Where AI processing happens is a design decision. This page explains the options, what current projects support, and what is still development or R&D.
Two of our projects work with local processing.
Auto LLM Benchmark compares local models on your tasks.
What it does: The same tasks are run on local models; answers, speed and quality grades are collected in one table.
Note: The compared models run locally; the AI evaluator (OpenAI GPT-4.1) is an external model.
Auto LLM BenchmarkTranscribes long audio recordings on your own computer.
What it does: Recognition runs with GPU acceleration and works fully offline after the initial setup.
Note: The optional text-processing scenario (AI structuring into a Google Docs article) uses an AI model and Google Docs.
Local transcription of large audio filesThese are reasons to consider it, not reasons to assume it.
Local is not inherently cheaper, faster or more accurate. Each of these has to be measured for the task.
A comparison of where processing happens, not a ranking.
| Approach | Where processing happens | What to check first | What stays external |
|---|---|---|---|
| Cloud AI | At an external provider | Provider terms, what data is sent, availability | The model and the provider’s infrastructure |
| Local AI | In a specified local environment | Hardware, model quality for the task, maintenance effort | Any other remote step in the workflow |
| Hybrid | Some steps local, some external | Which data goes where and how that is controlled | Every external step |
These terms are often mixed up.
A first approach to evaluate for each kind of need. Approaches can be combined.
| Need | First approach to evaluate |
|---|---|
| Stable output format | Prompting, examples, configuration, schema validation |
| Changing internal knowledge | Retrieval / RAG |
| Actions in organizational systems | Tools, APIs, bounded agent workflows |
| Domain-specific behavior | Evaluate whether adaptation is necessary |
| Persistent behavior change, enough suitable data, simpler approaches insufficient | A fine-tuning experiment |
| Deterministic task | Normal automation |
Fine-tuning is not the inevitable destination.
Each stage with its scope and how far current work supports it.
Supported by local comparison experiments.
Laboratory-scale local execution, not production hosting.
Measurement is supported; systematic optimization is developing.
A developing direction.
Planned LoRA / QLoRA R&D.
Supported by projects. No before/after fine-tuning results are claimed.
Experience from our projects; boundaries depend on each project.
A future operational stage that would need a separate assessment.
A dedicated laboratory AI compute environment is a concept for developing and demonstrating solutions. It is not an available managed service.
Demonstration concept Demonstration concept — no hardware, model, capacity or uptime is claimed
An application or automation server (a website, Telegram, n8n, a CRM) can call an AI inference server through an API. They can be different systems in different places.
Illustrative architecture — not a documented implementation
What does exist: laboratory-scale local experiments documented in the projects, such as local model comparison and local transcription.
A successful solution could eventually be reproduced in the organization’s own environment. Depending on requirements this may involve:
Models run in a specified local environment.
An architecture with access, isolation, storage and data-transfer controls.
Deployment within the organization’s own infrastructure.
Some steps local, some external.
Processing through an external provider.
The application and the inference run on different systems.
These are options to evaluate. The laboratory does not currently sell or operate all of them.
What the local-processing projects deliver.
The same tasks for different local models — quality and speed in one table.
In one run, about 60 hours of lectures were processed in about 5 hours on a GPU.
Directions, not available services.
A planned AI Model Engineering and R&D direction. It is not a launched, standard commercial service.
Before any future fine-tuning experiment we would require:
There is no price list and no purchase option for fine-tuning.
No. Private AI is about access, isolation, storage and data-transfer controls. Local execution is only one possible part of it.
Not by default. Cost, speed and accuracy have to be measured for your task.
Not as a standalone service at the moment. We build solutions that run in your environment or in the cloud — local, cloud or hybrid, depending on requirements.
No. LoRA / QLoRA is planned R&D, and simpler approaches are evaluated first.
Describe what is currently done manually, what data is involved and what result you would like to achieve. You do not need a technical specification or a chosen model to begin.
Please do not share trade secrets, source code, or sensitive personal data before an NDA (non-disclosure agreement) is signed.
