Deep learning

Text classification

ticket routing, sentiment, content moderation, intent

Assign labels to documents, tickets, or messages.

Probability & Bayes

Information theory

Odds & log-odds

Foundations

Naive Bayes is Bayes' theorem; cross-entropy is the loss for every text classifier since.

Frame the problem

Frame

Multi-class or multi-label. Decide what happens to ambiguous cases.

Labeling & data collection

Data & labels

Annotation guidelines and agreement first. Active learning cuts the labelling bill.

Modalities → vectors

Tokenization

word2vec & GloVebackground

Represent

Bag of words for the baseline; tokens into a pretrained encoder after that.

Split the data

Split

Group by author or source when documents repeat.

Dumb baseline

TF-IDF + linear

Naive Bayes

Attention & transformersbackground

Zero-shot

Prompt an LLM and measure it against the baseline before fine-tuning anything.

The LLM ladder

Fine-tuned encoder

A small encoder, fine-tuned, beats prompting at volume.

Masked & contrastive pretraining

Fine-tuning & LoRA

Model

TF-IDF plus logistic regression, or naive Bayes. Often enough.

Fine-tuning a pretrained model

Handle imbalance

Train

Content moderation is this playbook with extreme imbalance and adversarial drift: retrain often.

Classification metrics

Error analysis & significance

Evaluate offline

Per-class F1. Slice by length, language, and source. For moderation, measure recall on new attack patterns, not last month's.

Evaluate online

Nothing unusual here.

Serving & release

Inference optimisation

Monitor & retrain

Ship & monitor

Cost per call decides encoder versus LLM.

React Flow