Private AI and LLMs, built on site
Standaard Tech sends engineers into your company to prepare the data, choose and tune the models, and install them on servers you own. For founders, we build AI MVPs. You keep the data, the models and the code.
What we do
Private LLM on your own servers
Open-weight models on hardware you own or rent, connected to your documents and systems. Assessment, pilot on rented GPUs, installation and handover.
How a private LLM setup worksAI MVP development for founders
A working AI product your first users can try, built by senior engineers in Riga. Fixed scope and price, weekly demos, code in your repository.
How we build AI MVPsFree tools
The same estimates we use when we size a private LLM.
LLM VRAM calculator
How much GPU memory a model needs for its weights and KV cache, by size, precision, context length and number of users, and which GPUs fit.
LLM cost calculator: API vs self-hosted
What your token volume costs on an API, on a rented GPU server and on hardware you own, and the volume where self-hosting breaks even.
Guides
- NVIDIA DGX Spark for LLMs: what it runs
What the DGX Spark can run: which model sizes fit its 128 GB of unified memory, how fast it generates, and when a company needs a GPU server instead.
- Hardware for local LLMs: GPUs, Macs and servers
Which hardware runs local LLMs for a person, a team or a company: GPU memory, bandwidth, workstation cards, Macs, DGX Spark and GPU servers compared.
- Open-weight models for companies: how to choose
What open-weight models are, how they differ from open source, the main model families and licences, and how a company picks one to run privately.
- Private ChatGPT for business: your options
How companies get a private ChatGPT: enterprise plans, cloud deployments in your region or open-weight models on your own servers, and what stays private.
- RAG vs fine-tuning: which one your LLM needs
When to use retrieval-augmented generation, when to fine-tune a model, and when to combine them, explained for teams building private LLM systems.
- vLLM vs Ollama vs LM Studio: which to use
vLLM, Ollama and LM Studio compared for running local LLMs: who each tool is for, how they handle many users, model formats, APIs and production use.
How an engagement runs
Scoping call
We talk through the process, the data and your constraints, and tell you plainly whether AI fits.
On-site discovery
Our engineers work beside your team, read the data where it lives and agree the success metric.
Build and prove
We prepare the data, test models on your cases and run a pilot against the baseline before you buy hardware.
Handover
We install the system for good, train the people who run it and hand over the runbooks.
Questions
What is a forward-deployed engineer?
An engineer who works inside the client’s organisation instead of at a distance. At Standaard Tech, forward-deployed engineers learn your process and data on site, then build and run the AI system with your team.
Can a local model match GPT or Claude?
On a narrow task with your own data, an open-weight model with retrieval and fine-tuning often comes close enough, and you control cost, latency and privacy. On open-ended reasoning, frontier APIs still lead. We test both on your cases before you buy hardware.
What hardware will we need?
That depends on the model size, the number of users and the response time you need. Many single-team setups fit on one GPU server; larger rollouts need a small cluster. The LLM VRAM calculator gives a first estimate; we size the final setup from load measured in a pilot.
How do you handle GDPR and the EU AI Act?
Data and models stay in your environment, so no personal data goes to a third-party AI provider. We document data sources, access rights, model versions and evaluations, which is the record the EU AI Act asks for.
Where do you work?
With companies in Europe and the United States. Engineers come on site for discovery and installation when the work needs it.
Book a scoping call
Describe the process or product in two sentences. We reply within one working day with questions and times for a 30-minute call. Or email hello@standaard.tech.