Service · Software development

Custom LLM integrations in your own software.

A large language model integrated directly into your application, workflow or customer portal. Classify, summarise and search documents, or map free-form input to structured fields, without your data leaving the model provider for a training set.

Claude · GPT · GeminiPrivate LLMRAG & embeddingsEU data residency

An LLM is a building block, not a product.

A large language model (Claude, GPT-4o, Gemini, Llama) is, on its own, just an API call. The value only emerges once such a model is truly part of your workflow: it reads your documents, classifies your incoming emails, searches your knowledge base, or maps customer input to the right fields in your CRM. That is where our work comes in: the integration, the context, and the monitoring around it.

We build LLM integrations as part of a broader software project or as a standalone capability within an existing application. An AI agent is one specific form of this: a multi-step system that can call tools. An LLM integration is broader and often simpler: a single well-prompted call, an embedding search with RAG, or a fine-tuned classifier running in production.

Three flavours of LLM integration.

It depends on what the model needs to do, how sensitive the data is and how much custom work surrounds the model. We advise on which fits best in the first conversation, and we always check first whether a lighter option will do.

Compact project · fixed sprint budget

Single-call LLM feature

A clearly scoped feature in your application where a single well-prompted LLM call does the work: classifying an incoming email, summarising a long document, giving a review an automatic sentiment label, or turning free-form customer input into structured fields in your CRM. Cloud LLM from Anthropic, OpenAI or Google. Includes monitoring of token usage and quality.

Prompt engineeringClassificationSummarisationForm mapping
Mid-sized project · fixed sprint budget

RAG and semantic search

The LLM gets context from your data. We build a retrieval pipeline on your documents, knowledge base or product catalogue: embeddings, vector store, semantic search, and a chat or Q&A layer that answers only on the basis of retrieved sources. Source citations as standard, hallucinations kept to a minimum. Suited to internal knowledge bases, customer support and product search.

EmbeddingsVector storeHybrid searchSource citations
Larger project · fixed sprint budget

Private LLM on your own infrastructure

An open-source model (Llama, Mistral or a specialised variant) running in your own cloud or on-premise. Full data control, no calls to external APIs, no uncertainty about the training set. We set it up with vLLM, Ollama or a managed inference layer, including the monitoring, scaling and authorisation that go with it. More initial work, but for compliance-sensitive sectors often the only route.

Llama / MistralvLLM · OllamaGPU infrastructureOn-premise

What you get at the end.

A production-ready LLM integration, plus everything around it so you can manage it yourself and steer it on cost, quality and privacy.

  • The LLM integration itselfIn production and staging, running in your own cloud (GCP, AWS, Azure) or with us. For private LLMs, including the inference stack.
  • Prompt library and evaluation setVersion-controlled prompts, an eval set with test examples, and a short report showing how well the model performs your task.
  • Codebase and architecture documentationFull source code, build and deploy instructions, and an architecture overview explaining the model choice, caching strategy and data flow.
  • Cost monitoring and token dashboardInsight into tokens per call, cost per user or feature, and alerts when usage rises unexpectedly. Caching set up where it makes sense.
  • Compliance packDPIA where relevant, GDPR mapping of data flows, contractual agreements with the model provider on data retention and training, and an overview of how your integration relates to the EU AI Act.
  • Maintenance contract (optional)Monitoring of model quality, prompt maintenance, further development and migration if you want to move to a newer model. Fixed monthly price.

When an LLM integration is the right choice.

Four patterns in which we support clients. If you recognise one, we're happy to talk further. Not every use case needs an LLM; sometimes a classic rules engine or a specialised model is cheaper and more reliable.

Document flows

Incoming volume is too large to handle manually

Hundreds of invoices, contracts, emails or requests every day, currently read, sorted and forwarded by hand. An LLM can classify, summarise and extract the relevant fields — leaving your team time for the cases that genuinely need attention.

Unlocking knowledge

Your knowledge base isn't being found

Employees or customers search through a growing mountain of documents and just miss what they need. Semantic search with RAG lifts search results from keyword level to meaning level, with source citations so you know where the answer comes from.

Processing free-form input

Customers type, your system expects fields

Lead forms, support requests, product specifications — users write in free text, while your CRM or ERP wants structured data. An LLM integration maps free text to the right fields, with validation and a human check where needed.

Compliance & privacy

Data can't go to a cloud API

Do you work with medical, legal or financial data where the GDPR and sector regulations are strict? Then a private LLM on your own infrastructure is the right route — more expensive to set up, but without the privacy trade-offs of an external API.

What we use LLMs for in practice.

The applications we build most often. The common thread: a well-defined task, a measurable quality threshold and a logging layer that shows what the model did.

Documents

Classification and extraction

Automatically label incoming invoices, contracts, emails or requests and extract the relevant fields — amount, date, party, type, priority. For a large flow of incoming mail, that demonstrably cuts manual work and keeps the data in your back office consistent.

Text & language

Summaries and translations

Condense long documents into a few paragraphs, or produce domain-specific translations that understand legal or technical jargon. Far more reliable than generic translation APIs once you give the model context about your field.

Sentiment & reviews

Understanding reviews and support tickets

Automatically label customer reviews, NPS comments or support requests by sentiment, topic and urgency. Delivers a dashboard that lets teams spot trends before complaints escalate.

Search

Semantic search with RAG

A search function that understands what a user means, not just which words appear. It combines keyword search with embeddings and returns answers with source citations. A building block for both internal knowledge bases and product search.

Forms

Free-form input to structured data

A customer types what they're looking for in free text, and the LLM maps it to the right fields in your CRM, ERP or order system. Includes a human check for low confidence scores, so errors don't end up in production.

Product data

Generating product descriptions

For e-commerce catalogues with thousands of SKUs: consistent, SEO-friendly descriptions based on specifications. Tone of voice is safeguarded through a prompt library, with people staying in the loop for spot checks and final editing.

Code & SQL

Code and SQL generation in your own tooling

An LLM that generates queries or code snippets inside your application — for example a reporting tool where non-technical users can ask questions in plain language, which the model converts into safe SQL.

Privacy

Anonymisation and redaction

Strip personal data from logs, support ticket archives or training data before they are used further. An LLM recognises patterns that regex misses — names, addresses, policy numbers in free text.

Models and infrastructure we work with.

No vendor lock-in on a single model provider. For each use case we choose what fits on quality, price, latency and compliance — and we make sure that switching to a newer model later doesn't mean a complete rebuild.

Cloud LLMs

Anthropic, OpenAI, Google, Cohere

Anthropic Claude (Opus, Sonnet, Haiku) for nuanced language work and long context. OpenAI GPT-4o and o1 for broad use and complex reasoning. Google Gemini within Vertex AI if you already run on Google Cloud. Azure OpenAI if you run a Microsoft stack. Cohere for specific embedding and classification work.

ClaudeGPT-4o · o1GeminiAzure OpenAICohere
Open-source & private

Llama, Mistral, vLLM, Ollama

For scenarios where data must not leave your environment. We run Llama-3 (70B and smaller), Mistral variants and specialised fine-tunes via vLLM on GPU instances, or via Ollama for smaller models. This includes the operational layer: scaling, monitoring, authorisation and cost control.

Llama-3MistralvLLMOllamaGPU orchestration
Retrieval, evaluation & caching

Vector stores, evals, prompt caching

Postgres pgvector, Qdrant or Pinecone as the vector store, depending on scale and your existing stack. Evaluation frameworks to detect regressions when models are updated. We use Anthropic's prompt caching and OpenAI's cached prompts where they make sense — a caching strategy can significantly reduce token costs in production.

pgvectorQdrantPrompt cachingEval sets

How an LLM integration project works.

1

Introduction and use-case scan

A conversation in which we work out which task the model needs to perform, what data is involved, how sensitive that data is and what level of quality is acceptable. This is also where the first answer to the cloud-versus-private question emerges.

2

Prototyping and model selection

We test several models — Claude Sonnet, GPT-4o, Gemini, or a Llama variant — on your own examples. At the end you get a comparison on quality, latency and cost, plus advice on which model best suits which flow.

3

Building in sprints

A working build every two weeks. We build the integration into your stack, set up an evaluation set, put caching and cost monitoring in place, and configure logging so that we can prove which prompt version produced which answer. This includes any integrations with your existing systems.

4

Rollout, evaluation and maintenance

A phased rollout, starting with a small group of users so that we catch unwanted behaviour before it scales. After that comes ongoing maintenance: prompts keep evolving, models get replaced, and your eval set grows with them.

Frequently asked questions.

What clients typically want to know before we start — including the most frequently asked question: what does a private LLM cost?

How does an LLM actually work, in broad terms?
A large language model is trained on enormous amounts of text and has learned patterns from it. For every question, it predicts, word by word, what the statistically most plausible next piece of text is. It doesn't reason the way a human does; it recognises patterns. For your integration, that means: the more clearly you define the task and the more specific the context you provide, the more reliable the result. A well-built LLM integration rests on three pillars: carefully crafted prompts, relevant context (often via RAG), and an evaluation set that measures whether the model does your task well.
What does a private LLM cost compared with a cloud LLM?
There are no fixed amounts, but there is a clear logic. With a cloud LLM from Anthropic, OpenAI or Google you pay per token — for a well-defined use case that often stays modest, especially with caching. A private LLM on your own GPU infrastructure has higher fixed costs (hardware or GPU instances, maintenance, monitoring) but no more variable token costs. The tipping point comes when token volume gets so high that your own inference stack works out cheaper — or, more often, when compliance requires a private route regardless of price. We work through both scenarios for you in the first sprint so that the choice is based on facts.
Which model should we choose — Claude, GPT, Gemini or Llama?
That depends on the task, your data residency requirements and the balance between price and performance. Anthropic Claude is strong at nuanced text and long context; OpenAI GPT-4o is versatile and multimodal; Google Gemini is interesting within the Workspace ecosystem; Mistral and Llama are open source and suited to private deployment. We test several models on your own examples before choosing — a choice based on gut feeling or brand name is rarely the right one.
Will our data stay within Europe?
Yes, provided we choose it. Anthropic and OpenAI offer EU-region deployments through their Bedrock and Azure variants respectively, Google Gemini runs in EU regions via Vertex AI, and a private LLM on your own European cloud or on-premise keeps data within Europe by definition. We set the choice out in the contract and document the data flows for your GDPR records.
What about GDPR and the EU AI Act?
Under the GDPR, personal data that you have an LLM process requires a data processing agreement with the model provider, a DPIA for higher risks and clear documentation of which data ends up where. The EU AI Act adds a classification on top of that — depending on the application, your integration falls under minimal, limited or high-risk requirements. At the start we map your use case against both frameworks and build the integration so that you can demonstrably comply.
Will our data be used to train the model?
Not with the business API tiers we use. Anthropic, OpenAI (via the API, not the free ChatGPT app) and Google Vertex AI offer contractual guarantees that inputs and outputs are not used for training. We put those clauses in writing explicitly and, where there is any doubt, opt for a private LLM so that your data never reaches the model provider in the first place.
How long does a project like this take?
For a well-defined single-call feature, we can be in production within a few sprints. A RAG project with larger document corpora often needs a few extra sprints for the data pipeline and evaluation. A full private LLM setup including infrastructure takes several sprints. Either way, we work iteratively: after the first sprint you have a prototype you can steer from.
Do you work with our internal IT or data department?
Almost always. We share knowledge throughout the project, deliver a runbook for incidents and agree clear responsibilities. For organisations that want to take on LLM work more structurally, we like to combine this integration with a lightweight AI strategy, so that the first integration doesn't stand alone but is part of a bigger plan.

Talk to us about your LLM integration.

A no-obligation introductory call of half an hour. We listen to your use case, ask questions about data, compliance and quality, and give you direction you can actually use — even if the final advice is not to use an LLM at all. For larger projects, you can combine this conversation with an exploration of enterprise AI implementation.

Edit content