Fast video memory
This holds the most-used layers of the model and the users' context. More GPUs mean more memory and more people served at once, each with their own space.
bampa AI · on premises
We design and install AI servers that run top-tier open-source models for the whole company, on your documents, your code and your projects. No pay-per-use subscriptions, no data leaving the building.
Use cases
For the P200 the specified torque is 38 N·m, crosswise in two passes: first at 50%, then at full value. Since revision 4 this also applies to spindles with an HSK interface.
Illustrative example: every answer cites the document and the page.
Manuals, procedures, specifications and standards you can query in plain language, with the source cited.
An assistant on your private code: reviews, tests, documentation, legacy code to explain and migrate.
The model drives CAD through deterministic tools: parametric parts, assemblies, bills of materials, dimensional checks.
Schematics, netlists, footprints, routing rules, reading DRC and ERC errors, normalised BOMs.
Summaries, version comparisons, extraction of tables, clauses and parameters from thousands of PDFs.
Reading tenders, checking requirements, a first draft of the quote on your own templates.
Work instructions always at hand, consistency checks across documents, audit support.
The expert colleague available to new hires and internal help desks, around the clock.
Transcripts, minutes and technical translation, without recordings shipped to outside servers.
Workflows that read, check and report on their own, with permissions and a log of every action.
Why local
The cloud is convenient until it comes to manuals, drawings, code, quotes and contracts. For subcontractors, medical, defence or anyone under NDA, that's where the conversation ends. On premises, it starts again.
Documents, conversations and code never leave the company network. It works with the internet cable unplugged.
Forty years of procedures, solutions and tricks of the trade become searchable, even after the people who know them retire.
No charge per generated word: the hardware is depreciated, and the more you use it the better it pays.
Interchangeable open-source models: when a better one comes out, you switch with one line.
Dozens of people at the same time, with per-department permissions and a log of who saw what.
GDPR, trade secrets, customer clauses: if the data never leaves, half the questions disappear.
ERP, archive, code editors, CAD and EDA, through standard interfaces (API and MCP).
No usage caps, no queues, no service changing its terms overnight.
The model learns your vocabulary, your abbreviations and the way you write procedures.
The architecture
A large model isn't sized by the power of the graphics card, but by memory: how much there is and how fast it is. That's why we combine GPU memory with the RAM of an 8-channel server processor.
This holds the most-used layers of the model and the users' context. More GPUs mean more memory and more people served at once, each with their own space.
Hundreds of GB of error-correcting memory for models that don't fit in the GPUs, long contexts, document indexes and always-on agents.
RAM, GPUs and storage are added one at a time. Start with what you need today and expand when usage calls for it.
Try sizing it
Pick the model, how many people use it at once and how much text they work on each time. The figure is indicative: real sizing is done on your own cases.
Memory needed
0 GB
model 0 GB · user context 0 GB
The two-model workflow
When you need the most capable model on the market, use it only to invent the procedure, on a non-confidential example. Then the local model applies it to hundreds of real files that never leave.
understands the problem on a clean example
writes rules, tools, examples and tests
applies the procedure to all your data
checks format, consistency and errors
only the few ambiguous cases, anonymised
How to start
One or two days on site: documents, use cases, sizing, comparison with the cloud. You get a document, not a promise.
from €1,500Two weeks on real data with 5–10 users. At least 16 correct answers out of 20, with the source: otherwise we refund half.
from €3,500Sized hardware, models, archive, permissions, backup, two-level training.
on quotationNew models as they come out, new documents, monitoring. The system improves while you use it.
monthly feeIndicative prices, excluding VAT. Technical documentation and reference architecture on request, under a non-disclosure agreement.
Contact
A robot, an AI server for your company or a project that doesn't have a name yet: drop us a line and we'll reply within one working day.
info@bampa.netBased in Auronzo di Cadore and Bolzano, Italy.
The form opens your email app: your data never passes through third-party services. Privacy