Learn AI
The Qirava blog
Articles with voice and rigor on artificial intelligence, software quality, and method. Pick a section and start reading.
Fundamentals
12 articles
What is a token? The LEGO brick AI uses to build language
You type "cat" and assume the model reads the word. It does not. It reads chips. Here I show you which ones, why, and why that hits your wallet.
What embeddings are: turning words into coordinates
To a machine, 'king' and 'queen' aren't letters: they're points on a map. Here you'll see how that map gets drawn.
What a neural network is, explained as a chain of decisions
There are no real neurons in there. There are numbers tossing the ball around. I will sketch it out for you, layer by layer.
The context window: why a brilliant AI forgets too
The models that today write, summarize and code have an astonishingly short working memory. Understanding its size (and its shape) is understanding why AI sometimes loses the thread, and what to do about it.
AI, machine learning and deep learning: which is which
People treat them as synonyms, but they are not. They are three circles, one inside the other. Here it becomes clear which one contains which.
What artificial intelligence really is, explained without the jargon
Long before the term existed, the question did: can a machine do what we thought only we could? Here is the answer, no smoke and mirrors.
Show more · Fundamentals
What machine learning is for people who do not code
The machine is not thinking. It is tuning thousands of knobs until it gets things right. I will show you how, with an example you can follow in your head.
The six levels of context in AI: the ladder that decides whether your assistant remembers or improvises
There is no single way to "give context" to a model. There is a ladder with six rungs, from a loose prompt to a vector database. Here I show you which one you need and why climbing too high also has a price.
The machine that read too much
A large language model knows nothing in the way you know things. And yet something happened when we fed it nearly everything we have ever written. This is the story of what that thing is, with no mysticism and no smoke, and of why predicting the next word turned out to be far deeper than it seemed.
The vector space: how meaning becomes geometry
Before a machine can search by meaning, someone had to turn words into coordinates. This is the story of how meaning took on shape, direction and distance.
What is an AI agent and why it is not the same as a chatbot
A chatbot answers. An AI agent pursues a goal: it plans, uses tools, acts in steps and verifies. This is the difference that matters.
Types of artificial intelligence: narrow, general and the one that does not exist
The AI you use today is a specialist, not a universal genius. Telling the types apart saves you from misplaced fears and inflated expectations.
How it works inside
8 articles
How a neural network learns: the error that travels backward
For a machine, learning means getting it wrong and adjusting course. The method has an ugly name and a very elegant idea behind it.
What RLHF Is: How a Model Learns Its Manners
A raw model says useful things and terrible things in the same tone. RLHF is the upbringing that teaches it to prefer the useful ones.
Why AI hallucinates: it invents facts with total confidence
It hands you a quote, an author, a date. All false, all delivered with poise. It does not lie: it predicts. And sometimes the prediction is pure invention.
What a model's 'temperature' is and how it changes its answers
Raise the temperature and the AI turns creative and daring. Lower it and it becomes predictable. A dial almost nobody knows exists.
How a language model is trained, from scratch to assistant
First it reads half the internet. Then it gets taught manners. Two very different stages, and mixing them up breeds misunderstandings.
What is RAG and what is a vault: how to anchor an AI to YOUR documents
What is RAG, why an AI alone makes things up, and how a private vault with retrieval ties it to your real documents, not its blurry memory.
Show more · How it works inside
What quantization is: how to shrink a model without breaking it
A huge model will not fit on your laptop. Quantization squeezes it by rounding its numbers. How far you can squeeze before you lose it.
What are a model’s parameters (and why everyone talks about billions)
When you hear '70 billion parameters', what does that mean? The knobs the model tuned to get things right. Here is what they are.
Prompting and practice
9 articles
How to transcribe WhatsApp voice notes to text, step by step
Transcribe WhatsApp voice notes to text so you can search, summarize and stop replaying seven minute audios. Honest guide, with privacy.
How to turn a YouTube video into text, step by step
Turning a YouTube video to text helps you study, quote and summarize. The difference between auto captions and a clean transcript.
Few-shot: teach the AI with examples instead of instructions
Sometimes explaining the task is hard, but showing two examples makes everything clear. AI learns the way we do: by watching.
How to transcribe a meeting to text with AI, step by step
How to transcribe a meeting with AI: record or upload the audio, get text and Markdown with speakers and decisions, and what to do next.
How to use AI to write better without sounding like AI
Text straight from the factory sounds flat and repetitive. With the right instructions, AI edits instead of flattening.
What prompt engineering is and why it is not magic
They gave an engineering name to something closer to knowing how to ask. Useful, yes; mysterious, no. Let me bring the term back down to earth.
Show more · Prompting and practice
Asking AI to 'think step by step,' and why it sharpens the answers
A model in a hurry botches simple arithmetic. Force it to reason out loud and it grows far more accurate. The reason is a curious one.
How to write a good prompt: the structure that almost always works
The difference between a mediocre answer and a useful one is not the model. It is how you asked. Here is the recipe.
How to convert a PDF or Word file to Markdown to use with AI
Converting PDF to Markdown is not a whim: an AI understands structure far better than a raw PDF. Here is the how, step by step.
Training
1 articles
QA
26 articlesUnit tests: what they are and how to write your first one
The smallest unit of trust in your code. If you cannot test a function in isolation, you probably do not fully understand it either.
The four test levels and why their order matters
Unit, integration, system and acceptance are not the same test in different sizes: each level answers a different risk. Skipping one does not save time, it defers it.
Code coverage: the metric that lies hardest right when you trust it most
80% coverage does not mean 80% tested. You can run every line and verify nothing. Here is what coverage actually measures and why chasing it as a goal backfires.
Exploratory testing: when the script never finds the strange bug
Scripts test what you already imagined. The bugs that hurt live where nobody thought to look, and those are hunted down by a human exploring with judgment.
What software testing really is: the guide that separates signal from noise
Testing is not clicking around until something breaks. It is designing experiments that answer one concrete question: does this do what we promised? Here is the full map, no filler.
What is QA (Quality Assurance) in software, made simple
Quality is not "having no bugs". It is having a way to know, before you ship, whether the software does what you promised. That is QA, and it is not the same as testing at the end.
Show more · QA
ISO/IEC 25010: the map of what “quality software” actually means
When someone says “quality” without defining it, arguing is pointless. ISO 25010 gives names to the eight characteristics you can actually demand and measure in a product.
Functional vs non functional testing: what each one measures
A button working is functional. That button holding up under 5,000 simultaneous users is non functional. A product can pass every functional check and still be unusable in production.
Integration testing: when the parts pass alone but fail together
Every module passes its unit tests and the system still breaks. The bug lives in the seam between components, and only integration testing catches it there.
Performance testing: load, stress and spikes explained
It works perfectly with three users. The question is what happens on launch day with thirty thousand. You do not guess that: you measure it.
How to report a bug the developer will actually want to fix
"It doesn't work" is not a report, it's a complaint. A good defect report is reproducible, states expected versus actual, and cuts the fix time in half.
Verification vs validation: the question almost nobody knows how to answer
'Are we building the product right?' and 'are we building the right product?' are two different questions. Confusing them is why so many teams pass every test and still fail.
The life of a bug: from 'new' to 'closed' without losing it along the way
A defect is not fixed, it is managed. New, assigned, resolved, verified, reopened: each state exists for a reason, and skipping them is how 'fixed' bugs come back.
QA vs testing vs quality control: they are not the same
People use them as synonyms, but they are not. QA is the process that prevents, testing is the activity that verifies, and QC is the inspection that detects. Mixing them up leads you to hire wrong and measure worse.
What ISTQB Is and Whether the Certification Is Worth It for Testers
It is the de facto standard for talking about testing with a shared vocabulary. It will not make you a good tester, but it gives you the language the industry speaks.
The testing pyramid: why not everything should be E2E
Many teams have the pyramid upside down: mountains of slow, brittle E2E tests and almost no unit tests. The result is a suite that takes hours and that nobody trusts.
Line, branch and condition coverage: they are not all worth the same
Covering lines is easy and deceptive. Covering branches and conditions is where the real bugs hide. The technical difference that changes how much you trust your number.
Test case, scenario and script: three things that are not the same
Sloppy vocabulary breeds sloppy suites. Telling case, scenario and automation script apart is what keeps your suite readable a year from now.
End-to-end (E2E) tests: exercising the user’s entire journey
They mimic a real person doing the task from start to finish. They are the most realistic and the most expensive to maintain, which is exactly why they should not be the majority.
Regression testing: making sure fixing A does not break B
The most innocent change revives a bug you thought was dead. Regression testing is the safety net that warns you when the past breaks again.
Severity vs priority: why a critical bug can wait
Severity is how much it hurts; priority is when you fix it. Confusing the two makes teams put out small fires while the house burns down. Here is the matrix that sorts it out.
Black box vs white box: two ways of looking at the same code
One tests what the software does; the other tests how it does it. They do not compete: they complete each other. Picking just one leaves you with half the risk map.
Smoke testing: does the build even power on?
Before you test in depth, make sure the build is not dead on arrival. Five minutes of smoke testing save hours spent on something that never boots.
How to design test cases that actually catch bugs
Equivalence classes, boundary values, decision tables: techniques that turn 'what do I test?' into a repeatable method instead of intuition that runs dry.
The test plan: the document almost nobody reads and everybody needs
A test plan is not bureaucracy: it is the written answer to what you will test, what you will not, the criterion that tells you to stop and the risk you accept. Without it, you test blind.
Acceptance testing (UAT): when the decider is not the tester
Code can pass everything and still not be acceptable. UAT is the moment when the business, not QA, says 'this works'. And it is almost always done badly.
Method
8 articlesHuman in the loop: why the best AI keeps people inside
Human in the loop is not distrust of AI. It is designing when it decides alone and when a person reviews, so the result gains value without losing control.
What an AI workflow is and how to build one
An AI workflow is not a loose chat: it is a chain of steps where the model does some and people do others. Here is how you build it.
How to run an AI company with governed agents
How to run an AI company with governed agents: strategy from the chat, execution by a swarm, governance that lends judgment.
Deliberating without faking execution: governing agents under SDD
An agent that proposes is not an agent that did. Confusing the two is what fills your systems with work that looks finished but is not. Here I show you the exact line.
What Spec-Driven Development Is (and Why We Left Vibe Coding Behind)
You asked the AI to "see what comes out" and something came out. The problem is not what came out: it is that nobody wrote down what should have come out. The spec fixes that, not the prompt.
The Qirava factory: roles, skills and maturity gates (G0 to G3)
A lone agent improvises. A factory does not. Here I show you how a line of governed agents is organized: each role with its own package of skills, and nothing moves forward without clearing its gate.
Show more · Method
What is Qirava: a factory of services, not an app
What is Qirava, in plain terms: not one feature, but a factory that blends AI layers and people to deliver services that solve.
From prompt to spec: how a specification governs your agents
A prompt asks. A spec governs. Here you will see why a contract with acceptance criteria makes agents perform better and stop improvising.
Costs and business
6 articles
What is a payment gateway (and why you never store the card)
What a payment gateway is, what it really does when someone buys from you online and why the card number should never touch your server.
How to automate tasks with AI in a small business
You do not need a data department. With the right repetitive tasks, AI pays for itself in weeks. Where to begin.
What an AI API is and why you will end up needing one
The chat is for you; the API is for your product. It is the difference between using AI and building with it. The idea, no code required.
What is KYC: why a platform asks you to verify your identity
What is KYC and why it exists. Verifying your identity is not distrust: it is what keeps your money from getting mixed with someone else’s.
How to measure the return on investment of AI (without fooling yourself)
Saving time sounds good until you have to put it into money. A sober method for knowing whether AI pays off or just entertains.
AI service vs tool: what your company should actually buy
AI service vs tool is not the same choice: one you operate, the other delivers the result. How to choose without overpaying.
Ethics and bias
6 articlesThe real risks of AI: telling the urgent apart from the hypothetical
Between panic and denial there is a reasonable middle ground. These are the risks that already exist and the ones that are still speculation.
AI regulation, explained for people who are not lawyers
Europe classifies it by risk; others move more slowly. No legal jargon: what is changing and what you should watch as a user or a company.
Privacy and artificial intelligence: what happens to what you type
Every prompt is a piece of data that leaves your home. Sometimes it trains the model, sometimes it does not. Knowing the difference protects your information.
What AI bias really is and where it comes from
A model is not born prejudiced: it inherits its bias from the data. Knowing where that bias comes from is the first step to keep from amplifying it.
Algorithmic bias: three real cases that came at a steep price
These are not laboratory hypotheses. Production systems made unfair decisions at scale. What happened and what we learned.
What AI alignment is and why it worries its creators
A highly capable system chasing the wrong goal is a problem. Alignment is getting it to want what we want. It is not trivial.
History
2 articles
How ChatGPT Was Born: The Chain of Ideas That Made It Possible
It did not appear out of nowhere in 2022. Behind it lie a decade of papers, expensive bets and one unexpected turn. The short history of a phenomenon.
A brief history of AI: from Turing to the transformers
None of the machines that talk to us today would exist without an uncomfortable question posed in 1950, an optimistic summer in 1956 and two winters that nearly buried it all. This is the story of that stubbornness.