AI development

AI development company

Most AI projects fail on the parts around the model: retrieval quality, the evaluation harness, cost per request at real volume, and what the product does when an answer is wrong. We build AI features that hold up against those questions, and we will tell you when a problem does not need a model at all.

What's included

LLM applications

Assistants, copilots and generation features built on the Anthropic and OpenAI APIs, with versioned prompts, streaming and structured outputs.

Retrieval-augmented generation (RAG)

Document ingestion, chunking, embeddings and vector search on pgvector, with retrieval quality measured and tracked.

Document and speech processing

Extraction, classification and summarisation pipelines for contracts, invoices, support tickets and recordings, with a human review step where accuracy matters.

AI inside existing products

Adding an AI feature to a product you already run, released behind a feature flag so it can be measured against the version without it.

Evaluation and cost control

Evaluation sets, regression checks on prompt changes, caching, model routing and a cost budget per request, set before launch, not after the first invoice.

How we work

  1. 01

    Qualify the use case

    We check whether a language model is actually good at the task, and what accuracy the product needs to be useful.

  2. 02

    Data and retrieval

    We assess the source data, design ingestion and retrieval, and check that the right context is being found before we tune generation.

  3. 03

    Prototype and evaluate

    A working prototype with an evaluation set, so quality becomes a number you can track through every change.

  4. 04

    Make it production-ready

    Streaming, rate limits, fallbacks, caching, observability, and safe handling of prompt injection and sensitive data.

  5. 05

    Measured launch

    Release with usage, quality and cost instrumented, and a rollback that is one flag away.

  6. 06

    Tune

    Iterate on retrieval, prompts and model choice using the evaluation set and real usage.

Tools and technology we use

  • OpenAI
  • Anthropic
  • LangChain
  • pgvector
  • Python
  • Node.js

Frequently asked questions

Do you train your own models?
Rarely, and only with a clear reason. Most business problems are better served by retrieval, good prompting and a strong hosted model than by training and maintaining your own. When that applies to your project, we will say so.
Which models do you build on?
Mainly Anthropic's Claude and OpenAI's models, chosen per task. Routing lets a single feature use a cheaper model wherever quality allows. The provider sits behind an abstraction layer, so switching is a configuration change.
How do you stop it giving wrong answers?
We ground answers in your own data through retrieval, constrain outputs to structured formats where possible, evaluate against a fixed test set, and design the interface so users can see sources and correct the system.
What does it cost to run?
It depends on volume and model choice, so we agree a cost budget with you before building and treat it as a design constraint. We track cost per request and use caching and routing to stay within budget.
Is our data used to train models?
No. We use enterprise API tiers with zero data-retention terms where available, and design pipelines so sensitive fields are redacted before they leave your infrastructure.

Scope your AI product

Describe the task you want to automate or augment, and what a good answer looks like. We will tell you honestly whether AI is the right tool and what it would take.

Start the conversation