Deep learning

Information extraction

invoices, contracts, resumes, forms

Pull fields out of documents into a schema.

Probability & Bayes

Information theory

Foundations

CRFs and HMMs are graphical models over sequences; token-level losses are cross-entropy.

Frame the problem

Frame

Define the schema and what counts as a correct field. Decide the human-in-the-loop: review everything, or only low-confidence fields.

Labeling & data collection

Data & labels

Span-level annotation is slow. Weak supervision from existing databases and LLM pre-annotation cut the cost.

Modalities → vectors

Tokenization

Represent

OCR quality bounds everything downstream for scanned documents. Layout matters; keep positions.

Split the data

Split

By document template, so you measure generalisation to layouts you have not seen.

Dumb baseline

Structured extraction with LLMs

Sequence labeling

A fine-tuned token tagger when volume is high and the LLM is too slow or expensive.

Sequence labeling & NER

Graphical models

Model

Regexes and rules for the fields that are regular. An LLM with a schema and function calling for the rest. Measure before training anything.

Fine-tuning a pretrained model

Fine-tuning & LoRA

Train

Fine-tune a small model on LLM-labelled data once the schema is stable.

Classification metrics

Error analysis & significance

Evaluate offline

Field-level precision and recall, exact and fuzzy. Slice by template and source.

Evaluate online

Nothing unusual here.

Structured extraction with LLMs

Serving & release

Inference optimisation

Monitor & retrain

Ship & monitor

Type and range checks on every field; reject and route to a human rather than guess. New templates arrive silently: monitor confidence and rejection rates per source.

React Flow