Skip to content
mptr

AI · 2026

We put AI to work.

If you can tell AI could speed up your business, we'll help you find a sensible first step. It might be a customer support assistant, an internal knowledge base, document processing, or an assistant that works on your own data. We design it, integrate it with your systems, and run it in production.

Ask about an AI project AI security

What we do with AI

Three directions, one team.

LLMs in your product

RAG, assistants, document data extraction, semantic search. Claude, GPT, or open-source, depending on what fits technically and commercially.

  • RAG over your own knowledge base
  • Assistants in admin panels and customer portals
  • Document processing, structuring, summarizing
  • Semantic search instead of plain keyword search

AI on your own infrastructure

Open-source models on our GPU server or yours. Data stays in the EU, and in Hungary if you need it to.

  • Qwen3.8, DeepSeek V4, and other open-weight model deployment
  • Multilingual embedding models that also handle Hungarian well
  • Model serving with vLLM or TGI, access control and rate limiting
  • GDPR documentation of the data flow

AI agents

Multi-step AI agents for internal processes. Claude/GPT or self-hosted, with MCP integrations and human checkpoints.

  • Multi-step agents that use tools
  • MCP servers for your systems
  • OpenClaw-based agents that run locally
  • Internal process automation: content, support, data
  • Human approval at critical steps

What we use daily

Our tools, the ones we actually use.

We don't chase every trend. This list only includes tools we've used in production or can confidently take on work with.

Agent systems
LangGraph, CrewAI, OpenClaw, custom orchestration.
Locally run agents
OpenClaw deployment and customization — personal and internal agents that run on your own machine or server, not in the cloud. Markdown-based memory, messenger UI, any LLM as the backend.
Coding assistants
Claude Code, Cursor, Aider in our own development work. We know their limits, so we know what to use them for on client projects.
MCP integrations
MCP servers for client systems so Claude/GPT/local models see them as native tools.
Open-source models
Qwen3.8 and DeepSeek V4 deployment for clients with data-sovereignty constraints.

Our principle: if a trendy technology is missing from the list, it means we wouldn't yet take responsibility for a project built on it.

AI security

We don't just build it, we secure it too.

  • Prompt injection testing: before launch, we deliberately attack the AI application.
  • Data access in RAG systems: permissions, personal data redaction, logging.
  • Model access control: who can use which model with which data; rate limits, automatic spending caps.
Full security page

How it works

Task in, answer out. Feedback in, a sharper system.

PROBLEM input SOLUTION output FEEDBACK · REFINEMENT
  1. 01
    Input

    A question, a document or an event from your systems.

  2. 02
    Processing

    A model working on your own data: RAG, agents, rules.

  3. 03
    Solution

    An answer or an action, with human review at the critical points.

  4. 04
    Feedback

    Corrections flow back, so the system gets sharper week by week.

Got an AI idea? Start here.

Describe your idea briefly and we'll reply in writing: suggested approach and stack, estimated cost range, main risks.

Let's talk Services