Local AI development and automation on GB10 and DGX Spark
The machine is 20%.
What we build on it is 80%.
We don't just sell AI workstations. We develop the automations that make them useful.
A GB10 server sitting on a desk does nothing by itself. It becomes useful when you connect your documents, processes, and daily tools to it.
This is our secondary business: designing and developing these automations, entirely locally, on the machine we deliver to you.
The three levels of AI automation
Not all companies start from the same point. Here are the three levels we encounter most often, from the simplest to the most sophisticated.
The conversational assistant
The starting point. A web interface accessible from any workstation, with your models installed and your documents indexed.
Your teams ask questions, write, summarize, query the document database. It's the equivalent of a consumer tool, but hosted in-house.
This level is delivered pre-configured with the machine. It requires no specific development.
Automated workflows
Here, the AI stops waiting to be asked a question. It triggers itself based on an event.
A document arrives, a form is filled out, a time sounds: a processing chain starts, calls the local model, produces a result, and distributes it where needed.
We build these chains with n8n, an orchestrator we host on your machine. Everything remains local, including the data that flows between steps.
Custom business application
The most advanced level. When no off-the-shelf tool matches your process, we develop a dedicated application.
Interface adapted to your business, deterministic calculations where necessary, local model invocation only where it adds value. All installed on your machine, with no external dependencies.
Developed in Python, with the interface that suits your use: desktop application, internal web interface, or integration into an existing tool.
The tech stack we install
Here are the layers we deploy on a GB10 or DGX Spark workstation, from the machine to your daily use.
Each layer is containerized with Docker. This allows updating one component without breaking others, and reproducing the installation identically on a second machine.
Examples of what we build
Here are some types of achievements, presented without identifying clients.
Automated invoicing chain
An order is validated on an online store. A webhook triggers the generation of the document, its formatting, then its automatic sending to the client by email.
n8n, document generation service, mail server, all in Docker containers.
Business application with transcription and writing
A professional records a session. The application transcribes the audio locally, enters numerical data into a structured form, calculates indicators with deterministic code, then has the narrative parts written by the local model. The final document is generated ready for review.
Python, local Whisper, verifiable calculations, local language model, document generation.
Queryable document database
Hundreds of internal documents are segmented, vectorized, and indexed on the machine. Teams query the entire set in natural language and get answers with references to source documents.
Ingestion pipeline, vector database, local embeddings model, discussion interface.
Automated monitoring and reminders
A script checks an inbox, identifies unanswered cases after a defined period, prepares reminders, and sends a summary notification to the team's internal messaging.
Python, mail access, scheduling, notification to Telegram or corporate messaging.
Why the GB10 is an ideal base
The NVIDIA GB10 platform, also known as DGX Spark or ASUS Ascent GX10, has a decisive characteristic for this type of project: 128 GB of unified memory.
Concretely, this allows running several components simultaneously that, on a classic graphics card, would compete for video memory: a large language model, an embeddings model for document search, a transcription engine, and application containers.
Two other advantages are important for a machine that runs continuously: power consumption of approximately 240 W and a compact, silent format, which allows it to remain in an office rather than in a technical room.
Mini AI server NVIDIA GB10
- Chip NVIDIA GB10 Grace Blackwell
- Memory 128 GB LPDDR5X unified
- Power 1 petaFLOP in FP4
- System DGX OS, Docker ready to use
- Consumption approx. 240 W
- Format 150 × 150 × 51 mm
How a mission unfolds
- Initial exchange. We identify the process to be automated, the volume concerned, and confidentiality constraints. No commitment required.
- Scoping and quotation. We precisely describe what will be delivered, the necessary hardware base, and the cost. You approve before we begin.
- Development. We build the solution and test it on a machine identical to yours.
- Installation. On-site or remotely, according to your preference, with the entire pre-configured environment.
- Training. Hands-on training for the people who will use the tool daily.
- Follow-up. Adjustments, adding documents to the index, updating models, and future evolutions.
Other machines suitable for these developments
CoreAI 32 — RTX 5070 Ti 16 GBSimple automations and moderately sized document base.
2 442 €
CoreAI 64 — RTX 5090 32 GBHigh throughput for heavy processing and bulk transcription.
6 042 €
CoreAI Rack — 2 × RTX 5090 (64 GB)Team-scale automations, parallel processing.
11 221 €
CoreAI 128 Rack — 2 × RTX 6000 (192 GB ECC)Continuous production, numerous users, ECC reliability.
27 980 €
To discuss an automation project, write to us at contact@radiancesystems.eu. We respond within 3 business hours, in French, English, or Dutch.
Frequently Asked Questions
What does Radiance Systems do beyond selling hardware?
We develop the automations that run on the machines: n8n workflows, custom Python applications, queryable document databases, transcription and writing chains. Everything runs locally, without transmitting data to an external service.
What are the GB10 and DGX Spark?
These are the same NVIDIA hardware platform, based on the GB10 Grace Blackwell chip, with 128 GB of unified memory. DGX Spark is the name of NVIDIA's reference design; ASUS Ascent GX10 is a variant. This unified memory allows running multiple AI components simultaneously on a single compact machine.
Why use n8n rather than an online automation service?
Because n8n can be hosted on your own machine. Online automation services transmit your data through their servers. With n8n installed locally, documents, messages, and results remain within your infrastructure.
Can local AI be connected to our existing tools?
Yes, in most cases. Messaging, file storage, online store, internal tools with an API: these are the usual connection points of a workflow. We check feasibility during the scoping phase.
How much does custom development cost?
It depends entirely on the scope. A simple automation represents a few days of work; a complete business application requires more. We provide a precise quote after the scoping discussion, before any commitment on your part.
Do we have to buy the machine from you to benefit from development?
It's not mandatory, but it's simpler. When we provide the machine, we control the entire environment and can test the solution on an identical configuration before delivery. However, we also consider requests for existing infrastructure.
What happens if our needs evolve?
Solutions are designed to evolve. Adding documents to the index, modifying a step in a workflow, or replacing a model with a newer version is part of the follow-up we offer.
The examples presented describe types of achievements and are deliberately anonymized. The scope, duration, and cost of a project depend on your processes and constraints; they are defined during scoping and formalized in a quote before any start.
