AdvancedML model evaluation mistakes that make good numbers meaningless
Nine errors behind great offline metrics and bad production models: random splits on time series, threshold-free metrics, tuning on the test set.
Library
All articles
Free to read. Filter by topic or search by title.
Browse by tag (19)
Advanced
AdvancedData leakage: the kinds that survive your train/test split
Splitting correctly stops the obvious leak. Target encoding, feature stores, label windows and proxy features leak anyway — how to find each.
AdvancedProduction ML architecture: training-serving skew
The model is the small part. Training-serving skew, feature freshness, shadow deploys, drift detection, and monitoring that catches decay.
AdvancedEmbeddings mathematically: what cosine similarity measures
The geometry of embedding space: why you normalise, what dimensionality really buys you, and the anisotropy problem nobody mentions.
AdvancedThe attention mechanism, worked out by hand
Queries, keys and values with real numbers. Why the scaling factor is sqrt(d_k), what multi-head attention buys, and how causal masking works.
AdvancedTransformer from scratch in PyTorch, block by block
A working decoder-only transformer in 120 lines. Pre-norm, residual streams, why the MLP is 4x wide, weight tying, stable initialisation.
AdvancedCNNs from scratch in NumPy: convolution as a matrix multiply
Convolution forward and backward with no framework. The im2col trick that makes it fast, why the backward pass is a convolution too, and a net that trains.
AdvancedAgentic AI architecture: the loop and what breaks
An agent is a loop with tools and a stopping condition. Context growth, error handling, termination — and when to write a chain instead.
AdvancedMCP and tool calling: designing tools an LLM can use
Tool calling fails at the interface, not the model. Schema design, descriptions, error contracts, and what MCP actually standardises.
AdvancedLLM evaluation: grading a model with no right answer
Vibes do not scale and BLEU measures nothing you care about. Eval sets, LLM-as-judge without fooling yourself, and the biases that corrupt it.
AdvancedBuilding RAG properly: chunking, evaluation, reranking
Most RAG fails at retrieval, not generation. Build a golden set, measure recall@k, run chunking experiments, add a reranker when data says so.
AdvancedFine-tuning vs RAG: how to actually choose
Fine-tuning teaches behaviour, RAG supplies facts. A decision framework, the LoRA maths, real costs, and the cheaper ladder to climb first.
AdvancedVector databases under the hood: HNSW and IVF-PQ
ANN search trades accuracy for speed. How HNSW graphs and IVF-PQ quantisation work, which knobs move recall, and how to choose an index.
Newsletter
Get new articles
Free essays on learning, Python, and ML — no account required. We’ll only email when there’s something worth reading.