Skip to content

Requirements & installation

Hardware

Teleutias is delivered and configured on a dedicated workstation, sized according to the number of users and the expected workloads.

Reference configuration:

Component Indicative specification
GPU NVIDIA with 16 GB of VRAM (e.g. RTX 5060 Ti 16 GB)
RAM 32 GB or more
Storage NVMe SSD, 1 TB recommended (models + knowledge base + logs)
Operating system Windows or Linux, according to the client's IT preferences

Why a dedicated GPU

Local language models require a GPU to respond within timescales that suit daily work. CPU-only inference is technically possible but too slow for operational use: the GPU is not an option, it is the central requirement.

The machine can be supplied by Teleutias fully configured, or installed on the client's existing hardware if compatible.

Network

  • The interface is reachable from the company's local network via browser: nothing needs to be installed on users' PCs.
  • No exposure to the public internet is required. The machine accepts no inbound connections from outside.
  • For remote support and maintenance, an encrypted point-to-point remote access channel (mesh VPN, e.g. Tailscale) can be enabled, in agreement with the client, with limited permissions revocable at any time.
  • An internet connection is only needed during installation and for scheduled updates; day-to-day operation is entirely offline.

Commissioning procedure

  1. Initial analysis — gathering use cases, reference documents and the client's IT constraints.
  2. Machine preparation — installation of the inference engine, web interface and selected models; load testing.
  3. Assistant configuration — instructions, operational limits and knowledge base for each activated vertical.
  4. Connection to the client's network — LAN integration, user and permission setup.
  5. Kick-off session — user training on the assistant's capabilities and limits.
  6. Verification period — initial support with real but selected use cases, and configuration tuning.

Maintenance and evolution

Teleutias is not a piece of software you install once and leave untouched. Language models evolve quickly and the client's operational context changes: ongoing maintenance is what makes the assistant improve over time rather than age.

What it covers

Model evolution. Open-source models progress month by month: on the same hardware, what is available today will be surpassed within the year. We evaluate new models, test them against the client's actual use cases, and put them into production only after verification. The benefit is direct: the same machine, better answers.

Knowledge base updates. Internal procedures, circulars, manuals and communication templates change. If the loaded documentation is not kept current, the assistant quietly gets worse. Periodic content review is part of the service.

Configuration tuning. Instructions, operational limits and assistant behaviour are refined against real usage — against how the firm actually works, not how it was imagined during analysis. User reports of incorrect answers feed directly into this work.

Component updates. Inference engine, web interface and libraries are kept at supported versions, with incompatibilities checked in advance.

Adapting to business systems. New versions of ONE Start, ONE 200 or Odoo may change formats, error messages or data structures: the assistant is realigned accordingly.

Integration maintenance. When the assistant is connected to a business system via API, that connection is live software and needs active oversight: APIs change with versions, credentials and certificates expire, and regulatory formats (Swissdec, ISO 20022, QR-bill) are updated by law. Maintenance includes periodic verification that connectors respond and return consistent data, and realignment whenever the business system changes.

Why active oversight is necessary

An integration can degrade without any visible error: the assistant keeps answering, but on incomplete or stale data. This is why verification is scheduled and periodic rather than on request.

The integrations under active support are ONE Start, ONE 200 and Odoo. Other business systems with public APIs (bexio, for example) can be evaluated as a dedicated project.

Adapting to a new major version of a business system (a major release upgrade, for instance) is a separate project, quoted independently from ordinary maintenance of the existing connector.

Operational support. Help for users with day-to-day usage and with the questions that emerge from experience.

How it works

  • Interventions are scheduled and agreed with the client: no uncontrolled automatic changes.
  • Every new model is verified before going into production, against the client's use cases.
  • The client may request the status of installed components and versions in use at any time.
  • A periodic review takes stock of usage, results and possible improvements.

Why it matters

Those who use cloud AI get continuous improvement, but no control over their own data. Those who install a local system and then leave it untouched get the control, but an assistant that ages. Ongoing maintenance is what makes both possible: data that stays in-house, and an assistant that will work better next year than it does today.