A checklist for training a neural network that works
Overfit a tiny batch, watch the curves and change one thing at a time.
4 min read
Read →Resources
Plain-language guides on machine learning, LLM engineering and MLOps. Short reads, written to be useful on the day you need them.
Latest guide · Deep learning
A checklist for training a neural network that works
Overfit a tiny batch, watch the curves and change one thing at a time.
4 min read · Read the guide →A minimal evaluation harness for LLM apps
How to build a small, repeatable test harness for an LLM feature in an afternoon: cases, checks, scoring and a regression gate.
3 min read
Read →Chunking documents so retrieval works
How chunk size, overlap and metadata affect whether the right passage is found.
2 min read
Read →Data leakage: the reason a great model fails in production
How future information sneaks into training data, and how to spot it.
4 min read
Read →Document AI: OCR, extraction and validation
The stages of a document pipeline and why validation rules catch what confidence scores miss.
5 min read
Read →How text becomes tokens (BPE in plain words)
Why models read subword tokens, and why token counts drive cost and context.
4 min read
Read →How to choose between RAG and fine-tuning
A decision guide: what each approach changes, the questions that separate them, and how to test before you commit.
3 min read
Read →How to tell whether your RAG assistant is any good
A test-set approach that needs no code: retrieval, correctness, grounding and honest refusals.
4 min read
Read →Idempotent pipelines: safe to re-run
Why every batch job should give the same result when run twice.
4 min read
Read →IoU, precision–recall and mAP for object detection
How detectors are scored and why one mAP number is not enough.
4 min read
Read →MLOps for small teams: what actually matters
What a team of one to five people should do for reliable ML in production, and what can wait.
3 min read
Read →Monitoring a deployed model: four layers
Service health, data drift, prediction drift and delayed performance — what to watch and when to alert.
4 min read
Read →Never trust the model's JSON: validate and repair
Why structured output still needs validation, and a simple retry-with-errors pattern.
4 min read
Read →Precision, recall and choosing the right metric
Accuracy hides rare-class failures. Learn to pick a metric that matches the decision.
4 min read
Read →Prompt injection: what it is and how to reduce the risk
Direct and indirect injection, and why layered defences beat a clever system prompt.
4 min read
Read →Python and maths for ML: where to start
The small set of NumPy, pandas and maths ideas that unlock machine learning.
4 min read
Read →RAG, fine-tuning or prompting? A decision guide
A simple way to choose between three ways of adapting a language model to your work.
4 min read
Read →The KV cache: why LLM serving is a memory problem
What the KV cache stores, how big it gets and how paging and prefix sharing help.
4 min read
Read →Train, validation and test sets: why three?
Why an honest ML evaluation needs three datasets, and the mistakes that quietly ruin it.
5 min read
Read →What is an embedding? Meaning as numbers
A plain-language explanation of embeddings and why they power search over documents.
4 min read
Read →When fine-tuning (and LoRA) is worth it
The questions to ask before fine-tuning, and what LoRA changes about the cost.
4 min read
Read →GPU vs CPU for training: when the difference actually matters — CapabilityNext
A practical guide to when you need a GPU for machine learning, when a CPU is fine, and why environment setup is usually the real blocker for learners.
Read →
What retrieval-augmented generation actually is — CapabilityNext
A plain-language explanation of RAG: why you add a retrieval step to a language model, the common failure modes, and how CapabilityNext teaches it.
Read →