Generative AI development
Generative AI applications, built into the way your teams work
We build knowledge assistants, drafting tools and copilots that answer from your own documents and run where your data rules say they must.
Want the plain-language case for generative AI first? Our generative AI overview covers what it is good for. This page covers what we build for you and how.
What we build
RAG knowledge assistants
Answers from your policies, manuals, tickets and wikis, with a link to the source behind every claim. Access follows the permissions you already have.
Content and document generation
Proposals, reports, contracts and product copy drafted from your templates and data, then routed for human review before anything goes out.
Copilots inside existing apps
An assistant panel in your CRM, helpdesk or internal tool that drafts, summarises and looks things up, so people don't have to switch apps.
Document understanding
Classification, extraction and summaries for incoming documents, with confidence scores and a review queue for anything uncertain.
Fine-tuning and open-weight hosting
When an API model costs too much, or the data can't leave your estate, we fine-tune and host open-weight models with vLLM or Ollama on your infrastructure.
Evaluation and cost control
Test sets for answer quality, OpenTelemetry traces for latency and cost, plus routing and caching to keep the bill predictable.
How a knowledge assistant is put together
Most of the answer quality comes from retrieval. The model matters less than people expect.
Ingest
Connect the sources, split documents sensibly and keep metadata such as owner, date and access rights.
Retrieve
Hybrid search over a vector index such as NebulaDB, filtered by what the user is allowed to see.
Answer
The model answers only from what was retrieved and cites it. If nothing relevant comes back, it says so.
Measure
A test set of real questions with agreed answers, scored on every change, plus feedback from the people using it.
Data, quality and cost
Where it runs
Model APIs, Amazon Bedrock or Azure AI Foundry, or open-weight models on your own servers. You decide per use case.
Keys and secrets
API keys stay in your secrets store and reach the application through WSL Vault, not through config files.
Quality you can check
Answers cite their sources, and the test-set results live in your repository for anyone to rerun.
Cost you can predict
Per-request cost tracking, caching, and routing simple requests to smaller models. Our article on LLM observability and cost control shows the method.
Read the article: Cost you can predict
Questions about generative AI projects
We already use ChatGPT or Copilot. Why build something?
General assistants don't know your documents, permissions or processes. A built application answers from your sources, respects access rights, logs what it did and sits inside the tools your teams already use.
Where do our documents go?
Where your data rules allow. We pick providers and settings that match them, or host an open-weight model so the documents never leave your estate.
Can you add generative AI to our existing product?
Yes. We build the feature inside your codebase, behind your authentication, with tests and docs, so your engineers can maintain it.
How do you know the answers are right?
We agree a test set of real questions and correct answers with you before the build, and score every change against it. Answers that can't be grounded in a source are refused.
Pick one use case and we'll scope it
Drafting, knowledge answers or a copilot in an existing app. Tell us which one, and the documents involved.