead of shipping. Tool-shopping is procrastination wearing a work costume.
Six business models your stack supports, each with the co-founder questions: who pays, when, how much effort. Verdicts included.
Both your paying clients — Sorin and Valentin — are rental businesses. That's a vertical, not a coincidence. Client #3 in rentals costs a week; project #21 in a new field costs a quarter. Every idea above got its verdict by one filter: does it make the rental vertical stronger? Use the same filter on every idea you have this year.
Where technology moves in ~2 years and what each move does to your money. Honest label: informed bets, not certainties. Refresh this monthly in chat — a file can't watch the news.
Ten questions, all eight modules. Pass mark 8.
Live — saved on this device.
Take the module tests. This panel scores you as you go and names your weakest area.
Track 1 (your installed app) taught you how the technology works. Track 2 teaches what AI Strategy & Adoption roles are actually tested on: use-case prioritization, data readiness, governance, vendor assessment, and translating AI for humans. Built from a real posting — the Vestas AI Strategy & Portfolio Advisor role.
Foundation → frameworks → proof. Each week fits your hours.
Everything here feeds BOTH targets. Use-case prioritization = interview answer AND how you scope MyClienta client work. Governance = job requirement AND what lets you sell AI to serious Danish companies. Vendor assessment = job skill AND your own build-vs-buy decisions. Nothing you learn is single-purpose — that's why both-in-parallel is realistic at 5–7 hrs.
The core skill of the Vestas role, in its own words: "identify, scope, and prioritize AI use cases with clear business value."
Companies don't have an AI problem — they have a too-many-ideas problem. Everyone wants "AI for something." The advisor's job is being the adult who asks: which ideas are worth money, and which are toys?
Every proposed use case gets two honest scores. Value: money saved or earned, per year, if it works. Feasibility: is the data available and clean, is the process stable, is the tech proven for this job?
"Ensuring initiatives are grounded in data reality before resources are committed" — Vestas, verbatim. The unglamorous skill that separates advisors from enthusiasts.
AI is a chef. Data is the ingredients. Most failed AI projects didn't have a bad chef — they had an empty or rotten fridge, discovered after hiring the chef. The advisor's job: inspect the fridge before signing the chef.
You've lived this: the Kaeli veterinary project stalled on inaccessible CVR data, and Smaafolk needed a whole enrichment pipeline before the map was useful. That pipeline WAS data-readiness work. Now you have the professional name for it.
In every EU posting now — Vestas asks for "guidelines, guardrails, and responsible use practices." Your biggest gap; also the fastest to close, because it's rules, not math.
Governance sounds boring until you translate it: who is allowed to let a machine make which decisions, and who answers when it's wrong. That's it. Everything else is paperwork around that question.
"Supporting build-vs-buy decisions, including structured assessment of AI vendors" — you do this weekly by instinct. Here's the structure that makes it a credential.
Every vendor demo is a first date: everyone's charming, nothing's broken, and the price of leaving is invisible. The advisor is the friend who asks the unromantic questions.
Buy when the problem is generic (email, CRM, transcription) — your differentiation isn't there. Build when the process IS your edge or vendors would own your core data. And the third option interviews forget: wait — when the category is improving 50% a year, this year's purchase is next year's regret. Cite voice AI.
"Translating complex AI concepts into clear, actionable language" + "supporting AI literacy initiatives." Your strongest card — sharpened.
The advisor's real product isn't analysis — it's calibrated expectations. Too much hype and the project dies at first error; too much fear and it never starts. You manage the thermostat.
Structure that works: one live demo on THEIR real task (not a canned one) → the three-tier trust dial → hands-on with their own example → one page of do/don't (never paste client personal data into public tools). One hour, no slides beyond five. You've effectively been doing this for clients — now it's a named deliverable.
What's out there, your fit today, and the gap this academy closes. Paste your LinkedIn postings into the Tutor to map any specific one.
Pattern to notice: every emerging role rewards the same combination — operational experience + governance vocabulary + the ability to teach. That's precisely the stack this academy builds. Not an accident.
Ask anything — a term you don't get, a mock interview question, "explain simpler," or paste a job posting to map it against your profile. It knows your background and this curriculum.
Track 1 told you what an agent is. Track 2 told you how to advise about them. Track 3 opens the hood: the anatomy of an agent, prompt engineering, tools and MCP, memory, the end-to-end process of building anything, and how professionals test what they built. Written for someone who has been saying the words to Claude Code — so the words finally have machinery behind them.
Match / gap / action for each — from the actual postings you pasted.
Every posting wants the same trinity: ops experience + hands-on AI + the ability to explain. Nobody asked for a computer science degree. Two asked for governance awareness (Track 2), two asked for agent/MCP fluency (this track). The market is describing you — it just doesn't know your name yet. Your job in interviews is closing that last gap: translating MyClienta into their vocabulary.
Every agent on earth — Claude Code, a support bot, a research agent — is these six parts. No exceptions.
An agent is not a magic being. It's a loop wrapped around a language model, with hands attached. Learn the six organs and you can read any agent product's marketing and see exactly what's inside.
When the model "uses a tool," here is literally what happens: the model outputs a structured message — {"tool":"check_fleet","date":"Tuesday"} — your code sees it, runs the real database query, and pastes the result back into the conversation as text. The model then reads it and continues. The AI never touches anything. It writes requests; your plumbing does the work. Once you see this, "agentic AI" stops being mystical: it's a conversation where some replies are executed.
SimCorp lists it as a requirement. It's not magic words — it's writing job descriptions machines can't misread.
A prompt is a briefing for a brilliant contractor with amnesia. Everything they need must be in the briefing — role, task, rules, format, examples — because they know nothing about you and forget everything after.
Same model, same cost per attempt — wildly different reliability. Prompt engineering is cheap engineering: you're buying error-rate reduction with words.
"Advise functions on connectors to enterprise systems and MCP servers" — SimCorp, word for word. Here's what those words mean.
Before MCP, connecting an AI to each system was a custom-made cable — one per tool, per AI, built by hand. MCP (Model Context Protocol) is the USB standard for AI: one plug shape, so any AI can connect to any system that speaks it.
A company like SimCorp has dozens of internal systems and wants MANY AI use cases. Custom cable per pair = chaos and security holes. MCP = one governed doorway per system, with permissions controlled in one place. It's not a fancy feature; it's how AI access scales and stays auditable. Say that sentence in the interview and you're ahead of most candidates.
And you already live it: n8n's newer AI features, Claude's connectors, the tools Claude Code uses — MCP under the hood. You've been a user; now you're a speaker.
The most misunderstood part of every AI system — and where Track 1's SQL/vector module becomes engineering.
The model's only "memory" is the text currently in front of it — the context window. Think of it as a desk: whatever papers are on the desk right now, it knows; everything else in the universe, it doesn't. Memory design = deciding which papers your system puts on the desk, every single turn.
Not "how much memory?" but "what does this turn NEED on the desk?" Too little → the agent asks customers to repeat themselves (feels broken). Too much → slow, expensive, and the model gets distracted by irrelevant papers (real phenomenon — more context is not always better). Professional memory design is curation, not accumulation.
How anything AI gets made — the same eight steps whether it's a workflow, an agent, or a whole product. This is the process you've been doing blind with Claude Code; now it has names.
Thin slice first. Build the narrowest complete path — ONE call type, handled end-to-end, in production for one client — before adding breadth. A thin slice teaches more in a week than a full build teaches in a quarter, and it can start earning. Your instinct to build everything at once is the enemy; this rule is the cure.
Workflow first, agent only if forced. Start every design as a fixed workflow. Promote a step to agentic (letting the model choose) only when the paths genuinely can't be enumerated. Cheaper, more reliable, easier to debug — and when an interviewer asks "when would you use an agent?", this answer signals maturity: "as late as possible."
The professional's edge: anyone can build an AI thing; few can prove it works. Plus: what "multi-agent" really means.
Normal code is tested with exact answers: 2+2 must equal 4. AI output varies — same input, different words. So AI is tested like an employee, not a calculator: give it 50 realistic tasks, grade the results, track the score over time. That grading set is called an eval.
Evals are also your margin machine: they're what lets you safely downgrade steps to cheaper models (Track 2, C4) — because "cheaper" is only real if the score holds.
"Multi-agent" = several specialized loops handing work to each other, usually via an orchestrator — a manager agent that routes tasks: one agent reads the email, one checks the fleet, one drafts the reply, orchestrator assembles. Why split? Same reason companies have departments: a specialist with a short, focused job description outperforms a generalist with a 5-page one. Smaller prompts, clearer tools, easier testing per agent.
The honest counterweight for interviews: every handoff adds cost, latency, and a new failure point. The professional starts with ONE agent (or a workflow), splits only when one job description becomes overloaded. "Multi-agent" in a vendor pitch is a design choice to interrogate, not a quality badge to admire.
Everything an AI advisor/builder needs, and where it lives. What remains is not content.
Three tracks, ~14 hours of study, everything the four job ads ask for. From this point, more curriculum is procrastination with a syllabus. The order now: Track 1 scores → Stripe Live → weekly module + homework → portfolio → interviews. The academy's success metric was defined in C1 style: "Djoko explains any of this to a stranger, unaided, and lands interviews that test it." That number moves only when you press start.
Every grey code snippet and strange tool name from your Claude chats — wrangler, dist, push, branch, robots, secrets — decoded on your real systems: the academy, RentingPilot, jobalarm. Read cards in any order. No gate; the quiz at the bottom is for proving it to yourself.
Tap each card, read all four rows — same method as Level 0.
No gate on this track. But if you score under 5, the cards deserve a second pass.