Home / Engine room
Engine room · our own operationsHow we put AI to work ourselves.
Before we build something for you, we run it ourselves. Here you can see which processes at Punktexe run with AI, how we separate data according to the shell model and how we test our own agents. Everything comes from our own operations, no client references.
The shell model: the more sensitive, the further inside.
The more confidential a piece of information, the further inside it is processed, at the core with local models. We define for each application which data may leave which shell. Open a shell to see what runs there.
- Permissions per shell
- Budget per model
- Separation per area
- Log
Core · local models
- What runs
- Speech recognition with Whisper and local language models such as Llama or Qwen, on our own hardware.
- Which data
- Confidential material: drafts, personnel data, internal documents.
- Boundary
- We deliberately process confidential material with local models. Only what has been approved for each application goes outside.
Shell 1 · tools, APIs, research
- What runs
- Web search, interfaces to the website, email and file storage, our own scripts.
- Which data
- Public or approved content.
- Boundary
- Each tool receives only the access it needs, with its own key.
Shell 2 · EU and specialist models
- What runs
- European models and specialist models, for example for speech or documents.
- Which data
- Internal content according to classification. Personal data is masked beforehand where the process runs through the gateway.
- Boundary
- Approval for each application, costs and permissions via the gateway.
Shell 3 · frontier models
- What runs
- The most powerful cloud models for demanding tasks such as code, research and writing.
- Which data
- Only non-critical or cleansed content.
- Boundary
- No confidential data. Where the classification is unclear, no external processing is intended.
The gateway: the door between the shells
Our own AI gateway brings several models together behind one access point, local as well as European and cloud models. Routing, permissions, budget and logging sit in one place. If a local model fails, another local model takes over automatically. A model can be swapped without rebuilding the application.
What runs with AI at Punktexe
Seven processes from our own operations. Each one states how it is reviewed and where a person decides.
Tax return from end to end
Receipts arrive by email, are recognised, read and filed in a traceable way. From these, the documents for the annual accounts and the electronic submission to the tax office are produced. We show the process, not the figures.
Check: Review steps check the allocation. A person confirms the submission to the tax office.
This website, built automatically
The pages are created as source code in our own repository. Agents write drafts, the reviewer agent checks brand and facts, and publishing takes place via an interface. For AI search, we automatically generate a Markdown version of every page.
Check: The reviewer agent and automated checks review every change, and everything is traceable in the history. A person decides on larger changes.
Onboarding new staff
New staff and partners receive a curated package: a knowledge base, guides, and explainer videos and audio summaries generated from our own documents. The time saved goes into getting to know each other in person.
Check: Content is reviewed before it is passed on.
Speech to text, locally
Speak freely, and the computer turns it into an email, note or post in the desired tone. Speech recognition runs locally with Whisper. We build voice agents on the same technology.
Check: The text is read before it is sent.
Inbox and knowledge
Information sent by email to a dedicated address first lands in quarantine and is then filed in the right place in the knowledge store. Long audio recordings and documents are converted locally into searchable text.
Check: The quarantine checks content before anything is added to the knowledge store.
Sales
Several sources produce a cleansed target list with clear criteria. Drafts for first messages are created within guardrails and checked by the reviewer agent.
Check: The reviewer agent checks every draft, and messages are sent after approval.
Market monitoring
A daily overview of relevant public signals, read only, with suggestions on where it is worth joining the conversation.
Check: Whether and how we respond is our own decision.
Applied AI: what is possible in your company
Here we show what we use ourselves. Where AI can also work alongside your people on a factory site is shown in the floor plan.
How we test ourselves
What we build for others must pass with us first.
Attacks on our own agents
For our own agents, we have built a test matrix of more than 200 prompts in 16 attack classes, in English, German and Bavarian, mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS. We are currently setting up regular testing with it. This is our own quality assurance, not a testing service for third parties.
Masking personal data
When a request runs through the gateway, it attempts to mask detected personal data such as names, email addresses and telephone numbers locally before the cloud call. Detection can miss details, and not every one of our processes uses this route yet.
Credentials do not belong in source code
Before every commit, automated checks look for known patterns of credentials and private keys. If something is found, the commit is blocked.
Separate devices for separate tasks
Sensitive and experimental processes run on their own devices, separated from each other.
Regular housekeeping
A recurring check searches the knowledge store for secrets, breached separations and outdated content.
Our agents
They have names because we work with them. People make the decisions.
Trudi
Red teamWorks with our test matrix for attacks on our own agents. Regular testing is being set up.
Backy
BackupBacks up data and makes sure nothing leaves the house by accident.
Judy
Night shiftStarts scheduled runs every morning, even when nobody is sitting in front of the screen.
The Reviewer
Brand and factsChecks every text before approval: formal address, evidence, no empty promises.
Knowledge we share
Talks by Günther Kriele in a personal capacity, independent of client projects.
- 24th OWASP Stammtisch Augsburg, July 2026: AI firewalls put to the test. More than 200 prompts, some in dialect, and the question of when a hit really is a vulnerability.
- Compliance as code with OSCAL: describing and checking requirements in machine-readable form.
On the road: where we speak, listen and learn
One talk of our own and events we attended.






Which of these do you need?
Get in touch with no obligation or suggest a meeting straight away. We will respond personally.
Where do you stand? Take the AI check (4 minutes) →