AI Engineering, Explained by Magic. Rahul Parwal

spells learned {{ learned }} / 7

← back

AI Engineering,
Explained by Magic

Seven spells, one map. From your first wand to thinking like a system designer. No Muggle prerequisites.

Rahul Parwal

platform 9¾ · the crossing

Most of us were raised by Muggles

Harry spent eleven years with the Dursleys before he knew magic was real. Most engineers meet AI the same way.

The Muggle World

software engineering

✦Machines are literal. add(2, 3) is 5, forever.

✦Same input, same output. A test passes or it fails.

Dependable. Predictable. Entirely without magic.

The Wizarding World

ai engineering

✦Magic responds to you. Wording and context shape the result.

✦The same spell never lands the same way twice.

Non-deterministic. That one fact reshapes everything you build.

This map is how a Muggle-born learns magic properly.

node 1 · the {{ g.llm }}

The Wand Chooses the Wizard

✦Everything you build stands on one thing: the large language model, or LLM.

✦Start here, before any framework or tool. No PhD required.

Harry never chose the holly-and-phoenix wand. It chose him. Which wand fits the task sets what your system can do, what it costs, and how fast it runs.

At Ollivanders

Every wand a different core. Every model a different training.

in the real world

Dragon heartstring

code & duels

Phoenix feather

deep reasoning

{{ bm.deepseek }}GPTOpenAI

Unicorn hair

long context

Willow, ten inches

small, fast, cheap

The Elder Wand

mighty, slow, costly

whatever tops the board this month

the fitting · try before you buy

Every Ollivanders Wand, Tried

a client's brief

40,000 support tickets. Answers back in under a second. Nothing may leave their own servers.

node 2 · making it useful

A Wand Alone Is Not a Wizard

Every AI product is one LLM wrapped in three things.

tools

Owls & the Floo Network

✦The wand only knows what it learned at school. Tools are its owls: they fly out and come back with answers.

{{ g.mcp }} is the Floo Network.

knowledge

The Hogwarts Library

✦Your world is invisible to the model. {{ g.rag }} is an open-book exam, against a library you stock.

Retrieval: Madam Pince fetches the book. Generation: you answer with it open.

What you don't shelve matters too.

real world pgvector Qdrant Pinecone

memory

The Pensieve

✦Every new chat is a fresh Obliviate. Short-term is the conversation. Long-term is the vials you keep.

Semantic: facts about a person. Episodic: specific events.

real world summary buffer user profile

{{ ico.wand }}+{{ ico.owl }}+{{ ico.bookopen }}+{{ ico.pensieve }}=

an AI Agent. What must it do, know, and remember?

node 3 · proving it works

Evals Are Your O.W.L. Exams

Ask it once and it looks fine. Ask it again and you get something else. So you set a paper, and you mark it.

The Lockhart Problem

confident · charming · wrong

"Can I claim a bereavement refund after I've flown?"

one question, asked twice

run one

"Yes - just apply within 90 days of travel."

✗ invented. no such rule.

run two

"No - it has to be arranged before you fly."

✓ the actual policy.

Air Canada's chatbot gave the left-hand answer to a real passenger, who booked on it, sued, and won.

ordinary wizarding level

Mark Sheet · Refund Answers

the rubric · what every answer must do

✓ Quotes the policy it is answering from 20 / 20

✓ Says "I don't know" instead of guessing 20 / 20

✗ Never invents a rule that does not exist 0 / 20

Poor · 40 / 60

Two out of three were fine. The paper still fails - and now you know the exact line to fix.

grades

Outstanding

Exceeds

Acceptable

Poor

Dreadful

Troll

Who marks 60 answers a day?

Snape does, without mercy. LLM-as-judge: one model grades another's answers against your rubric.

Writing the rubric is the craft

What must it always do? What must it never do? The rest is bookkeeping.

real world {{ ico.claude }}{{ ico.googlegemini }}{{ ico.deepseek }} the judge is just another model

node 4 · scaling up

When One Wizard Isn't Enough

Multi-agent systems are Dumbledore's Army. Powerful, and summoned far too often.

{{ ico.sword }}Specialisation

✦One wizard juggling every skill gets sloppy.

✦Split it like the trio: Hermione researches, Ron drafts, Harry stress-tests.

real world {{ ico.langchain }} LangGraph CrewAI AutoGen

{{ ico.broom }}Parallelisation

✦Some work needn't happen one step at a time.

✦Five searches? Send five of the Army at once. Faster, and sharper.

real world fan-out map-reduce async batch

{{ ico.warning }} The trap: summoning an army the moment one agent struggles.

one wizard

1 way to fail

the whole army

summoned on reflex

{{ ico.users }} {{ ico.users }} ✗ {{ ico.users }} {{ ico.users }} {{ ico.users }}

5 agents · 4 handoffs 9 ways to fail

The Room of Requirement only appeared when they truly needed it. Apply the same test before you summon a second agent - often the fix is a better prompt, or a better eval.

node 5 · defence against the dark arts

He Who Must Not Be Named

Ship it and two things go wrong: the spell misfires, or someone bends it.

The Mirror of Erised

when the spell misfires

{{ ico.quill }}Erised → hallucination · fluent, invented

{{ ico.flame }}Amortentia → sycophancy · agrees because you pushed

{{ ico.compass }}Confundus → context drift · long threads lose instructions

{{ ico.hat }}Pure-blood prejudice → bias · the data's blind spots

Nobody attacked you. Lockhart wandered back in.

The Dark Lord's Playbook

when someone bends it

{{ ico.swirl }}Imperius → prompt injection · hidden instructions take over

{{ ico.mask }}Polyjuice → jailbreaks · rules talked around

{{ ico.eye }}Legilimency → data theft · secrets coaxed out

real world OWASP LLM Top 10

Protego · The Guardrails

what you build instead

A guardrail is a rule the model cannot talk past.

{{ ico.funnel }}Validate in, filter out

{{ ico.bookopen }}Ground and cite · no source, no claim

{{ ico.key }}Least privilege · only the wands it needs

{{ ico.users }}Human in the loop · spends, sends, refunds, deletes

{{ ico.checklist }}Audit every tool call

The model is never the last line of defence.

node 6 · life after launch

The Classroom Is Not the Battlefield

✦An AI product is built once, then improved every day in production.

✦McDonald's AI drive-thru aced the lab. Real life brought accents, traffic noise, three orders from one car.

Monitoring is your Marauder's Map. "I solemnly swear I am watching production." Every footstep, live.

The AI Ops Cycle

{{ ico.compass }} Design

{{ ico.gear }} Build

{{ ico.scroll }} Evaluate

{{ ico.rocket }} Ship

{{ ico.eye }} Monitor

{{ ico.cycle }} Feed signals back. Repeat.

The Marauder's Map

real world LangSmith Langfuse OpenTelemetry

mischief managed

no monitoring

? ? ?? ? ?? ? ?? ? ?? ? ?

answer ✓ and nothing else

i solemnly swear

traced, every footstep

{{ ico.quill }}prompt {{ ico.bookopen }}retrieval {{ ico.wand }}model {{ ico.owl }}tool call {{ ico.scroll }}answer

answer ✓ wrong tool · 3 retries · 4.2 s

node 7 · thinking

"Books! And Cleverness!
There Are More Important Things."

- Hermione Granger

Anyone can spin up an agent in thirty minutes. The skill that compounds is system design, a grand name for thinking.

{{ ico.wand }}Which wand?

{{ ico.owl }}Which owls, books, vials?

{{ ico.scroll }}What does good mean here?

{{ ico.sword }}Do you truly need an army?

That judgment separates wizards from wand-owners.

mischief managed

The Whole Map, Assembled

{{ ico.wand }}1 · The Wand LLM

2 · The Kit {{ ico.owl }}Tools {{ ico.bookopen }}RAG {{ ico.pensieve }}Memory

{{ ico.scroll }}3 · O.W.L.s Evals

{{ ico.bolt }}The AI Agent

{{ ico.users }}4 · The Army Multi-agent

{{ ico.snake }}5 · Defence Syndromes · Attacks · Guardrails

{{ ico.map }}6 · Marauder's Map Production

{{ ico.book }}7 · Thinking System Design

the same map, in muggle tools {{ bm.claude }}{{ bm.googlegemini }}{{ bm.meta }}{{ bm.mistralai }}{{ bm.deepseek }}{{ bm.qwen }}{{ bm.ollama }}{{ bm.modelcontextprotocol }}{{ bm.langchain }} the names churn. the map doesn't.

Stop memorising roadmaps. Learn the map once, and you'll know how to build with whatever drops next.

Rahul Parwal