Our local AI lab in Barcelona

What is in the lab

Component	What it is	What we use it for
Inference cluster	Two NVIDIA DGX Spark units joined by a 200 Gb/s link	Open frontier models of 180 to 320 billion parameters, with up to a million tokens of context
GPU server	Three NVIDIA RTX cards (one 3090 and two 4060 Ti)	A very fast coding model, voice, transcription, vision and image generation
Private gateway	An OpenAI-compatible API, with accounts per person and per agent	So that every user and every agent has their own key, budget and usage dashboard
Agents	Hermes Agent, OpenCode, Pi and Cursor	Development, system operations, reports and office tasks

What it does every day

• Private chat for several people, with reasoning, coding and vision models, with no subscriptions and no dependence on the cloud.
• Document search with cited answers.
• Meeting transcription that tells speakers apart, in Catalan, Spanish and English.
• Agents that work overnight: they review, write code, prepare reports and raise the alarm if something fails.
• A custom model in training with LoRA fine-tuning for a Barcelona educational project on music and maths.

What we have learnt so that your AI does not fail

• Choosing the model according to the number of users, not according to whichever ranking is in fashion.
• Measuring before promising: real speed, power consumption and what happens when ten people ask at the same time.
• Sharing the GPUs properly, with automatic scheduling between tasks, to get the most out of the hardware.
• Switching models without interrupting the service: in 2026 we updated the main model three times in five days without anyone noticing.
• Clear rules for agents: human confirmation for any important change and no touching production without permission.

Come and see it

If you are in Barcelona, we will show it to you in person. If not, on a video call with the system running live. Book a demo.