Services
AI agents and LLM integration
Dahlinius builds language models into your systems: AI agents that do work, MCP servers that give models controlled access to your data, and integrations with Azure OpenAI, Anthropic or other providers. In your own code.
What we build
Models from Anthropic, OpenAI via Azure and open models, chosen by requirements on data handling, cost and quality.
- AI agents that handle cases and run workflows within clear limits
- MCP servers that expose your systems to any LLM in a controlled way
- LLM integration in existing .NET, Java or Python systems
- Search over internal documents with the right permissions
- Evaluation, logging and cost control
How it works
We start with a concrete use case, not a strategy. A short review of data and permissions, then a working prototype in your code within a few weeks.
What works gets hardened: error handling, evaluation, logging and operations in Azure.
Experience
Skogsradar has a RAG-grounded assistant over geodata. AgentPurge runs a custom MCP server with over 16 tools and agent loops across Anthropic, OpenAI, Gemini, Groq and Ollama. Both are live today.
For Azure clients we work with Azure AI Foundry, Azure OpenAI and Semantic Kernel. The same patterns as in the products, with stricter limits.
Data handling
Your data can stay in your environment. Azure OpenAI in a European region, clear limits on what the model may do, logging of every call. Framework agreements with CSN and the City of Sundsvall for the public sector.
Frequently asked questions
Which LLM providers do you work with?
Anthropic, OpenAI via Azure OpenAI and open models. The choice is driven by data handling, cost and quality.
What is an MCP server?
An open standard, Model Context Protocol, that lets a language model call tools and read data in a controlled way. Build the integration once, switch models freely.
Can you build AI into existing systems?
Yes, that is the most common case. In .NET, Java or Python, in the code you already have.
How do you keep an agent from doing the wrong thing?
Limit its tools, require confirmation for actions with consequences, log everything and evaluate against real cases before it acts on its own.
Do you have a use case?
Describe it in a few lines and we will assess what is reasonable to build.