Service · Web development

Custom generative AI solutions.

A working AI feature in your own software: text, images, code or audio generated on demand. Not stand-alone experiments, but an integration that fits your workflow, your security requirements and the European AI Act.

Text generationImage generationRAG & your own dataGDPR & AI Act

Generative AI = AI that creates new content.

Generative AI is the umbrella term for models that produce new content from a prompt: a blog post, a product description, an image, a piece of SQL, a speech clip or a summary of a long document. The difference from classic AI: the output is not a score or a label, but new content you can use straight away. For your end user, that usually means a visible creation or generation UI in your product: a "write for me" button, a suggestion field, an image generator or a summariser that works in one click.

For organisations, the question is rarely "can this be done?" and almost always "how do we do this securely within our existing systems?", a question we hear from many clients in exactly those words. We build that layer: from initial use-case selection to production-ready integration with your application, your data and your compliance frameworks. We work custom, the code stays in your repo, and we also deliver an AI strategy and roadmap if you need direction first before we build.

Our typical clients are content agencies, marketing teams, e-commerce organisations with large product catalogues, creative agencies, media companies and publishers, training and learning companies, customer service organisations, and product teams that want to add an AI feature to their existing SaaS. The use cases vary widely, from automated product descriptions to legal contract generation to question-answering bots on internal documentation, but the pattern is the same: a model that creates content, a UX that keeps humans in control, and an architecture that doesn't leak your data without your say-so.

Three flavours of generative AI.

Most enquiries we receive fall into one of these three categories. Which one fits depends on where the generation becomes visible to your end user and how much the model needs to know about your data.

Compact project · fixed sprint budget

AI feature within existing software

You already have a product and want to add a generative layer to it: a "write for me" button, a summariser, a suggestion field, a translation flow or an alt-text generator. We build it into your stack, with a prompt layer you can adjust yourself, streaming output so users don't wait for a long response, and cost monitoring so a viral moment doesn't blow up your monthly budget.

Prompt engineeringAPI integrationStreaming UXCost monitoring
Mid-sized project · fixed sprint budget

Generation on your own knowledge base (RAG)

The model doesn't write out of thin air, but answers based on your own documents, product data, policies or helpdesk history. Retrieval-Augmented Generation with a vector store, source citation and an evaluation set that you can extend yourself. Related to our custom LLM integrations, but specifically focused on generative content flows.

RAG architectureVector storeSource citationEvaluation set
Larger project · fixed sprint budget

Multi-modal generation platform

A custom platform or feature suite in which multiple models work together: text, image, audio and code. Often for marketing, e-commerce or media organisations producing branded content at scale. We build the orchestration between models (Stable Diffusion for images, Claude or GPT for text, ElevenLabs or similar for audio), a review workflow in which editors or brand managers can approve content, and cost controls per user, project or campaign. For public-facing output, we handle the transparency labels required by the AI Act.

Multi-modelImage & audioReview workflowBrand guardrails

What you get at the end.

A working generative AI solution, embedded in your application, plus everything around it so you can manage and fine-tune it yourself.

We work custom: no white-label SaaS, no wrapper around a chatbot product. The code lives in your repository, the choice of model is yours, and you can switch providers at any time without us having to rebuild anything.

  • The generation feature itselfProduction and staging, running in your own cloud (GCP/AWS/Azure) or hosted by us, with streaming UX and latency monitoring.
  • Codebase + prompt libraryFull source code in your repo, plus a structured library of prompts and system messages that you can maintain yourself.
  • Model abstraction layerA thin internal API so you can switch between Claude, GPT, Gemini, Llama or Mistral without touching the application.
  • Evaluation set + monitoringA test set with expected outputs, hallucination detection, a cost dashboard per user or feature, and alerts when the model behaves unusually.
  • Compliance packDPIA template, AI Act classification of your use case, output labelling where required, and clear documentation for your data protection officer or compliance officer.
  • Managed service contract (optional)Monitoring, model updates, prompt tuning, ongoing development and evaluation of new models. Fixed monthly price, four response-time tiers.

When generative AI is the right choice.

Four patterns we often step into. If you recognise one, we'd be glad to talk through how we would approach it in your situation.

Content at scale

Marketing or e-commerce content

You produce blog posts, social posts, product descriptions, alt texts or multilingual variants in such volumes that writing them by hand no longer scales. AI produces the first draft based on your product data, tone of voice and SEO guidelines; an editor reviews and publishes.

Service & support

Making customer contact smarter

Tickets are summarised automatically, replies are suggested based on your knowledge base, and sentiment and urgency are made visible. Your service team stays in control and keeps the final decision, but handles more cases per day without quality dropping.

Document flows

Generating contracts and reports

Quotes, contracts, audit reports or policy documents are largely filled in by a template plus AI based on structured input. Lawyers or consultants review and finalise, instead of writing every paragraph from scratch.

Product features

AI features in your own software

You have a SaaS product and your users now expect there to be "something with AI" in it — a writing assistant, a suggestion, a summary. We build it in neatly, without disrupting your existing UX.

HR & recruitment

Job postings and CV screening

Job ad texts in a few variants, anonymisation of CVs for more objective screening, or suggested interview questions based on a job profile. Entirely feasible, but exactly the type of use case where AI Act classification needs to be spot on — high-risk when it is used to make decisions about candidates.

Learning & training

Personalised learning paths

Generating quiz questions from study material, progress coaching with personal feedback, or an AI tutor that explains things at the learner's level. Frequently requested by training and learning companies that want to enrich their course library at scale.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

How a generative AI project works.

1

Use case workshop

We spend half a day with your team to determine which generative application delivers the most value. Not the most spectacular one, but the one that fits your data, your users and your risk appetite. We build nothing without a workshop — we too often see impressive demos that fall apart in production.

2

Scoping & model selection

Together we choose the model (cloud API, self-hosted open source or a hybrid), define the evaluation criteria, and draw up a DPIA assessment if personal data is involved. By the end: a concrete scope, a plan and the first screen flow with a working prototype prompt.

3

Build in sprints with a human in the loop

A working build every two weeks. We test not only whether the feature runs, but also whether the output is good enough — using an evaluation set you help put together yourself. Hallucinations, prompt injection and output leakage are actively tested in this phase, not only in an audit afterwards.

4

Rollout, monitoring and further development

A phased rollout, first to a small group, with monitoring of costs, latency, errors and user satisfaction. After that, ongoing management: evaluating new models, adjusting prompts, and expanding the evaluation set as you learn more about what works well.

Frequently asked questions.

What clients usually want to know before we start with generative AI.

What exactly is generative AI, and how does it differ from an AI agent?
Generative AI is an umbrella term for models that produce new content (text, images, audio, video or code) from a prompt. An AI agent is more specific: a system that carries out multiple steps on its own, calls tools and finishes tasks. Many generative solutions have no agent layer: they generate something once, a person reviews it, done. Which one you need depends on how autonomous the system has to be.
How do I integrate generative AI securely into existing systems?
We work to a few fixed principles: personal data is not sent to the model without pseudonymisation, prompts and outputs are logged in your own environment, prompt injection is actively filtered, and we choose providers with EU data residency, or a self-hosted model where needed. We also build an output layer that filters sensitive information before content leaves your application. Compliance is not a separate step at the end; we make it part of the architecture.
Which model do you choose: cloud API or self-hosted open source?
It depends on your requirements. Cloud APIs (Claude, GPT, Gemini) are often the strongest and the quickest to deploy, but your data leaves your infrastructure and costs scale with usage. Open-source models (Llama, Mistral, Qwen) run in your own cloud: more control, less convenience, higher infrastructure costs but fixed monthly outgoings. For most organisations a hybrid approach works: non-sensitive flows in the cloud, sensitive flows on a model of your own. We advise based on your data, use case and data classification, and build an abstraction layer so you can switch providers later without having to rebuild the application.
How do you handle copyright on generated content?
The legal position is still unsettled, so we advise conservatively. For text and code we follow the current consensus: in most jurisdictions, model output is not protected by copyright, so exclusivity cannot be claimed. For images, we choose models whose training data status is clear and we set up logging so that provenance can be demonstrated. For questions about output IP or commercial reuse, we coordinate with your legal adviser; we are engineers, not lawyers.
What do you do about hallucinations?
Hallucinations can't be eliminated, but they can be greatly reduced. We use Retrieval-Augmented Generation so the model draws on your own documents rather than on its training. We add source citations so users can check where an answer comes from. We build an evaluation set with expected outputs and run it automatically. And for high-risk flows we keep a human in the loop: the AI proposes, the human publishes.
What about the EU AI Act and GDPR?
Generative AI falls under the AI Act, with specific transparency requirements: AI-generated output must be recognisable as such, deepfakes must be labelled, and high-risk applications carry heavier obligations. For GDPR, what matters is which data the model sees and where it is stored. We carry out a DPIA on every major project, classify your use case under the AI Act, and handle labelling where required. We are not a law firm, so for heavier legal questions we coordinate with your data protection officer or lawyer, but we deliver the technology audit-ready.
What determines the cost of a generative AI project?
Three things determine the build cost. One: the scope of the feature — single-shot text generation is quicker to build than a RAG architecture with its own vector store, evaluation set and source citations. Two: integration complexity with your existing systems, especially when multiple sources or legacy APIs are involved. Three: the compliance package — an internal tool for twenty employees falls under a different regime than a feature for a hundred thousand end users. There are also ongoing costs: model API usage, infrastructure costs if you self-host, and maintenance. After an introductory meeting, we'll give you a concrete estimate based on your situation and data volumes.
Will our employees need training to work with this?
Almost always a short introduction, rarely extended training. We build the feature so that it feels intuitive to your users — a writing assistant should work as naturally as a spell checker. For your IT team and any admins we do provide a manual, and we hand over knowledge in the final sprint so your team can adjust prompts themselves afterwards. For more substantial AI projects, we also point clients to our enterprise AI implementation page for the organisational side.

Talk to us about your generative AI solution.

A no-obligation introductory meeting of half an hour. We listen to the use case you're unsure about, advise on model choice and compliance, and tell you honestly when generative AI isn't the right solution yet. No sales talk — nobody wants that.

Edit content