Autonomous process agents
- Automation of multi-step business logic across ERP systems, databases and APIs.
- Elimination of repetitive manual entry, invoice verification and operational cross-checking.
20 years of enterprise architecture × modern applied AI
Deterministic AI Systems for Small and Mid-Sized Companies
Helping small and mid-sized companies eliminate manual operational drag. We bring the discipline of large enterprise systems — without their budgets or their timelines.
Why Kraus AI
Three reasons the project does not stop at the slide deck.
20 years in enterprise systems
Native understanding of ERP systems, procurement, logistics and master data management. What works in large corporations, delivered at a scale that fits you. You will not have to explain approval workflows to us.
Zero-hallucination architecture (RAG)
An ungrounded model answers even where it has nothing to go on. On an internal note that does not matter; on the reading of a contract or a policy, someone answers for it. That is why every answer is linked to a specific document or database record, with the page cited.
Working prototype in two weeks or less, at an agreed price
De-risk your investment. We build a functional, testable prototype on your real company data in two weeks or less, at a price agreed up front.
What we build for you
Three engagement formats, depending on where you are.
Private knowledge systems (RAG)
- Instant search and structured synthesis across thousands of internal documents, manuals and contracts.
- Built with production-grade vector stores, rerankers and strict permission models.
AI architecture sprint: working prototype in two weeks or less
- Rapid scoping and implementation of a working PoC tailored to your specific process bottleneck.
- Delivered with architecture blueprint, source code and clear ROI metrics.
How the engagement works
Three steps from the first call to production rollout.
- Step 1
Process scoping call (20–45 min)
Identifying the highest-ROI bottleneck, evaluating technical feasibility and agreeing on a measurable criterion by which you will judge the prototype.
- Step 2
Prototype sprint, two weeks or less
Building and testing the solution on representative client data.
- Step 3
Rollout and handover
Production deployment, API integrations, security validation and staff training.
Two weeks is our prototype, not your project. Validation runs at your company between the second and the third step — typically two to three months, after which you decide about going into production.
When it is time to call us
Not every company needs what we build. Where you stand decides who helps you most — and for the first two stages that is not us.
- 01
You have not started yet
Nobody in the company uses AI systematically, and there is no record of who is already trying it privately.
- What you need
- Paid accounts and a single page of rules on what data may go in.
- Who handles it
- You can do this yourself
- 02
Everyone does it their own way
People have e-mails drafted or documents searched, but nobody knows which of it holds up.
- What you need
- Training and a shared library of verified approaches.
- Who handles it
- A trainer or a workshop
- 03
The approach is proven, moving data by hand slows you down
You have measured that it works. The data still travels between two windows by hand.
- What you need
- Wiring existing tools together, for as long as they are enough.
- Who handles it
- An automation specialist
- 04
You have hit your own systems and the question of liability
The process reaches into ERP, a database or an API. Someone answers for mistakes, and the result has to be provable, not just convincing.
- What you need
- An architecture blueprint and an implementation shaped to your data and permissions.
- Who handles it
- This is where we start
What we build on
We build LLM-agnostic solutions — the model is a replaceable part, not a dependency. We pick the framework to fit the process, not the trend.
- Agent frameworks
- DeepAgents (LangChain)
- LangGraph
- PydanticAI
- Google ADK
- Claude Agent SDK
- OpenAI Agents SDK
- MCP (Model Context Protocol)
- Language models
- Claude (Opus 5, Sonnet 5)
- GPT
- Gemini
- Open models (Llama, Qwen)
- Self-hosted: Ollama, vLLM
- Data and retrieval
- PostgreSQL + pgvector
- Qdrant
- Elasticsearch
- Rerankers
- PDF parsing and OCR
- Observability and tuning (LLMOps)
- Tracing of every model call
- Response quality evaluation
- Prompt tuning driven by traces
- Prompt regression tests
- Cost and latency monitoring
- LangSmith, Langfuse, OpenTelemetry
- Operations and integration
- Python
- TypeScript
- FastAPI
- Docker
- REST / OData / SAP
- n8n (workflows and orchestration)
- On-premise and private cloud
Application logic stays decoupled from any single model, so switching providers or moving to a local model does not mean rewriting the solution. We assemble the stack around your security, operations and budget constraints. If wiring existing tools together is still enough for you, we will say so. Once it is not, we build on top of it — n8n handles the surroundings of the process here, not the decision logic.
CASE STUDY / LIVE ASSET
Kecalo: what a finished system looks like, not a demo
- Verifiability instead of trust: every answer cites the document, article and page. When an answer is disputed, a tool in the admin shows which passages it came from and at what similarity score.
- We hand over an operable system: your own staff change the knowledge base, bot behaviour and permissions from the admin. Changing behaviour requires neither us nor a new deployment.
- Measurable quality, your data stays yours: latency, cost and rating on every query, retention periods and erasure under GDPR inside the application. We never train models on client data.
kecalo — Preview of the Kecalo interface
What is the notice period in the supplier framework agreement?
The notice period is three months and starts on the first day of the month following delivery of the notice.
[Article 12.3, Page 8]
Source verified
Who is behind it
Ing. Tomáš Kraus
Founder and principal architect
Over 20 years of enterprise consulting, business process design and systems architecture in SAP. Today an AI engineer focused on deterministic LLM implementations where mistakes are expensive.
Languages
- Czech (native)
- German (fluent, business)
- English (fluent)
- Italian (conversational, business)
Operating regions
- Czech Republic
- DACH (Germany, Austria, Switzerland)
- International
What people ask us
Six questions that come up almost every time. We answer them before you have to ask.
What is left with us if the engagement ends?
Source code, the architecture blueprint, documentation, prompts and configuration. The solution runs on your infrastructure or in your cloud, not on ours. Another supplier or your own people can take it over.
Does this tie us to one model or one supplier?
Application logic stays decoupled from the model, so switching providers or moving to a model running on your own premises does not mean rewriting the solution. Approaches and prompts are handed over with the code.
What does the whole thing cost if the prototype works?
The prototype price is confirmed up front. The scope of the rollout that follows is decided on its result and on the criterion we agreed during the first call — not on an estimate made at the start.
Where does our data go?
To your infrastructure or your cloud tenant. We do not train models on client data. Access to documents is governed by a permission model, so each user sees only what they are entitled to.
Can someone trick the agent with text planted in a document?
It is a real risk for any system that processes incoming documents and e-mail. We keep instructions separate from the data being processed, the agent holds only the permissions a given step needs, and sensitive actions are confirmed by a person.
How will the people whose work this touches react?
The agent takes over the repetitive part; the decision stays with a person. For every process we set out in advance where the checkpoint is and who signs off the output — without that the system would not clear your own internal approval anyway.
Who we are not for
- A first attempt at AI. Start with paid accounts and your own experiments, that pays off sooner.
- A one-off output. A text, a deck or a quick analysis is something you can do in a chat without us.
- A tool that serves five people once a month. It will not pay for itself.
Let's discuss your high-friction processes.
Tell me where the drag is worst. On the first call we work out whether automating it is worth it.