Natural Language Processing
Turn unstructured text into structure, meaning and searchable knowledge your systems can actually query. Eight engagements, each measured on your own documents before anything is promised.
Most text problems are definition problems.
Ask three experienced people to label the same two hundred documents and they will disagree on a fifth of them, usually because the categories were never pinned down rather than because anyone was careless. Every accuracy figure quoted before that disagreement is measured is a figure about nothing. So we start by writing the label scheme with your team, measuring whether they agree with each other, and fixing the definitions until they do.
After that the work is unglamorous and reliable: a model that is right often enough to be useful, a confidence threshold set from what a miss and a false alarm each cost you, and a review queue that only sees the cases the model is unsure about. Where the distinction you want is not actually present in the text, that finding arrives in week two rather than month four.
What we build
Each is a standalone engagement with its own scope, price and output. Most clients use two or three in sequence.
Text Classification and Entity Extraction
Labels and entities extracted from text at volume, with a scheme your experts agree on and per-class accuracy.
Read more →Sentiment and Intent Analysis
Aspect-level sentiment and intent tuned to your domain, calibrated against your own reviewers.
Read more →Document Summarisation
Summaries written for a named reader and decision, checked for faithfulness against the source.
Read more →Semantic and Vector Search
Search that finds the right thing, measured on your real queries with hybrid retrieval and reranking.
Read more →Knowledge Graph Construction
A graph built around the questions you cannot answer today, with entity resolution and provenance per fact.
Read more →Contract and Legal Document Analysis
Clause extraction, deviation flagging and obligation tracking, cited back to the page for fast verification.
Read more →Resume Parsing and Candidate Matching
Parsing and matching that assists shortlisting, with bias testing, explanations and the decision left with a person.
Read more →AI Translation and Localisation
Translation at volume with enforced terminology, quality tiers and human review where the content demands it.
Read more →How they fit together
You do not need all twelve. Most programmes follow one of these paths depending on where the uncertainty sits.
Structure from raw text
Classification and entity extraction with agreed label schemes and measured agreement, plus aspect-level sentiment and intent that tells you what customers are reacting to, not just whether they are cross.
Finding and condensing
Semantic and vector search built as hybrid retrieval with reranking and judged against expert relevance ratings, and document summarisation checked for omission rather than for fluency.
Documents that carry risk
Contract and legal document analysis tuned for recall so nothing material is missed, and resume parsing and candidate matching that ranks rather than rejects, bias-tested before launch.
Knowledge and languages
Knowledge graph construction derived from the questions your systems cannot answer today, with provenance on every fact, and AI translation and localisation with quality tiers by content risk.
FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
How much labelled data does an NLP project need?
For fine-tuning a pretrained model, usually several hundred to a few thousand examples per class, with more for rare classes and subtle distinctions. Consistency matters more than volume: a thousand examples labelled to an agreed scheme beat ten thousand labelled to three different interpretations, which is why the scheme and the agreement measurement come first.
Can a large language model do this without training data?
Often it can get you to a usable first version with no labelled data at all, and for low-volume or exploratory work that is the right answer. At volume the economics change: a small fine-tuned classifier is typically far cheaper per document, faster and more stable in its outputs. We size both before choosing, and say which one your volumes justify.
How is this different from the Generative AI pillar?
Generative AI and LLM engineering is about producing text and handling open-ended reasoning. The work here produces structured, measurable output over large volumes of existing text: a label, an entity, a score, a ranked list, a fact with a source. Many of these builds use language models inside; the difference is that the output is checkable against a ground truth.
Does this work in languages other than English?
Yes, and the honest caveat is that quality varies by language and by task. Multilingual models are strong across major European and Asian languages and thinner for low-resource ones, so we measure on your own text in each language you need rather than quoting an aggregate benchmark, and we say where a language needs its own labelled data.
Who owns the models, datasets and label schemes?
You do, in full, on final payment: fine-tuned models, the labelled dataset, the label scheme and annotation guide, the agreement measurements, evaluation sets, deployment code and monitoring. The labelled data and the guide are usually the most durable assets, because they make every later model faster and cheaper to build.
Start with a conversation.
Thirty minutes, no charge, no deck. Tell us what you are trying to build with language models and we will tell you which of these engagements fits, or whether none of them do.