RAG system design & build
Design and delivery of retrieval-augmented generation systems over your own documents, data and knowledge bases, with evaluation built in from day one.
8–14 weeksfrom €65,000
Independent AI consulting · EU, UK, US and Canada
LLM and RAG systems that survive contact with production.
Rubra Digital is an independent consultancy specialising in retrieval-augmented generation, LLM evaluation and AI governance. I work with regulated and document-heavy organisations, and I measure what I build.
The gap
The demo took two weeks. The production system is still not live.
Most teams I meet have already built something impressive: a retrieval prototype that answers questions well, in a demo, on the clean documents, for people who know what to ask.
Then it meets the whole corpus, real users, access control, and someone asking how you know it is accurate. That is where projects stall. Not on model capability, but on document parsing, evaluation, permissions, observability and the failure handling a demo never needed.
Crossing that distance is the work. I do it, then hand the system over so your team can carry it forward.
Services
Six things I do, and do properly
Each has a defined scope, a stated price band, and evaluation evidence at the end.
LLM evaluation & QA
Turn "it seems better" into a number. I build evaluation harnesses, labelled test sets and CI gates so you can change an LLM system without breaking it.
4–8 weeksfrom €38,000
AI agent engineering
LLM agents that use tools, call your APIs and complete multi-step work, with the guardrails, observability and human checkpoints to deploy them safely.
10–16 weeksfrom €75,000
LLMOps & AI platform
The infrastructure under your AI features: gateways, prompt versioning, tracing, cost controls, caching and pipelines that let several teams ship safely.
6–12 weeksfrom €55,000
EU AI Act readiness
Classify your AI systems, close the gaps and produce the technical documentation the EU AI Act requires, led by an engineer rather than a lawyer.
5–10 weeksfrom €42,000
AI strategy & discovery
A short, evidence-led engagement that tells you which AI use cases are worth building, which are not, what each will cost, and in what order to do them.
3–5 weeksfrom €22,000
How I work
What you get from a specialist instead of a firm
For a broad transformation programme, a large firm is genuinely the right choice. Hire me when the problem is specific and technical.
01
Evaluation before retrieval
The labelled test set gets built in the first two weeks, before any retrieval code. It is unglamorous work and it turns every later decision into a measurement instead of an argument.
02
One person, start to finish
The engineer who scopes the work is the engineer who builds it. No account layer, no handoff to a team you meet at kickoff, no discovering in month three that the estimate came from someone who had not seen your data.
03
Built to be handed over
I do not sell managed services. A consultant who depends on running your system has the wrong incentives when designing it, so the measure of a finished engagement is that you do not need me afterwards.
04
Regulation treated as engineering
Most of what the EU AI Act asks for is logging, evaluation evidence and documented oversight. That gets built in rather than written up afterwards by someone who never saw the code.
Answers
Straight answers to the questions I get asked
No gated PDFs. The useful version of what I would tell you on a call.
- What is retrieval-augmented generation (RAG)?RAG combines a search step over your own documents with a language model that generates an answer from what it retrieved. It gives models access to private and current information without retraining, and makes every answer traceable to a source.
- Should we use RAG or fine-tuning?RAG adds knowledge; fine-tuning changes behaviour. Choose RAG for private or changing information, fine-tuning for format, tone or narrow tasks. They are complementary, not alternatives, and most teams should try RAG first.
- How much does it cost to build a RAG system?Expect €60k–€180k to build a production RAG system and €500–€4,000 per month to run one. Document quality and source system count drive the build cost; query volume drives the running cost, and it is usually the smaller number.
- Is a RAG system GDPR compliant?RAG is generally more GDPR-tractable than fine-tuning because personal data stays deletable and access-controlled in an index rather than baked into weights. You still need a lawful basis, a DPA with your model provider, a valid transfer mechanism, and usually a DPIA.
- How do you evaluate a RAG system?Split evaluation into retrieval metrics (recall@k, MRR, nDCG against known-correct passages) and generation metrics (groundedness, citation accuracy, task completion). 150–300 curated cases in CI catches real regressions.
- How do you reduce hallucinations in an LLM system?Fix retrieval before prompts: most hallucinations are the model answering from context that never contained the answer. Then constrain generation with citations and refusal, verify claims for high-stakes output, and gate on measured groundedness.
Where I work
Regulation is not the same in every market
An EU AI Act obligation, a Quebec Law 25 explanation right and a Colorado AI Act duty are different problems. These pages set out what changes where.
- Germany
- Netherlands
- United Kingdom
- France
- Switzerland
- Ireland
- Spain
- Sweden
- Denmark
- Belgium
- Austria
- Norway
- Poland
- Italy
- United States
- Canada
Data residency in the EU, UK, Switzerland, Norway, the US and Canada. Fully on-premise deployment where sovereignty rules require it.
Insights
Notes from the work
20 July 2026
The EU AI Act: what engineering teams actually have to build
Past the legal summaries: the concrete instrumentation the EU AI Act requires, article by article, and how much of it good engineering already covers.
2 July 2026
Most RAG failures are retrieval failures
When teams report hallucination in a RAG system, the model is usually not the problem. Here is how to tell the difference and what to do about it.
12 May 2026
Your evaluation set is the real product
The labelled evaluation set outlasts your model, your framework and probably your architecture. It is the most durable asset an AI project produces.
Frequently asked questions
What does Rubra Digital do?
I design, build and evaluate production LLM and retrieval-augmented generation systems. Most of the work is with organisations in regulated or document-heavy sectors: financial services, legal, healthcare, manufacturing and the public sector.
Is this a one-person consultancy?
Yes. You work directly with the person who scopes, builds and hands over the system. For larger programmes I bring in specialists I have worked with before, and I say so up front rather than presenting them as staff.
Where do you work?
Remotely across the EU, the UK, the United States and Canada, with on-site workshops where being in the room helps: discovery, architecture review and handover. Working hours cover EU and UK in full, with afternoons overlapping US Eastern and Central.
How does an engagement start?
With a 30-minute call to work out whether this is the right fit. After that, either a three to five week discovery engagement or straight to a build if the problem is already well defined. If you do not need a consultant, I will say so.
What does it cost?
Discovery engagements start at €22,000. Production builds typically run between €55,000 and €180,000 depending on scope. Defined scopes are quoted fixed price, with the assumptions written down before you commit.
Do you work alongside an in-house team?
Usually, and it is the arrangement that works best. The aim of every engagement is that your team can run and extend the system without me. Handover sessions, architecture decision records and the evaluation harness are deliverables, not extras.
Get a straight answer on your AI roadmap
A 30-minute call with the engineer who would do the work, not a salesperson. You will get an honest read on what is worth building, what is not, and roughly what it costs.
No NDA needed to talk. EU and UK hours in full, with afternoons overlapping US Eastern and Central.