Generative AI development

Generation features measured against your content, not a demo

Extraction, summarisation, search and drafting built on your own material — with structured outputs, source citations and a review step where the result matters.

  • Grounded in your documents
  • Schema-validated outputs
  • Evaluated on real examples
  • Human review before publishing

Written scope and a fixed estimate within two business days. NDA on request.

Hosted models

Fastest to ship, best quality per token, no infrastructure to run. Data leaves your environment under the provider terms you accept.

Self-hosted models

Data stays inside your infrastructure. Lower per-token cost at volume, higher operational burden and usually a quality trade.

The data-residency decision is made before the model choice, not after.

Capabilities

What we build

Each of these is scoped around a measurable output: fields extracted correctly, answers grounded, drafts accepted after edit.

Document extraction

Invoices, contracts and forms turned into structured fields, with confidence scores and a review path for anything uncertain.

Content generation

Product descriptions, summaries and first drafts generated from your material and your tone rules — with a human publishing step.

Search over your content

Semantic search and answer generation across documents, help centres and internal wikis, returning the passages used.

Summarisation

Long transcripts, reports and threads condensed to the decisions and action items, not a generic abstract.

Classification and tagging

Consistent labels across large content sets, with an agreed taxonomy and an audit sample you can check.

Custom model behaviour

Prompt systems and, where it genuinely helps, fine-tuned models for narrow tasks with stable output requirements.

Quality

Four steps that separate a demo from a feature

The test set is built from your real inputs before the pipeline is finished, so “it works” has a number attached.

  1. 1

    Define “good”

    Write the output requirements as a checklist a person could grade — length, tone, required fields, refusals.

  2. 2

    Build the test set

    Real inputs from your business, including the awkward ones. Fifty good examples beat a thousand synthetic ones.

  3. 3

    Build the pipeline

    Retrieval, prompting, validation and the interface. Structured output enforced with a schema where possible.

  4. 4

    Measure and iterate

    Score the test set, fix the failures in prompts or retrieval, and keep the score as a regression test.

Prompting vs fine-tuning

When training a model is worth the cost

ConsiderationPrompting + retrievalFine-tuning
Time to first versionDaysWeeks
Output quality for a narrow taskGood with good promptingHigher, with training data
Data requirementExamples for evaluationA labelled dataset, often thousands of items
When the task changesEdit the promptRetrain
Where it makes senseMost business use casesHigh volume, stable, narrow tasks

Output rules

What every generation feature ships with

  • Structured outputs validated against a schema, not parsed out of prose
  • Every generated claim traceable to a source passage where facts are involved
  • A refusal path for inputs outside the supported scope
  • Human review before anything customer-facing is published
  • Versioned prompts, so a behaviour change is a reviewable diff
  • Cost and latency ceilings measured per request, not discovered on the invoice

Stack

What we build with

OpenAIAnthropicOpen-weight modelspgvectorEmbeddingsChunking strategyJSON schemaGuardrailsEval harnessWeb appInternal toolAPI

Proof

Products where content is the product

Study guides, a video learning library and a language app — content-heavy products we designed and built.

FAQ

Generative AI questions

Do we need to fine-tune a model?

Usually not. Most business use cases are solved with prompt design, retrieval of your own content and output validation. Fine-tuning helps when the task is narrow, the volume is high and the required output is tightly constrained — and it needs a labelled dataset that most teams do not have yet.

How do you keep generated content accurate?

By grounding it: the model works from retrieved passages and returns them alongside the output, structured fields are schema-validated, and anything customer-facing has a human approval step. We also measure a failure rate against your own examples instead of asserting it works.

Can we keep our content out of third-party models?

It depends on the provider terms you accept. We can use zero-retention endpoints on hosted providers, or run an open-weight model inside your infrastructure when the data cannot leave. That decision is made during scoping.

What does it cost per request?

It varies with model, prompt size and output length, so we measure it on your actual workload during the build and set ceilings. A feature with unpredictable per-request cost is a design problem, not a billing surprise to hand you later.

Can this work with our existing CMS or documents?

Yes, if the content can be exported or reached by API. The ingestion and cleaning step is usually more work than the model integration, and it is where content quality problems surface.

Request a quote

What content work should the model take on?

Send a sample of the inputs and what a correct output looks like. We will tell you whether prompting, retrieval or training is the right approach, and what it costs to prove it.

No obligation. We will tell you if we are not the right team for the job.

  1. 1

    Tell us the goal, platforms and must-have features.

  2. 2

    We reply within two business days with scope, approach and a fixed estimate.

  3. 3

    Sign an NDA first if you prefer — nothing about your idea is public.