AI for Research programmes

Corpus digitisation

Turning archives, collections and field material into machine-readable data.

• PDF/document to structured JSON

  The first pass over any corpus: structure, reading order, tables and figures recovered as data rather than flat text.

• Handwriting recognition

  Manuscripts, field notebooks, ledgers and correspondence, where accuracy is measured per field rather than per page.

• Table extraction

  Statistical tables out of scanned reports and historical returns, ready for analysis.

• Multilingual translation

  Two hundred languages including many with no commercial service, which matters for area studies and low-resource material.

• Duplicate / near-duplicate text detection

  Deduplicating a scraped or aggregated corpus before it biases a result.

Annotation and extraction

Producing the structured data an analysis actually runs on.

• Named entity recognition

  People, places, organisations, dates and domain-specific entities labelled with their positions in the text.

• Fact / relation extraction

  Statements — subject, relationship, object — that load into a knowledge graph and can be queried and cited.

• Text classification / tagging

  Coding a corpus against your own scheme, with inter-rater agreement measured against your coders.

• Sentiment / intent analysis

  Discourse and reception studies at a scale manual coding cannot reach.

• Structured data extraction from text

  Free-written sources turned into a dataset with the fields your analysis needs.

• Text summarization

  Literature triage across thousands of papers, with the source retained for every claim.

Audio and video corpora

Interviews, recordings, broadcast archives and field footage.

• Audio to text - transcription

  Interview and fieldwork transcription with timings, in roughly a hundred languages.

• Speaker diarization

  Who spoke when, which is what makes an interview transcript analysable rather than an undivided block.

• Speech language detection

  Labelling a multilingual collection, including recordings that switch language mid-way.

• Audio classification / tagging

  Bioacoustics, soundscape and monitoring work — species, events and conditions labelled with their times.

• Audio semantic search

  Finding the moment in a recorded collection by what was said or by what was heard.

• Video semantic search

  The same over footage, with the file and timecode returned so a claim can be checked.

• Speech/music source separation

  Recovering usable speech from field recordings made in noisy or musical settings.

Images and observation

Imaging collections, instrument output and observational footage.

• Image classification

  Coding an image collection against your own categories, with accuracy measured on your data.

• Image segmentation

  Exact shapes, which is what allows area, coverage and size to be measured rather than merely counted.

• Open-vocabulary object detection

  Finding what you can describe but have no labelled examples of — the usual state at the start of a study.

• General object detection and tracking

  Behavioural and movement analysis from recorded footage, with paths rather than per-frame counts.

• Pose estimation (human)

  Movement studies where a skeleton can be kept and the footage discarded.

• Crowd counter

  Attendance and density estimates from images and footage, without identifying anybody.

• Visual similarity / duplicate image detection

  Grouping an image collection and finding the same object photographed across sources.

Ethics and disclosure

What an ethics committee will ask about before the data can be used or shared.

• PII detection and redaction

  Anonymising a corpus so it can be shared or deposited, measured on what it misses rather than on average accuracy.

• Document redaction

  Producing a release copy where removed material is genuinely gone, not hidden under a rectangle.

• Face detection / face count

  Finding faces so they can be blurred, without identifying anybody — the usual requirement for footage in a publication.

Research programmes usually arrive with their own model — trained on the group’s own data, published alongside a paper, or inherited from a collaborator. It runs on Super-GPU infrastructure with the same pinned versions and per-run records as our own services.

We help you select the right hardware, containerize and deploy the workload, build the processing pipeline and define the infrastructure required to scale it.

Your model. Your data. Your deployment choice.