AI software development: LLMs, RAG and agents in production
AI features that hold up in production
AI proofs of concept are easy. The hard work starts when a feature has to handle real documents, real users and real compliance requirements. That is where we come in.
We build LLM integrations, RAG architectures and agentic workflows with evaluation and monitoring, so quality is measured rather than guessed. Where an agent does not hold up, we say so.
Data stays with the client: we work with European and self-hosted models, GDPR-compliant and with the EU AI Act in mind.
The typical situation
Many AI initiatives fail on the question of which parts of a process are suited to automation and which are not. Specialist processes add sensitive data, such as client data at patent law firms.
How we approach it
- Understand the process first: At Wayra we mapped the assessment workflow together with the team before defining any tools.
- Hybrid instead of full automation: classic workflow steps with n8n and LLM stages where they hold up. That keeps the process traceable and under the control of the business experts.
- Costs in view: Rule-based pre-filters remove false positives before any LLM costs arise.
- AI in the real workflow: Patent Studio has 4 AI use cases in the product, from a domain assistant to a drafting tool for patent claims.
- Control over data: All AI features are built so that sensitive client data stays under control.
7,235 patents have gone through the opt-out process via Patent Studio, and 5 law firms use it actively. The AI sprint with Wayra took 1 month.
AI projects like these often qualify for public funding. Check your project with the funding check.
How it works
We approach AI projects in three phases. Each phase ends with a result you use to decide whether and how to continue.
- Consulting and architecture, from one week: We first understand the process and work out which steps are suited to automation and which are not. The result is a decision paper with architecture, options and an effort estimate.
- Prototyping and pre-study, typically 2–6 weeks: We define the tools and assess where LLM stages hold up and where classic steps are better. One example is our one-month AI sprint with Wayra. The result is a clickable prototype, a risk assessment and a solid implementation plan.
- Engineering and operations, months to years: We build the feature with evaluation and monitoring, take it into production and keep quality and costs in view. The result is software in production, documented and ready to hand over to your team.
At Wayra, the sprint ended with the handover of a hybrid solution. Patent Studio is in production with 4 AI use cases. You can find an overview of the phases under Services.
Technologies
We choose models without vendor lock-in. These are the technologies we use in AI projects:
- Models: OpenAI and Azure OpenAI, European and self-hosted models
- RAG and orchestration: LangChain and ModelFusion
- Vector search: Azure AI Search with a chunking and embedding pipeline
- Workflows: n8n for classic steps, agentic workflow patterns for LLM stages
- Documents: OCR with AWS Textract
- Applications: React, Tauri and Astro
We mapped the process for Wayra together with the team in Excalidraw.
“Introducing Patent Studio was a turning point for our firm. Saving more than a third of our time is not just a number. It is time we now invest in strategic advice for our clients.”
Frequently asked questions
Does our data stay with us in AI projects?
Yes. Data stays with the client: we work with European and self-hosted models, GDPR-compliant and with the EU AI Act in mind. We choose models without vendor lock-in. At Patent Studio, all AI features are built so that sensitive client data stays under control. More under Security & compliance.
Do you automate the whole process with agents?
Not always. Where an agent does not hold up, we say so. At Wayra we first mapped the assessment workflow together with the team. The result was deliberately hybrid: classic workflow steps with n8n and LLM stages where they hold up. That keeps the process traceable and under the control of the business experts.
How do you measure the quality of AI features?
With evaluation and monitoring, so quality is measured rather than guessed. Cost control is part of it: rule-based pre-filters remove false positives before any LLM costs arise. At Patent Studio, the AI features run in the law firms' real workflow, not as a demo. The goal is traceable quality instead of demo effects.
How quickly can we get to a result?
Consulting starts at one week; a pre-study typically takes two to six weeks. The AI sprint with Wayra took one month: process mapping, requirements for agentic tools and handover of a hybrid solution. After that you decide whether and how to continue. AI projects often qualify for public funding, see the funding check.
- LLM integrations and RAG architectures
- Agentic workflows and automation
- Evaluation, monitoring and cost control
- Sovereign hosting in the EU and Germany
- GDPR and the EU AI Act
- Model choice without vendor lock-in
- AI features that run in production
- Traceable quality instead of demo effects
- Data sovereignty stays with the client
- Honest limits where AI does not hold up
- A fast path from prototype to product