Technologies
Agents and AI engineering
A model on its own answers questions. An agent reads the purchase order, checks it against the contract, updates the ERP and asks a person when something looks wrong. We build these with LangGraph, the Claude Agent SDK, the OpenAI Agents SDK and MCP, with retrieval, logging and evaluation set up before anything goes live.
4 tools
Agent frameworks
LangGraph
Multi step agents with clear states, retries and human approval points.
Claude Agent SDK
Agents that use tools, files and code with Claude models.
OpenAI Agents SDK
Agents with handoffs and guardrails on OpenAI models.
CrewAI
Small teams of role based agents for research and drafting tasks.
3 tools
Tools and retrieval
MCP
Standard connectors that give agents safe access to your systems and data.
LangChain
Loaders, prompts and integrations for retrieval pipelines.
LlamaIndex
Indexing and querying large document collections.
1 tools
Evaluation and monitoring
Langfuse
Tracing every agent step, cost and response so issues can be found and fixed.
What we build with it
Typical work in this stack
- Back office agents
- Matching invoices to purchase orders and goods receipts, chasing missing documents and preparing entries for approval.
- Order and support agents
- Taking orders or answering queries on WhatsApp and email, and handing over to a person with full context.
- Knowledge assistants
- Retrieval over manuals, SOPs and contracts, with access rules so people only see what they are allowed to.
- MCP servers for your systems
- Connectors that let approved AI tools read and act on your ERP, CRM or database within set limits.
- Evaluation suites
- Test sets and automated checks that run on every change, so quality does not quietly slip.
How we choose
We start with the simplest thing that works. Many tasks need only a single model call with good retrieval. When a process does need several steps, we use LangGraph or the Claude Agent SDK because they make each step visible and let us put a person at the points where a mistake would be costly. Multi agent setups come last, only when one agent genuinely cannot hold the job. Before launch we agree what good looks like, build an evaluation set from your real cases, and keep running it after release.
Questions
Good to know
Other parts of the stack
Will the agent act without anyone checking?
Only for low risk steps you approve. Payments, customer commitments and record changes go to a person until accuracy is proven.
How do you stop wrong answers?
Answers are grounded in your documents with citations, checked against an evaluation set, and the agent says when it does not know.
What is MCP and do we need it?
MCP is an open standard for connecting AI tools to systems. It helps when several AI tools need the same safe access to your data.
Start a conversation
Tell us what you are trying to build.
Share the problem in a short call. You leave with a clear view of the approach, the team and the cost.
A specialist replies within one working day. The first conversation is free.