We set up SI in your business, improve existing solutions and develop software that uses models purposefully. From local models with freely available weights to frontier-model APIs, with suitable guidance, model training and an introduction for your team.
Individual quote · defined pilot scope · documented handover
Capabilities, data processing and effort belong together. We assess the options against your tasks and decide with you how the solution should run.
Run suitable models on your computer, server or an agreed infrastructure of your own. Runtimes such as Ollama can be one building block.
Freely available model weights can make it easier to get started. Hardware, setup, energy, updates and support still require resources.
Connect capable provider models for demanding tasks, for example through OpenAI, Anthropic or Mistral AI. The specific choice is made within the project.
Features, availability and terms depend on the provider and model. Our estimate uses the conditions applicable to your offer.
Local processing and provider APIs can serve different tasks. A combination needs clear rules about which data may go where.
A local model does not automatically make an entire application offline. Search, extensions, logs and external services must also be considered.
Open weights do not automatically mean an unrestricted open-source licence. Each model’s terms remain authoritative. Background: Hugging Face on licences · Ollama on local operation and cloud features.
Start with one clear question or have us accompany a project throughout. We agree the specific scope in your offer.
We start with your task, existing systems and everyday examples. We compare suitable models by output quality, response time, data flows and total cost.
Outcome: a reasoned choice and a manageable pilot.
Set up a local runtime, server, interface or provider API. Match authentication, roles, access keys, logging and data processing to your use case.
Outcome: a configured environment with clear responsibilities.
Bring in handbooks, internal documents or approved specialist information. We plan source attribution, updates and permissions throughout retrieval and response generation.
Outcome: knowledge integration with traceable sources and access.
Adapt suitable models to recurring tasks: prepare data, plan training and evaluate against independent test cases. We clarify model licences, provider support and data usage rights first.
Outcome: a tested adaptation approach compared with the starting solution.
Integrate assistants, document processing and SI functions into your application. With documented APIs, bounded actions, approvals and understandable error handling.
Outcome: an application that fits your agreed workflow.
Train colleagues on their actual tasks: useful prompts, source checking, handling confidential data and understanding the limits of results. With matching instructions and exercises.
Outcome: your team can use the introduced solution with informed judgement.
Check quality using representative cases, test unwanted data access and manipulative inputs, and define human review and fallback paths. The test plan follows the intended use.
Outcome: documented tests, known limitations and acceptance criteria.
Monitor usage, failures, response times and consumption within the agreed scope. Test model changes, updates and new data before introducing them deliberately.
Outcome: traceable operations with a budget and improvement plan.
Making new documents available, steering responses and training a model are different tasks. We choose the approach based on your intended result.
Clear task descriptions, examples, response formats and deliberately authorised tools can already help. We test whether they can achieve the agreed quality.
The application retrieves relevant material and gives it to the model as context. This allows current sources to inform responses. Permissions, retrieval quality and source maintenance are part of the plan.
Adapt a suitable model to a task using selected examples. We check data quality and rights, separate training and test data, and compare the results with the unchanged model.
Fine-tuning depends on the model and licence and does not automatically improve results. Knowledge integration is not model training. Technical background: Hugging Face PEFT · RAG · Model Cards for documenting suitability and limitations.
Clarify the goal, users, data, responsibilities and budget. These become an offer with scope, schedule and deliverables.
Implement a bounded use case and test it with agreed examples. The existing way of working provides the comparison.
Discuss results and open questions. If you adopt the solution, receive the agreed configuration, documentation, team introduction and acceptance tests.
Continue with an agreed support scope if you wish. Review quality, costs and changes regularly with a clear record.
You receive an individual quote. Consulting, setup, data preparation, development, testing and training are priced within the agreed scope. Hardware, hosting, API consumption and ongoing support are shown separately. We do not promise blanket savings or a particular result; the pilot provides the basis for your decision.
Select what is already clear. We will clarify the open questions together. Receive a summary for your enquiry without committing to a model or a price.
Yes, if the model licence, hardware and task fit. Freely available weights do not make the entire operation free. We first check what your existing equipment can handle and what setup and support will require.
No. Depending on the task, better instructions, suitable search or knowledge integration may be enough. For commissioned training, we use only appropriate, lawfully available data. Purpose, access, retention and deletion are clarified within the project.
Yes. Examples include internal assistants, document-processing functions and connections to existing specialist systems. We agree the scope, source-code handover, rights to custom developments, documentation and operations explicitly in the offer. Third-party models and components remain subject to their own licences.
That depends on the integration. We define which actions are allowed, who must approve them, and how errors are detected and handled. Responsibilities and fallback paths are part of the design.
Practical exercises using the agreed tasks, clear rules for handling data and ways to check results. A suitable handbook and an introduction to the new workflows are included in the agreed training scope. We clarify the scope and audience beforehand.
Technical background reviewed on 9 October 2026. The linked provider documentation explains individual building blocks; we assess the suitability of a specific solution against your requirements.