/ BLOGS

MY NOTES

Notes, experiments, and

things worth knowing.

A 3D diagram of an AI agent: a central robot on a glowing platform wired in a loop to its labelled capabilities — Planning, Memory, Tools, Action, Feedback, Data, and Goals — the model-in-a-loop-with-tools shape the article describes.

Agentic AI, demystified: it's tools and a loop

"Agent" is the most overloaded word in AI right now. Strip the hype and it's something an engineer can reason about: a model in a loop that can call tools and read the results.

Aug 26, 2026·2 min read

Read article→

◆ All articles

A learning roadmap laid out as a path: mathematical foundations, programming, classical machine learning, deep learning, and specialisation, with hands-on projects running alongside every stage.

Career & Learning

Aug 28, 2026·4 min read

A roadmap for learning AI

There is no shortage of AI courses; there is a shortage of a coherent order to learn them in. A grounded roadmap — math foundations, programming, classical ML, deep learning, then specialisation — built around projects over tutorials.

CareerLearningAI

A layered application-security shield showing defensive controls stacked together: input validation, authentication and access control, encryption, dependency scanning, and logging and monitoring.

Cybersecurity

Aug 28, 2026·6 min read

Application security: a defender's checklist

Most breaches exploit a short list of well-understood weaknesses. This is a defender's tour of that list — access control, injection, misconfiguration, dependencies, secrets — and the concrete habits that shut each of them down.

CybersecurityAppSecWeb Security

Research

Aug 28, 2026·4 min read

How to read an AI research paper

Research papers are dense by design, and reading them front to back is the slowest way in. The three-pass method, what to look for in each pass, and how to find the contribution and the limitations without drowning in detail.

ResearchLearningPapers

A landscape of AI hardware: a data-centre GPU and a tensor accelerator on one side, and a small edge device with an on-chip neural processing unit on the other, connected by the flow of models from cloud to edge.

Future Tech

Aug 27, 2026·5 min read

AI hardware and the edge: why silicon shapes what's possible

Every leap in AI has ridden a leap in hardware. GPUs, TPUs, and NPUs are not interchangeable boxes — their design decides what models are feasible and where they can run. A grounded look at the silicon underneath the models.

Future TechAI HardwareEdge

A smartphone running a neural network locally on its own chip, with no connection to the cloud, illustrating on-device AI inference for privacy, low latency, and offline use.

Aug 26, 2026·4 min read

On-device AI: why the model is moving to the edge

More and more AI runs directly on your phone, laptop, or watch instead of a distant server. Here is why on-device inference is worth the trouble, the techniques that make it fit, and the trade-offs you accept in return.

AIEdge ComputingOn-Device

Two line charts rising together to suggest a causal link, with a hidden confounding variable revealed underneath as the true common cause of both.

Data Science

Aug 25, 2026·4 min read

Correlation is not causation — and why the difference matters

Two things moving together does not mean one causes the other. That single confusion drives bad decisions in business, science, and machine learning. Here is why it happens, the traps that create it, and how to think straight.

Data ScienceStatisticsCausal Inference

The retrieval-augmented generation pipeline: a user query drives retrieval from a vector store / knowledge base, whose returned documents ground an LLM's generation to produce the final answer. A band along the bottom lists the payoffs — relevant context, reduced hallucinations, up-to-date knowledge, transparency, and better answers.

Aug 24, 2026·2 min read

RAG without the hype: retrieval is a search problem

Retrieval-augmented generation gets discussed as if it were a new kind of model. It isn't — it's a pattern, and most of the quality comes from treating it as a search problem.

AILLMsRetrieval

A diagram of data leakage: information from the test set leaking backwards into training and preprocessing, producing an inflated evaluation score that collapses in production.

Machine Learning

Aug 24, 2026·4 min read

Data leakage: the bug that fakes success

The most dangerous bug in machine learning does not throw an error — it makes your model look brilliant in testing and then fail in the real world. Here is what data leakage is, the forms it takes, and how to design it out.

Machine LearningEvaluationPitfalls

A machine-learning pipeline diagram: raw data ingested, cleaned, and transformed into features, split into train and test sets, used to train and evaluate a model, then deployed and monitored with a feedback loop.

Data Science

Aug 22, 2026·4 min read

The machine learning data pipeline, end to end

A model is a small part of a real ML system; the pipeline around it — how data is collected, cleaned, split, fed in, evaluated, and monitored — is where projects succeed or quietly rot. A tour of the stages and their traps.

Data ScienceMLOpsPipelines

A collage of core computer-science ideas: a Big-O growth chart, a hash table, a binary tree and a graph, a call stack, and stack-versus-heap memory regions.

General Computing

Aug 21, 2026·5 min read

Computer science fundamentals that compound

Frameworks churn every few years, but a handful of fundamentals keep paying off for a whole career: complexity, data structures, recursion, memory, and concurrency. Here is why they compound — and the mental models that stick.

Computer ScienceFundamentalsAlgorithms

The compiler pipeline: source code flowing through a lexer into tokens, a parser into an abstract syntax tree, semantic analysis, an intermediate representation, an optimizer, and a code generator producing machine code.

Programming

Aug 20, 2026·5 min read

How a compiler works: from source text to running code

A compiler turns text you can read into instructions a machine can run, and it does it in a pipeline of well-defined stages. Understanding those stages demystifies error messages, performance, and language design all at once.

ProgrammingCompilersLanguages

A distributed system diagram: a load balancer fanning requests to replicated application servers, a cache layer, and a partitioned, replicated database, with redundancy at every tier.

System Design

Aug 19, 2026·5 min read

System design: building for scale and failure

Scalable systems are not built by adding servers; they are built by assuming every part will eventually fail and designing so the whole keeps working. A tour of the core ideas — scaling, load balancing, caching, replication, partitioning, and the consistency trade-offs underneath them.

System DesignScalabilityReliability

An isometric map of the modern AI technology landscape: a connected globe ringed by a neural-network brain, server racks, a robotic arm, an AI accelerator chip, cloud infrastructure, a developer laptop, and analytics dashboards.

Aug 18, 2026·2 min read

The Claude 5 family — and what a tier like Fable 5 means for builders

A new model generation reshuffles the quality / latency / cost trade-off. Here's the way I actually pick a tier — and where a fast, lightweight model like Fable 5 earns its place.

AILLMsModel selection

A web-performance dashboard: a page-load waterfall of network requests, a rendering timeline, and the three Core Web Vitals gauges for loading, interactivity, and visual stability.

Web Development

Aug 18, 2026·4 min read

Web performance: from request to paint

Fast sites are not an accident — they come from understanding what happens between a request and the first useful paint, and removing what blocks it. A practical tour of the critical rendering path and the Core Web Vitals.

Web DevelopmentPerformanceCore Web Vitals

The lifecycle of a web request: a browser resolving a domain through DNS, opening a TCP and TLS connection to a server, exchanging an HTTP request and response, and rendering the returned HTML into pixels.

Web Development

Aug 17, 2026·4 min read

What happens when you type a URL and press enter

It feels instant, but pressing enter kicks off a precise chain of events — DNS, TCP, TLS, an HTTP exchange, and the browser's render pipeline. Following that chain is the fastest way to understand how the web actually works.

Web DevelopmentNetworkingHTTP

A convolutional neural network diagram: an input image passing through stacked convolution and pooling layers, feature maps growing deeper and smaller, ending in a classifier that outputs a label.

Deep Learning

Aug 16, 2026·5 min read

Convolutional networks for images, explained

Why do convolutional networks dominate vision? Because they bake the right assumptions about images directly into their architecture. A walk through filters, pooling, receptive fields, and the depth that made them work.

Deep LearningCNNsComputer Vision

Image Processing

Aug 15, 2026·5 min read

Image processing from first principles

Before a neural network ever sees an image, decades of deterministic tricks can already sharpen it, find its edges, and clean it up. A tour of classical image processing — pixels, histograms, filters, and morphology — and why it still matters.

Image ProcessingFundamentalsFiltering

A 3D data-visualisation dashboard: bar charts, a scatter plot, line and donut charts, a world map, a heatmap, and a 3D surface — the kinds of plots that make high-dimensional data, including embedding spaces, legible.

Machine Learning

Aug 14, 2026·2 min read

Embeddings, explained: coordinates for meaning

Vector databases sound exotic until you see the one idea underneath them: an embedding is just coordinates for meaning, and search becomes geometry.

Machine LearningEmbeddingsSearch

The JPEG pipeline visualised: a photo split into 8x8 blocks, transformed into frequency coefficients, most of them zeroed out by quantization, then packed into a compact bitstream.

Image Processing

Aug 13, 2026·5 min read

How JPEG compression works: from pixels to a smaller file

JPEG shrinks a photo to a fraction of its size by throwing away detail your eye barely notices. Here is the pipeline — colour transform, the discrete cosine transform, quantization, and entropy coding — and where the loss actually happens.

Image ProcessingCompressionJPEG

One street photo shown three ways: a single class label, labelled bounding boxes around each object, and a coloured per-pixel segmentation mask — illustrating classification, detection, and segmentation.

Computer Vision

Aug 12, 2026·5 min read

Classification, detection, segmentation: the three questions in computer vision

A vision system can tell you what is in an image, where it is, or exactly which pixels belong to it — three tasks of rising precision and cost. Here is how each one works and the model families behind them.

Computer VisionObject DetectionSegmentation

A camera viewing a night street scene through translucent feature-detection planes, with a neural network turning the image into labelled bounding boxes and coloured segmentation masks.

Computer Vision

Aug 11, 2026·5 min read

Computer vision: how machines learn to see

A camera turns the world into a grid of numbers. Computer vision is the work of turning that grid back into meaning — edges, objects, scenes. Here's the mental model, from pixels to convolutions to learned features.

Computer VisionDeep LearningConvolutional Networks

Machine Learning

Aug 10, 2026·2 min read

Explainability isn't optional: shipping models you can defend

A score without a reason is a liability. Notes from building credit-risk scoring where every prediction has to be explainable, not just accurate.

Machine LearningExplainabilityMLOps

Deep Learning

Aug 6, 2026·2 min read

Attention, in plain English: why transformers won

Nearly every model in the current wave is a transformer, and the transformer's one big idea has a plain-English name: attention. Here's the intuition, no matrix math required.

Machine LearningDeep LearningTransformers

A diagram of an efficient AI model: a compact neural-network core on a chip that demands less data, less memory, and less compute, which unlocks deployment to mobile, edge, IoT, and embedded devices.

Aug 2, 2026·2 min read

Small models, big leverage: the case for efficient AI

The frontier gets the headlines, but distillation, quantization, and small task-specific models are where a lot of real production value is quietly showing up.

AIMachine LearningSystems

MY NOTES

Agentic AI, demystified: it's tools and a loop

◆ Trending now