Custom AI development for real business use
We design and build AI products and features around a defined problem, the information available and the people who will depend on the result. The work includes the software around the model, not just a prompt and a demonstration.
- 30,000+ conversations handled by Holp.ai since launch
- Senior only the people scoping the work also build it
- Live product experience developing and operating AI in use
Where AI development is worth considering
AI is useful when a problem depends on language, judgement or a large amount of information that is difficult to handle with fixed rules. It still needs a clear job, a reliable source of information and a way to recognise when the answer is uncertain.
We start by understanding the process: what people are trying to do, what a good answer looks like, which mistakes matter and where a person must stay involved. Only then do we decide which model or approach fits.
If conventional software is the clearer or more economical answer, we will recommend that instead.
What the work includes
A dependable AI feature needs product decisions, software engineering and an evidence-based way to test the model. Depending on the job, that can include a maintained knowledge base, retrieval-augmented generation (RAG), semantic search, evals and guardrails.
Use case and evaluation
Define the task, assemble representative examples and agree how we will judge whether the result is useful.
Knowledge bases and RAG
Structure approved source material into a maintained knowledge base, then use embeddings and semantic search to retrieve relevant passages. Retrieval-augmented generation (RAG) gives the model that material when it produces an answer.
Application development
Build the interface, accounts, permissions, integrations and operational tools that turn a model into usable software.
Model and tool integration
Choose and integrate the model based on the job, using structured outputs and tool calling where the application needs the model to retrieve data or take a controlled action.
Evals and guardrails
Build repeatable evals from realistic examples, put guardrails around inputs and outputs, and define when the system should refuse, ask for clarification or hand over to a person.
Monitoring and improvement
Review traces from live use, measure retrieval and answer quality, monitor latency and cost, and improve the product as its information and underlying models change.
We built and run Holp.ai
Holp is a custom-built customer service AI product that answers questions from an organisation’s own information. It has handled more than 30,000 conversations since launch for organisations including Castle Howard, Jerry Green Dog Rescue and the UK Green Building Council.
Running the product means dealing with the work after the first release: improving answers, reviewing conversations, supporting customers and adapting as the underlying AI changes.
Read the Holp.ai case study →Questions about AI development
Can you add AI to software we already use?
Yes. An AI feature often makes more sense inside an existing process than as a separate product. We can review the current application, its data and integrations, then identify the smallest useful addition.
Which AI model do you use?
That depends on the task. We compare candidate models using evals built from representative cases, then consider answer quality, speed, privacy, availability and cost before choosing one. The surrounding application should make it possible to change provider when there is a sound reason.
Can you turn an AI prototype into a live product?
Yes. We can assess what the prototype proves, what can be retained and what needs engineering before real users or sensitive information are involved. See our prototype to production service.
What if we are not ready to build?
Start with AI consultancy. We can assess the use case, the information involved and the commercial case before recommending a product, prototype or custom build.
Discuss an AI development project
Tell us about the problem, the information available and the people who will use the result.