ENNL
Free consultation
Home / Services / LLM Optimisation
Service. LLMs

LLM Optimisation

LLM optimisation makes a generative AI system you already run accurate, grounded and affordable: better retrieval, evaluation sets that catch regressions, and token costs measured on your real workload. It starts with an audit of the assistant you have.

Get real, reliable value out of large language models, grounded in your own knowledge, tuned to your use case, and kept accurate, fast and affordable. We make GPT-class models actually work for your business, not just demo well.

Last updated: 14 September 2026

Book a free consultation → See all services
In short

LLM optimisation makes a generative AI system you already run accurate, grounded and affordable: retrieval that finds the right passage, prompts and evaluation sets that catch regressions, and token costs measured on your real workload. Start with a €2,500 audit of the assistant you have.

What it is

Make LLMs accurate, grounded and affordable

A raw LLM is impressive and unreliable in equal measure: it makes things up, drifts off-topic, and the bills add up fast. LLM optimisation is the work that turns that raw capability into something dependable: grounded in your own documents, constrained to your domain, evaluated for accuracy, and tuned for cost and speed.

Whether you're building a customer-facing assistant, an internal knowledge tool or a content workflow, we make the model behave, and prove that it does.

What's included

From impressive demo to dependable tool

RAG & knowledge grounding

Connect the model to your own documents so answers are based on your truth, with sources.

Prompt & system design

Carefully engineered prompts and guardrails that keep the model on-task and on-brand.

Fine-tuning

Tune a model on your own data for the cases where prompting alone isn't enough.

Evaluation & guardrails

Automated testing for accuracy, hallucination and safety, before and after launch.

Cost & latency tuning

Smaller models, caching and routing to cut spend and speed up responses.

Assistants & copilots

Chatbots, copilots and agents wired into your tools and your data.

How it works

Grounded, measured, then optimised

Four fixed steps. You see something real at the end of each one.

  1. A large team meeting in a boardroom1

    Define

    We pin down the use case, the knowledge sources and what "correct" means.

  2. Archive binders lined up on an office shelf2

    Ground

    We connect your data with RAG and shape the prompts and guardrails.

  3. A person checking a document by hand3

    Evaluate

    We test for accuracy, hallucination, cost and speed against real questions.

  4. A dashboard of metrics open on a laptop4

    Optimise

    We tune, cut cost and harden it for production, then keep watching.

What you walk away with

An LLM you can actually rely on

FAQ

Questions, answered

Why not just use ChatGPT out of the box?

For general questions, do. For your business, a raw model doesn't know your data and will confidently get things wrong. Grounding and guardrails are what make it reliable enough to put in front of customers.

Do you fine-tune or use RAG?

Usually RAG (grounding in your documents) first, it's less expensive, faster to update and more transparent. We fine-tune when the use case genuinely needs it.

How do you stop hallucinations?

We ground answers in your sources, add guardrails, and run automated evaluations so we can measure and reduce wrong answers rather than just hope.

Which models do you use?

Whatever fits (OpenAI, Anthropic, open-source models like Llama, or a mix) chosen for accuracy, cost, privacy and where the data is allowed to go.

Can it run privately for sensitive data?

Yes. Where data can't leave your environment, we can use private or self-hosted models so nothing sensitive goes to a third party.

Do you send our data to OpenAI or other third parties?

Only if you choose to use a cloud API. We can build entirely on open-weights models running in your own infrastructure if data must stay on-premises. We recommend the architecture that fits your security and cost requirements.

How do you reduce hallucination risk beyond basic RAG?

Through source citation in every answer, multi-step output validation, task-specific fine-tuning where needed, and automated evaluation suites we run before and after deployment.

Is this compliant under the EU AI Act?

We classify each LLM application against EU AI Act risk tiers at scoping. Most business-assistant use cases are low or limited risk: we document the basis, implement the required transparency notices and stay current as the regulation evolves.

Who can we hire for LLM optimisation in the Netherlands?

Crux Digits is a boutique applied-AI firm based in the Utrecht region, working with clients across the Netherlands and the EU. Because we are a small senior team, the engineers who scope your LLM work are the ones who build it, with no offshore hand-off. If you want a hands-on Dutch or EU partner to make a model accurate and grounded rather than a large consultancy, we are worth a conversation.

How do we get started on an LLM project, and what does it cost?

We start with a short paid discovery to scope the use case, your data and the risks, then quote a fixed price so there are no surprises. Senior engineers do the build, you own the source code and IP, and we work EU AI Act and GDPR aware by default. The Production-ready MVP is a fixed €20,000; larger production builds are scoped after that first step.

Related reading

RAG vs fine-tuning

What are AI agents?

AI agents in production

AI agent vs chatbot

The production AI stack

Related case studies

AI Discharge-Summary Automation (Clinical NLP)

Draft and clarify discharge summaries fast, cutting turnaround and documentation-error risk while the clinician keeps sign-off.

Scope

What is included, and what is not

So you can price the gap before a call rather than after one.

In scope

  • Answers grounded in your own documents. Retrieval over your sources, so the system cannot invent a policy, a price or a clause that does not exist.
  • An evaluation set with a pass mark. A fixed set of real questions with the answer your experts accept, run on every change, so a prompt tweak cannot quietly make things worse.
  • Refusal behaviour you chose. What it must decline, what it escalates, and what it says when the source does not cover the question.
  • The code, models and prompts. In your repository, under your licence, with the documentation. You can hand the whole thing to another supplier and have them continue.

Not in scope

  • Writing the source documents. If the policy is out of date or contradicts itself, the model will surface that faithfully. Fixing the source is yours, and it is usually worth doing first.
  • Third-party licences and usage costs. Model and API usage, and any SaaS the solution calls, are billed to you directly so you see the real running cost. We set a ceiling and an alert rather than an estimate.
What you can verify
One real engagement

AI Discharge-Summary Automation (Clinical NLP)

What we found. Producing and interpreting discharge summaries by hand was slow and error-prone, holding up the discharge process and pulling clinicians away from care. The group needed faster, clearer documentation without compromising the clinician’s final review.

−60%Discharge-process turnaround time
ClearerDischarge summaries, fewer errors

Real result from a Crux Digits engagement; client name withheld pending permission. Decision support only: the clinician retains sign-off on every summary.

Read the full case study →

Want an LLM you can actually trust?

Tell us what you want it to do, we'll show you how to make it accurate and affordable, in a free consultation.

Book a free consultation →