Artificial intelligence has taken over corporate conversations—often with more marketing hype than real substance. Here, we choose to put expertise first. This article is intended for specialists in particular, while offering everyone a glimpse of the excellence and rigor that fuel our projects every day.
a deliberate test-and-learn approach
As part of our AI initiative built around seven pillars (Training our people, Adapting our services, …), we launched several studies and pilot projects focused on how AI can be used in our day-to-day consulting activities.
One of these studies focused on setting up a dedicated “AI architecture” enabling us to work on confidential data while understanding the entire technological stack end-to-end—from hardware to user interface—without external dependencies.
The long-term objective is to support our clients with these topics, particularly with the architectural challenges of self-hosted AI within their Information Systems.
early discoveries
Our initial experiments led us to deploy an infrastructure hosted by a “local” managed-service provider.
This setup allowed us to deploy open-source LLMs and make them available to our teams as part of internal challenges and use-cases.
Key lessons learned:
- limited response times
- limited choice of models we could deploy
- dependency on the provider when adapting the infrastructure (understandable and even welcome for a built-and-operated environment, but more problematic during exploratory phases)
next step
To also address operational issues (LMOps), we decided to invest in professional-grade hardware designed for AI workloads, starting with NVIDIA’s DGX Spark and then the DGX Station.
The DGX Spark, powered by the NVIDIA GB10 Grace Blackwell superchip, delivers up to 1 petaflop of AI performance in a compact format with 128 GB of unified memory (we are also closely evaluating the DGX Station and awaiting the release of the GB300 chips, along with details on how they operate and their reliance on external GPUs).
These infrastructures will allow us to develop and test AI models with up to 200 billion parameters locally, with the option to interconnect systems for even larger workloads.
We chose to focus on architectures based on the Grace Blackwell superchips, such as the DGX Spark (1 petaflop, 128 GB unified memory), to benefit from a unique CPU/GPU convergence and maximize high-volume AI processing. We prioritize unified NVIDIA hardware due to the ease of integration offered by CUDA, anticipating a true technological shift we strongly believe in.
Note: We could also have opted for an AMD / Ryzen AI Max+ architecture. However, it requires using FastFlowLM backends instead of Ollama.
backend
Our tests currently cover several state-of-the-art open-source models, including:
- Mistral AI (high-capacity multimodal models)
- Llama 4 (Mixture-of-Experts architecture supporting several million tokens)
- OpenAI’s GPT-OSS, optimized for self-hosting
- Apertus, the Swiss model developed by EPFL – ETH Zurich – CSCS
We do not intend to build our own models; instead, we aim continuously to deepen our expertise by staying close to open-source innovation.
frontend
We are exploring solutions that, beyond offering an ergonomic user portal, provide robust user and access-right management.
We expect a frontend solution to:
- support both on-premise and cloud-based models
- enable fine-grained API access control
- support RAG workflows on large document corpora without exposing those documents externally
We began our tests with Ollama Web UI and are actively evaluating alternatives such as LibreChat, AnythingLLM, LobeChat, and Text Generation Web UI.
why this approach
This strategy enables us to:
- Understand the technology: we maintain genuine expertise and stay aligned with weekly advances
- Control our data: everything remains on our infrastructure when handling sensitive or confidential information
- Experiment safely: tests take place entirely in our own environments
- Recognize limitations: we evaluate the real constraints and operational costs of AI tools. Some processes may not require AI at all. We are developing methodologies and tools to rigorously measure the added value of AI compared with simple automation, avoiding over-engineering
- Empower our teams: providing a safe, dedicated environment for learning and experimentation
what comes next
Preparing for the future: this experimentation phase will enable us to bring our expertise to clients and support them in defining and implementing their AI strategy.
Our goal is twofold:
- help you choose the most suitable models for your use-cases and integrate them effectively
- provide concrete, experience-based feedback focused on the real value created for the business
No marketing promises, just concrete results backed by our expertise.
mehr Insights
Tricentis Transform London 2026: Agentic AI is redefining the boundaries of Quality Engineering
company newsdata & AIquality assurance & testing
Oktober 30, 2026
Upcoming Itecor Testing Breakfast – AI for Software Quality and Value Creation
data & AIquality assurance & testing
Oktober 12, 2026
[press release] Itecor appoints Amine Mansour to accelerate the impact of data and AI
company newsdata & AI
Juni 15, 2026
