Short answer
kalAIDoCreate deploys large language models on your own servers or in a private cloud, so AI agents can work with sensitive data without sending it to public model APIs. We handle model selection, infrastructure sizing, the governed agent platform on top and ongoing operation.
On this page
When public model APIs are not an option
For many organizations the question is not whether AI agents are useful, but where the data goes. Public administrations, defence suppliers, hospitals, banks, insurers and industrial companies with sensitive engineering knowledge often cannot or will not send documents to an external model provider, regardless of contractual assurances. Some are bound by classification rules, some by professional secrecy, some simply by a board decision to keep strategic data under their own control.
Open-weight language models have become capable enough for a wide range of agent tasks: classification, extraction, summarization, retrieval-augmented answers and tool use. Running them yourself is realistic, but it is a different discipline from calling an API. It involves hardware, model serving, security hardening, evaluation and operations.
Deployment options
| Option | Where models run | Typical fit |
|---|---|---|
| On-premises | Your own data centre, on GPU servers you own or lease | Highest control, classified or strictly regulated data |
| Private cloud | Dedicated tenancy with a European cloud or hosting provider | Control without operating hardware yourself |
| Hybrid | Sensitive tasks on private models, uncritical tasks on public APIs | Balancing capability, cost and data protection |
| Air-gapped | Isolated network without internet access | Security-critical environments |
In a hybrid setup, the platform routes each task to the right model according to data classification rules. Where text must go to an external model, Textwash Pro can anonymize it locally first, so personal data never leaves your network.
What we take responsibility for
Deploying a model is only half the job. Agents on top need the same governance as anywhere else. kalAIDoCreate delivers the full stack:
- Model selection and evaluation: we test candidate open-weight models against your real tasks and languages, including German, and document the trade-offs between quality, speed and hardware needs.
- Infrastructure sizing: we derive GPU, memory and storage requirements from expected load and latency targets, and plan for redundancy.
- Serving layer: high-throughput inference servers, quantization where appropriate, load balancing and access control.
- Agent platform: the kalAIDoCreate enterprise AI platform runs in the same environment with SSO, role-based permissions, connectors to SAP, SharePoint, CRM and other systems, and complete audit trails.
- Security hardening: network segmentation, secrets management, no outbound calls unless approved, logging into your SIEM.
- Operations: monitoring, model updates after evaluation, capacity planning and incident response through managed AI agents.
Your IT keeps ownership of the infrastructure, the network and access policies. Your business owners decide which tasks agents may perform. We make sure the models and agents behave within those limits and prove it with logs and regular evaluations.
A realistic project path
Discovery. We clarify use cases, data classes, regulatory constraints and the existing infrastructure. Often the answer is a hybrid design rather than everything on-premises.
Proof of capability. AI Code Engineers benchmark two or three models on your own examples on temporary hardware, so the hardware purchase is based on evidence.
Build. Infrastructure is set up, the platform installed and the first agent connected to your systems. Where custom applications or interfaces are needed around it, OnlineAppCreator can deliver done-for-you engineering with SLAs.
Go-live and operation. The first agent goes live with monitoring of quality, latency and resource use. Additional agents reuse the same infrastructure.
Governance documentation, including risk assessment and EU AI Act considerations, is covered by our AI governance and compliance service.
Trade-offs to be honest about
Self-hosted models need hardware investment and operational effort. The strongest proprietary models may still outperform open-weight models on some complex reasoning tasks. We address this with careful task design: agents built around clear responsibilities, structured tools and retrieval rarely need the very largest model. Where they do, a hybrid route with anonymization is often the pragmatic answer.
What you get
Full control over where data and models live. Predictable behaviour, because models change only when you approve an update. Independence from a single provider's pricing and policy decisions. And a governed agent platform that auditors and data protection officers can inspect end to end.
For public bodies with sovereignty requirements, see AI for state agencies; for privacy-first chat use cases, the GDPR-compliant AI chatbot.
FAQ
Are open-weight models good enough for enterprise AI agents?
For many tasks, yes, especially classification, extraction, summarization and retrieval-based answers. We evaluate on your own data before you commit to hardware.
Can we run everything without internet access?
Yes. Air-gapped deployments are possible. Model updates are then imported through your controlled software supply process.
Who operates the GPU servers?
Your IT or your hosting provider operates the hardware. We can operate the model serving, platform and agents, or train your team to do so.
Can we combine private and public models?
Yes. In a hybrid setup the platform decides per task based on data classification, and Textwash Pro can anonymize text before it reaches an external model.