PE & VC operating partners
Find where AI pays off in a portfolio company — then get it built and running.
Current agentic-AI and digitization work for a PE/VC firm.
grait, pronounced “great”, with AI inside. Mariusz Kreft
AI that runs your operations — and improves how it runs them.
grait is the AI consulting practice of Mariusz Kreft (Berlin & Poland): AI strategy, AI agents that take over repeat work in your processes and learn from it, Odoo-based solutions and private AI infrastructure — advised, built and run in production.
You write a few lines → we talk it through → you get a short written recommendation: what to automate, where it runs, what to do first.
who it's for
Three teams I work with today
Manufacturers & Mittelstand
Odoo as your business backbone, AI agents for the repeat work, private AI on hardware you control.
Current agentic-AI and digitization work for a Swiss manufacturing company.
Engineering teams
Hands-on AI engineering training for senior developers: agentic coding, orchestration, memory, evaluation.
Current recurring training for senior developers at Markant (Switzerland), a retail company.
services
What I do for operating teams
I advise leadership teams, build agentic workflows and Odoo-based systems, and run them in production — in your cloud or on hardware you control.
advise
AI strategy & architecture
Know where AI pays off — before you spend on it.
For Owners and managing directors, PE/VC operating partners and portfolio companies, manufacturers and SMEs adopting AI.
- A use-case map ranked by value and feasibility
- A target architecture: cloud, private or hybrid
- A build / buy / run decision and a roadmap
- A C-level presentation your board can act on
Proof Current work for a PE/VC firm and a Swiss manufacturing company; transformation programs in retail, automotive, banking and pharma (freelance).
Ask about this →build · flagship
Self-improving agent workflows
Operations run by herds of AI agents that capture lessons and improve their own workflows.
For Teams buried in repetitive cross-system work, or that outgrew n8n- or Zapier-style automation.
- Custom workflows run by herds of AI agents, with human approval where it matters
- A lesson-capture loop that reviews and improves the workflow over time
- An audit-proof decision trail: a signed, hash-chained log of AI decisions
- Can work on top of your Odoo — see below
Proof The Machine.Machine (machinemachine.ai) fleet: 22 AI agents running, orchestrated with m2herd. Previously n8n; now custom workflows run by herds of agents.
Ask about this →build · erp
Odoo-based solutions
Your business on one backbone, with AI agents doing the repeat work around it.
For Manufacturers, Mittelstand and SMEs that want one business system instead of scattered tools.
- Odoo ERP as the business backbone
- Custom addons for how you actually work
- AI agents wired to it for the repeat work
- Examples of what I build: a recruiting stack, dealflow, ERP for SMEs
Pairs with the agent workflows above and, when data must stay in-house, the private AI setup next to it.
Ask about this →run
Private AI infrastructure
Own your AI: LLMs on hardware you control, so your data never has to leave the building.
For Companies whose data cannot leave the building — legal, finance, health, industrial IP — and teams whose API bills keep climbing.
- Hardware sizing for your workloads
- An inference stack (vLLM / SGLang) behind an OpenAI-compatible API
- Secure access, monitoring, watchdogs and runbooks
- Ongoing operation, if wanted
Proof Machine.Machine operates an 8-node NVIDIA DGX Spark cluster and 4× RTX 3090 GPU servers, self-hosted.
Ask about this →teach · side door for engineering leads
AI engineering training
Developers who use AI on their real work, and leaders who can judge it.
For Senior developers, engineering leads and leadership teams who need to judge AI, not just hear about it.
- Agentic coding on your team's real work, not toy examples
- Multi-agent orchestration, memory and evaluation
- Workshops and C-level presentations for leadership
Proof Current recurring hands-on AI engineering training for senior developers at Markant (Switzerland), a retail company. Also trained executives and coaches in transformation programs (freelance).
Plan a training →how it works
How an engagement starts
Every engagement starts with an AI diagnosis of one process.
- 01
Diagnose
You write a few lines and we talk it through on a call. You get a short written recommendation: what to automate, where it should run, what to do first.
- 02
Build
The workflow, Odoo extension or private AI setup from the recommendation, built with your team.
- 03 ↻
Run & improve
Run in production, and keep improving how it runs.
Not sure which fits? Start with an AI diagnosis — in German, English or Polish.
Request an AI diagnosisproof
Proof you can inspect
Current engagements, past clients, and the systems that run right now.
Current engagements
current · PE/VC
A PE/VC firm
Agentic AI systems and digitization.
current · manufacturing
A Swiss manufacturer
Agentic AI systems and digitization.
current · retail
Markant (Switzerland)
Recurring hands-on AI engineering training for senior developers.
Before: clients, programs & product roles
- appBauer clients
- Procter & Gamble · Breitling · ADAC · Telekom
- Transformation
- Programs for corporates in retail, automotive, banking & pharma (freelance); trained executives and coaches
- Product roles
- Babbel · KIEDIS · unhamster · Neofonie (WePad)
For technical reviewers: what runs right now
Machine.Machine operates an 8-node NVIDIA DGX Spark cluster, 4× RTX 3090 GPU servers and 22 AI agents, all self-hosted; 134 repos on GitHub.
- herdr-factory-loop-skillm2herd · flagship
Herds of AI coding agents through herdr in a spec-driven factory loop;
m2herd evolveturns failed runs into reviewed lessons. - fleet-playbookoperating model
The operating model for autonomous AI organisations: identity, memory, autonomy loops, governance.
- openclaw-m2-memory-skillagent memory
Long-term memory for agents: Qdrant + BGE-M3 embeddings.
- ask-fable-skilldelegation
Delegates hyper-complex tasks to a Claude Code worker inside herdr.
point of view
Recursive self-improvement for operating businesses
Most automation does the work. This kind of system also learns from the work: agent fleets capture their own lessons, then review and amend their own workflows — while humans keep the governance. Mariusz Kreft experiments with this in the Machine.Machine fleet, and the building blocks are open source.
credentials
15+ years from boardroom to production
Computer scientist, product leader and entrepreneur: from his own product company (appBauer) and Babbel to transformation programs for corporates — building AI products since before generative AI went mainstream.
- – present
- Founder & AI Systems Architect · Machine.Machine Vertical AI adoption for SMEs: industry-specific agents and workflows with an audit-proof AI decision trail.
- current
- AI Consultant & Tech Partner · grait.io Agentic AI systems and digitization for a PE/VC firm and a Swiss manufacturing company.
- current
- AI Engineering Trainer · Markant (Switzerland) Recurring hands-on AI engineering training for senior developers at a retail company.
- –
- Product Lead · KIEDIS
- –
- Co-founder & Chief AI Officer · unhamster · Timmerman Partners
- –
- Agile & Transformation Coach (freelance) — corporates in retail, automotive, banking & pharma
- –
- Senior Product Manager · Babbel
- –
- Founder · appBauer — Procter & Gamble, Breitling, ADAC, Telekom
- –
- Product Manager · Neofonie
- 2004 – 2009
- Diploma, Computer Science in Media · Beuth University Berlin
Full CV → projects, toolbox, industries
Want this experience on one of your processes? Start with an AI diagnosis.
Request an AI diagnosisfaq
Questions buyers ask first
What happens when I request an AI diagnosis?
You write a few lines about one process: what it is, which systems you use today and where the data must live. We talk it through on a call. You then get a short written recommendation: what to automate, where it should run and what to do first. From there you decide whether to build it with grait.
Do you build on Odoo?
Yes. grait builds Odoo-based solutions: Odoo ERP as the business backbone, extended with custom addons for how you actually work and wired to AI agents that do the repeat work around it. Examples of what I build: a recruiting stack, dealflow and ERP for SMEs.
Can the AI run on our own hardware, so our data stays in-house?
Yes. grait designs and operates self-hosted LLM inference on hardware you control, from a single GPU server to a multi-node cluster, behind an OpenAI-compatible API. Machine.Machine runs its own AI the same way: an 8-node NVIDIA DGX Spark cluster and 4× RTX 3090 GPU servers, all self-hosted.
Do you work with PE/VC firms and portfolio companies?
Yes. grait currently builds agentic AI systems and digitization for a PE/VC firm. The usual starting point is an AI diagnosis of one process, or AI strategy work: a use-case map ranked by value and feasibility, a target architecture and a build / buy / run decision.
Do you work with manufacturers?
Yes. grait currently builds agentic AI systems and digitization for a Swiss manufacturing company. Industries worked in: retail, automotive, banking, pharma, private equity, manufacturing, environmental and urban planning, media and publishing, and EdTech.
Do you train developer teams?
Yes. Mariusz Kreft currently runs recurring, hands-on AI engineering training for senior developers at Markant (Switzerland), a retail company. The training grait offers covers agentic coding, multi-agent orchestration, memory and evaluation. He also runs workshops and C-level presentations for leadership teams.
Which languages do you work in?
German, English and Polish, all fluent. Based in Berlin and Poland.
How is this different from n8n or Zapier?
I built automations with n8n in the past. Today the workflows are custom and run by herds of AI agents, orchestrated with m2herd. The difference: the agents capture lessons from their work, and the fleet reviews and amends its own workflows over time, while humans keep the governance.
Who is Mariusz Kreft, and how do you say “grait”?
Mariusz Kreft is a Berlin- and Poland-based computer scientist, product leader and entrepreneur with more than 15 years in software, AI and digital transformation, and the founder of Machine.Machine (machinemachine.ai). grait is his AI consulting practice. It is pronounced like “great”; the “ai” in the middle is intentional.
Where can I see your work?
Machine.Machine operates an 8-node NVIDIA DGX Spark cluster and 4× RTX 3090 GPU servers and runs 22 AI agents, all self-hosted. The code is in the GitHub organization github.com/machine-machine (134 repos); start with herdr-factory-loop-skill (m2herd), fleet-playbook, openclaw-m2-memory-skill and ask-fable-skill.
contact
Have a process that should run itself?
Write a few lines: which process, how big the team is, which systems you use today (e.g. Odoo or another ERP), where the data must live, and when you need it. In German, English or Polish.
You decide afterwards whether to build it with grait.
Hi Mariusz, Your role: Company type (PE/VC portfolio / manufacturer / retail / other): The process that should run itself: Systems today (ERP/Odoo, tools): Team size: Where must the data live (cloud / EU / CH / own servers / don't know): Timeline: Preferred language (DE/EN/PL):
No mail app? Copy the template and send it to hi@grait.io.