Deep dives, tutorials, and thoughts on AI.
19 of 30 posts published · One AI concept every day, all of Ramadan 2026.
Machine learning explained from zero — what it is, how it differs from normal programming, the three types, and why it now powers everything.
Read more →The complete journey from raw data to a deployed model — data collection, cleaning, feature engineering, training, evaluation, and deployment.
Read more →Loss functions, gradient descent, and learning rate — the universal engine behind every ML model from linear regression to GPT-4.
Read more →Linear regression from scratch — the simplest and most foundational ML model. OLS, MSE, R², and why it still powers production systems.
Read more →How a linear model learns to output probabilities using the sigmoid function, log-loss, and what the decision boundary really means.
Read more →Distance metrics, the curse of dimensionality, choosing K, and its surprising link to vector search in modern LLMs.
Read more →Bayes' theorem from scratch, the independence assumption, why it works despite being wrong, and text classification in production.
Read more →Gini impurity vs entropy, overfitting, pruning, feature importance — the foundation of every modern ensemble method.
Read more →Bias and variance explained through darts, learning curves, and every model unified into one framework.
Read more →L1, L2, Elastic Net, and Dropout — the tools that prevent overfitting in every model with intuition, math, and code.
Read more →How asking 500 imperfect models instead of one perfect model produces the most reliable predictions in classical ML — bootstrap sampling, feature randomness, and OOB validation.
Read more →Bagging makes models vote in parallel. Boosting makes them learn in sequence — each new model fixing exactly where the last one failed.
Read more →The engineering masterpieces that made gradient boosting 10–100× faster — second-order gradients, histogram splits, and early stopping in production.
Read more →The craft that separates good ML engineers from great ones — creating, transforming, and selecting features that make every model smarter.
Read more →A logistic regression unit, stacked and made non-linear — the bridge from classical ML to deep learning. Neurons, backprop, ReLU, and vanishing gradients from scratch.
Read more →How machines learned to see — convolutions, filters, pooling, receptive fields, and residual connections explained from first principles.
Read more →How a model pre-trained on everyday photos achieved 95% accuracy on brain tumour classification with just hundreds of examples.
Read more →Classification tells you what. Detection tells you where. Segmentation tells you which exact pixels. YOLO, IoU, NMS, and U-Net explained.
Read more →Teaching machines to read — tokenisation, bag-of-words, TF-IDF, and the foundational techniques that turn language into computable numbers.
Read more →The first serious attempt at sequence modelling — how RNNs process text word by word, why gradients vanish over long sequences, and the exact mathematical proof that led to the LSTM.
Read more →Forget gates, cell states, and the gradient highway — how LSTM and GRU solved the vanishing gradient problem and powered a decade of NLP breakthroughs.
Read more →BPE, WordPiece, SentencePiece — how GPT, BERT, and LLaMA split text into tokens, why the choice matters more than you'd think, and how to build a tokenizer from scratch.
Read more →king − man + woman ≈ queen. Word2Vec, GloVe, and sentence transformers — the geometry of meaning that powers semantic search and RAG retrieval.
Read more →The architecture that powers GPT-4, Claude, and Gemini — self-attention, Q/K/V, multi-head attention, and the full transformer block built from scratch.
Read more →How transformers understand word order without recurrence — sinusoidal encodings, rotary position embeddings, and the techniques powering long-context models.
Read more →Why BERT reads the whole sentence but GPT only looks left — masked language modelling vs causal language modelling, and when to use each architecture.
Read more →A deeper look at how each major tokenizer algorithm works under the hood — merge rules, vocabulary construction, and why the same word tokenizes differently across models.
Read more →Fine-tuning a 7B model on a single GPU — full fine-tuning, LoRA's low-rank decomposition, 4-bit QLoRA quantisation, and when each approach is the right call.
Read more →Giving LLMs long-term memory — the full RAG pipeline from document ingestion and chunking to embedding, vector search, and grounded generation.
Read more →Zero-shot, few-shot, chain-of-thought, system prompts, and prompt injection attacks — the complete picture of how to talk to language models and how they can be misled.
Read more →