Artificial intelligence is software that performs tasks normally requiring human judgement — recognising images, understanding language, making predictions. Modern AI mostly means machine learning: systems that learn patterns from data rather than following rules written by hand.
Fig. 01 — Artificial Intelligence
Swipe sideways to see the whole diagram →
Step by step
Collect data
A model is only as good as its training data. Biased or thin data produces a biased or unreliable model.
Choose a model
The architecture defines what patterns the system can represent — from simple regression to a transformer with billions of parameters.
Train
The model makes predictions, measures its error against known answers, and adjusts its internal weights to reduce that error. Repeat millions of times.
Validate
Performance is tested on data the model never saw. A model that memorised the training set but fails here is overfitted.
Infer
The trained model is deployed and makes predictions on new inputs. This is what happens every time you use an AI feature.
Monitor and retrain
Real-world data drifts. Models degrade quietly over time unless they are measured and refreshed.
Real-world example
An email spam filter was never given a list of spam words by a programmer. It was shown millions of labelled emails and worked out the patterns itself — which is also why it can be fooled by spam written unlike anything in its training data.
Where people get this wrong
AI is either treated as magic or dismissed as autocomplete. Both skip the mechanism. A model adjusts internal weights to reduce error on examples it was shown, which explains both its impressive results and its confident mistakes on anything unlike its training data.
Common interview questions
What is the difference between AI, machine learning and deep learning?
AI is the whole field. Machine learning is the subset that learns from data. Deep learning is the subset of ML using multi-layer neural networks.
What is overfitting?
A model that has memorised its training data, performing excellently there and poorly on anything new.
What is the difference between supervised and unsupervised learning?
Supervised learning trains on labelled examples. Unsupervised learning finds structure in unlabelled data, such as clustering customers by behaviour.
Does AI actually understand what it produces?
Current systems model statistical relationships in their training data extremely well. That produces useful and often impressive output, but it is not comprehension in the human sense.
Inside the Druvexaa simulation
The simulation trains a small model in front of you, showing the error falling over successive passes, then feeds it an input unlike anything in its training set so you can watch it fail confidently.
Machine learning builds systems that improve at a task by finding patterns in data, rather than by following rules a programmer wrote. The rules are discovered from examples instead of specified in advance.
Fig. 02 — Machine Learning
Swipe sideways to see the whole diagram →
Step by step
Supervised learning
Training on labelled examples. Given enough photos marked cat or dog, the model learns the boundary between them.
Unsupervised learning
No labels. The model finds structure on its own — grouping customers by behaviour, for instance, without being told the groups.
Reinforcement learning
The model acts, receives a reward or penalty, and adjusts. This is how game-playing and robotics systems are trained.
Features are the input
The measurable properties the model actually sees. Choosing good features has historically mattered more than choosing the algorithm.
Training minimises error
The model predicts, the error is measured against the known answer, and internal weights shift to reduce it. Repeated across the dataset many times.
Validation is the honest test
Performance is measured on data held back from training. A model that scores well in training and poorly here has memorised rather than learned.
Real-world example
A bank's fraud model was never given a list of fraud rules. It was shown millions of labelled transactions and learned which combinations of amount, location, time and merchant pattern correlate with fraud — including patterns no analyst had noticed.
Where people get this wrong
Model choice is treated as the important decision. In practice, data quality determines almost everything. A simple model on clean, representative data reliably beats a sophisticated model on biased or thin data — and a model trained on data that under-represents a group will fail that group confidently, which is where most real-world harm comes from.
Common interview questions
What is the difference between supervised and unsupervised learning?
Supervised learning trains on labelled examples with known answers. Unsupervised learning finds structure in unlabelled data.
What is overfitting and how do you detect it?
The model has memorised training data rather than learning general patterns. It shows as high training accuracy alongside poor validation accuracy.
Why is a separate validation set necessary?
Measuring on training data tells you what the model memorised. Only unseen data reveals whether it generalises.
What is a feature?
An individual measurable property used as model input. Feature selection strongly influences what the model is able to learn.
Inside the Druvexaa simulation
The AI simulation trains a small classifier in front of you with the error curve falling in real time, then lets you deliberately skew the training set and watch the same architecture produce confidently wrong answers on the group you removed.
Deep learning uses neural networks with many layers to learn features automatically from raw data. Each layer builds a slightly more abstract representation than the one before it, which is why it works so well on images, audio and language.
Fig. 03 — Deep Learning
Swipe sideways to see the whole diagram →
Step by step
Neurons and weights
Each neuron takes inputs, multiplies them by learned weights, sums them, and passes the result through an activation function.
Layers build abstraction
Early layers detect edges. Middle layers combine edges into shapes. Later layers combine shapes into recognisable objects.
Forward pass
Data flows through the network and produces a prediction.
Loss measurement
The prediction is compared against the correct answer and the difference is quantified as a loss.
Backpropagation
The loss is propagated backwards, and every weight is nudged in the direction that would have reduced the error.
Repeat at scale
Millions of iterations across large datasets, which is why GPUs matter — the operations are matrix multiplications that parallelise extremely well.
Real-world example
A network trained to recognise faces was never told what a nose is. Inspecting its middle layers shows detectors for edges, then curves, then eye and nose shapes — features it constructed on its own because they helped reduce the error.
Where people get this wrong
Depth is treated as the goal, so more layers is assumed to be better. Beyond a point, extra layers make a network harder to train and more likely to memorise the training set. The reason deep learning displaced earlier methods is not depth itself but that it removed manual feature engineering — the network learns what to look at, which is a different claim from deeper being stronger.
Common interview questions
What is backpropagation?
The algorithm that computes how much each weight contributed to the error and adjusts it accordingly, working backwards from the output layer.
Why are activation functions necessary?
Without a non-linear activation, stacked layers collapse mathematically into a single linear transformation and depth adds nothing.
Why are GPUs used for deep learning?
Training is dominated by matrix operations, which GPUs execute in parallel across thousands of cores far faster than a CPU.
What is the vanishing gradient problem?
Gradients shrink as they propagate back through many layers, so early layers barely update. Techniques such as ReLU activations and residual connections address it.
Inside the Druvexaa simulation
The AI simulation visualises activations layer by layer as an image passes through, so you can watch early layers respond to edges and later layers respond to whole objects — the abstraction hierarchy shown rather than described.
Generative AI produces new content — text, images, audio or code — rather than classifying existing content. Language models do this by repeatedly predicting the most plausible next token given everything before it.
Fig. 04 — Generative AI
Swipe sideways to see the whole diagram →
Step by step
Tokenisation
Text is split into tokens — roughly word fragments. The model works with token numbers, not letters.
Context window
The model sees a fixed maximum span of tokens at once. Anything beyond it is invisible, which is why very long conversations lose earlier detail.
Next-token prediction
The model outputs a probability distribution over possible next tokens and samples from it.
Sampling settings
Temperature controls how adventurously it samples. Low values give predictable output; high values give varied and less reliable output.
Iteration
The chosen token is appended to the input and the process repeats, which is why output appears word by word.
Images work differently
Diffusion models start from noise and remove it step by step, guided by the prompt, until an image emerges.
Real-world example
Ask a model for a summary and it does not retrieve one. It generates a plausible sequence of tokens conditioned on your text — which explains both the fluency and the fact that a confidently stated detail can be entirely invented.
Where people get this wrong
Fluency is read as knowledge. The model optimises for plausible continuation, not for truth, and a fabricated citation is generated by exactly the same mechanism as a correct one — with the same confident tone. This is why generative output needs verification for anything factual, and why the errors are hardest to catch precisely where the writing is best.
Common interview questions
What is a hallucination in a language model?
Confidently stated output that is not grounded in fact. It arises because the model optimises for plausibility rather than accuracy.
What is a context window?
The maximum number of tokens a model can consider at once. Content outside it cannot influence the output.
What does temperature control?
How randomly the model samples from its probability distribution. Lower values produce more deterministic output.
How is generative AI different from a search engine?
A search engine retrieves existing documents. A generative model produces new text that may or may not correspond to any real source.
Inside the Druvexaa simulation
The AI simulation shows a prompt being tokenised and then generates the response one token at a time with the probability distribution visible at each step, so the mechanism behind fluent-sounding output is exposed rather than hidden.
Agentic AI describes systems that pursue a goal across multiple steps, deciding what to do next and using tools to do it. Where a generative model answers, an agent plans, acts, observes the result and adjusts.
Fig. 05 — Agentic AI
Swipe sideways to see the whole diagram →
Step by step
A goal instead of a prompt
The instruction describes an outcome rather than a single output — book the trip, reconcile the report, fix the failing test.
Decomposition
The agent breaks the goal into steps it can actually perform.
Tool use
It calls external capabilities — a search, a database query, a file operation, an API — instead of only producing text.
Observation
The result of each action feeds back in. A failed step is information, not an ending.
Iteration
The loop repeats, with the plan revised as results arrive, until the goal is met or a limit is reached.
Guardrails
Because an agent takes real actions, it needs limits: approval before irreversible steps, a cap on iterations, and clear scope over what it may touch.
Real-world example
Asked to summarise last quarter's support tickets, a generative model can only work with text you paste in. An agent queries the ticket system, notices the date filter returned nothing, corrects the query, pulls the records, and then writes the summary.
Where people get this wrong
Agentic is used as a synonym for more capable, which misses what actually changed. The difference is that the system takes actions with consequences. A generative model that is wrong produces a bad paragraph; an agent that is wrong sends the email, deletes the file, or makes the purchase. Capability is the smaller half of the story — the important half is that errors now have side effects.
Common interview questions
What distinguishes an agent from a chatbot?
An agent takes actions through tools across multiple steps and adapts based on results. A chatbot produces a response to a prompt.
What is tool use?
Allowing a model to call external functions — searches, APIs, file operations — so it can gather information and change state rather than only generate text.
Why do agents need guardrails?
Because their errors have real consequences. Approval gates, iteration limits and scoped permissions contain the damage from a wrong decision.
What is human-in-the-loop?
A design where the agent pauses for human approval before consequential steps, keeping autonomy for reversible actions only.
Inside the Druvexaa simulation
The agentic simulation traces one goal through the full loop, showing each planned step, the tool called, the result returned, and the point where the plan is revised after a failure — including the approval gate before any irreversible action.
These topics are written to the same depth. A dedicated Druvexaa simulation is planned for each — until then, every one links to the closest existing simulation and quiz.
Natural Language Processing is the field concerned with making machines work with human language — understanding it, generating it, translating it, and extracting meaning from it.
Fig. 06 — Natural Language Processing
Swipe sideways to see the whole diagram →
Step by step
Tokenisation
Text is split into words or sub-word pieces, because models operate on discrete units rather than raw characters.
Embeddings
Each token becomes a vector of numbers positioned so that related meanings sit close together.
Context matters
Modern models read the whole sentence at once, so bank in river bank and bank account produce different representations.
Common tasks
Sentiment analysis, named entity recognition, translation, summarisation, question answering and classification.
Why it is hard
Language is ambiguous, contextual and culturally loaded. Sarcasm, idiom and implied meaning defeat purely statistical approaches regularly.
Real-world example
A support system routing tickets does not match keywords. It embeds each message and compares meaning, so 'I cannot log in' and 'password not accepted' land in the same queue despite sharing no words.
Where people get this wrong
Understanding is inferred from correct output. The model has learned statistical relationships between tokens, which produces genuinely useful results and also produces confident nonsense on inputs unlike its training data. Treating fluency as comprehension is how people end up trusting an answer that was never grounded in anything.
Common interview questions
What is an embedding?
A numerical vector representing a token or piece of text, arranged so that semantically similar items are close together in that space.
Why is tokenisation necessary?
Models operate on discrete units with fixed vocabularies. Tokenisation converts arbitrary text into that fixed set.
What is named entity recognition?
Identifying and classifying spans of text as people, places, organisations, dates and similar categories.
Inside the Druvexaa simulation
No dedicated simulation yet. The AI simulation shows tokenisation and the prediction step directly, which is the mechanism underneath every task described here.
Computer vision is the field concerned with extracting meaning from images and video — recognising what is in a picture, where it is, and what is happening across frames.
Fig. 07 — Computer Vision
Swipe sideways to see the whole diagram →
Step by step
Images are numbers
A photograph is a grid of pixel values. Everything a vision model does starts from that grid.
Convolution finds patterns
A small filter slides across the image looking for a specific pattern — an edge, a corner, a texture — and produces a feature map.
Layers build complexity
Early filters find edges. Later layers combine those into shapes, then into recognisable objects.
Pooling reduces size
Downsampling keeps the strongest signals and discards precise position, which makes the model tolerant to small shifts.
Task determines the output
Classification produces one label, detection produces boxes, segmentation produces a per-pixel mask.
Data is the bottleneck
Vision models need large, varied, well-labelled datasets. Performance usually improves more from better data than from a better architecture.
Real-world example
A camera reading number plates runs detection to locate the plate, then character recognition on that crop. Splitting the problem this way is why it works in varied lighting where a single end-to-end guess would fail.
Where people get this wrong
High benchmark accuracy is read as reliability. A model trained largely on clear daytime images can collapse at dusk or in rain while still reporting high confidence. Vision systems fail most dangerously on conditions that were under-represented in training, and the confidence score gives no warning.
Common interview questions
What is a convolution in this context?
A small filter applied across the image to detect a local pattern, producing a feature map of where that pattern occurs.
What is the difference between detection and segmentation?
Detection returns bounding boxes around objects. Segmentation labels every pixel, giving exact object shapes.
Why is data augmentation used?
Rotating, cropping and adjusting brightness creates variation that makes the model robust to conditions it would otherwise never see.
Inside the Druvexaa simulation
No dedicated simulation yet. The AI simulation visualises activations layer by layer, which is exactly the edge-to-shape-to-object progression that convolutional vision models build.
Prompt engineering is the practice of writing input that reliably gets useful output from a language model. It matters because the same question phrased two ways can produce very different quality.
Fig. 08 — Prompt Engineering
Swipe sideways to see the whole diagram →
Step by step
Be specific
Vague instructions produce vague output. Naming the audience, length and purpose narrows the space the model is sampling from.
Zero-shot and few-shot
Zero-shot means asking directly. Few-shot means including two or three worked examples, which usually improves consistency dramatically.
Ask for reasoning
Requesting step-by-step working improves accuracy on multi-step problems, because intermediate steps become part of the context.
Specify the format
Asking for a table, JSON or a fixed structure removes ambiguity and makes output usable by other software.
Give constraints, not just goals
Stating what to avoid is often more effective than adding more description of what you want.
Iterate
Treat a prompt as something to test and refine against real cases, not something to get right in one attempt.
Real-world example
Asking for 'a summary' returns something generic. Asking for 'a four-line summary for a non-technical manager, no jargon, ending with the decision needed' returns something usable — the same model, a different instruction.
Where people get this wrong
Prompting is treated as a trick to unlock hidden ability. It does not add knowledge the model lacks. What good prompting does is remove ambiguity about the task, which is why it helps most on format and consistency and helps least on facts the model never learned.
Common interview questions
What is few-shot prompting?
Including a small number of worked examples in the prompt so the model infers the pattern and format expected.
Why does asking for step-by-step reasoning help?
Intermediate steps enter the context and condition later predictions, which improves accuracy on multi-step problems.
Can prompting fix factual gaps?
No. It clarifies the task but cannot supply knowledge the model does not have. Retrieval is the tool for that.
Inside the Druvexaa simulation
No dedicated simulation yet. The AI simulation shows how the input tokens condition each prediction, which is the mechanism explaining why phrasing changes the output so much.
RAG connects a language model to a searchable source of documents. Relevant passages are retrieved first and placed in the prompt, so the answer is grounded in real material rather than produced from memory alone.
Fig. 09 — Retrieval Augmented Generation (RAG)
Swipe sideways to see the whole diagram →
Step by step
Documents are chunked
Source material is split into passages small enough to be useful and large enough to keep meaning intact.
Chunks are embedded
Each passage becomes a vector and is stored in a vector database.
The question is embedded too
The query is converted the same way, so it can be compared against stored passages by meaning rather than keyword.
Nearest passages are retrieved
The most semantically similar chunks are pulled back — typically the top three to ten.
The prompt is augmented
Those passages are inserted alongside the question, giving the model the material it needs in front of it.
The answer is generated
Because the source text is in context, answers can cite it and are far less likely to be invented.
Real-world example
A company assistant answering policy questions does not have the handbook in its training data. RAG retrieves the two relevant paragraphs and the model answers from those — which is also why it can quote the exact clause.
Where people get this wrong
RAG is assumed to eliminate hallucination. It reduces it substantially but does not remove it. If retrieval returns the wrong passages, or none, the model will often still produce a confident answer from its own weights. Retrieval quality — chunking, embeddings, ranking — is where most RAG systems actually succeed or fail, not the model.
Common interview questions
Why use RAG instead of fine-tuning?
RAG updates instantly when documents change, cites sources, and costs far less than retraining. Fine-tuning suits changing style or behaviour, not changing facts.
What is a vector database?
A store optimised for finding the nearest embeddings to a query vector, enabling search by meaning rather than exact keywords.
What is chunking and why does it matter?
Splitting documents into passages. Chunks that are too small lose context and too large dilute relevance, so it strongly affects retrieval quality.
Inside the Druvexaa simulation
No dedicated simulation yet. The AI simulation demonstrates how added context changes the model's predictions, which is precisely the effect retrieval is engineering deliberately.