--- title: "CSIDIA at a glance: AI agents on your company’s ontology and private AI with open-weight models" description: "CSIDIA builds a system that brings your company’s people, documents, case files, processes and data into a single ontology, governed by AI agents that work for people; and it trains and fine-tunes open-weight AI models on your task, running them privately on our NVIDIA infrastructure or on-premise. What we do, who it is for, how we work, and direct answers." url: https://csidia.com/en/at-a-glance/ lang: en publisher: CSIDIA S.r.l. (https://csidia.com/) --- At a glance # CSIDIA at a glance. CSIDIA S.r.l. (Vignola, Modena, Italy) builds, for a company or public body, a system that brings its people, documents, case files, processes and data into a single ontology, that is a shared model of the organisation, and has it governed and connected by AI agents that work for people: within each person’s rules and permissions, with a person’s confirmation before the actions that matter. Examples of these agents are a personal assistant that coordinates a system of agents, the agent of each client on a professional firm’s platform, and an agent that runs an organisation’s day-to-day case files. Alongside the system, CSIDIA trains and fine-tunes open-weight AI models on the client’s task, with its own data, and runs them privately: on CSIDIA’s NVIDIA compute infrastructure, or on-premise, on infrastructure at the client’s site that CSIDIA designs, installs and manages. On that task, a model trained this way works better than a general-purpose model; what it learns from the data stays with the client, and there are no per-token fees to a closed-model vendor. The facts ## Who, what, where, for whom. **Who** CSIDIA S.r.l., Via Giacomo Leopardi 20, 41058 Vignola (MO), Italy. VAT IT04190980369. **What** A system that brings people, documents, case files, processes and data into a single ontology, governed and connected by AI agents that work for people, with rules, permissions and confirmations. Alongside, open-weight AI models trained and fine-tuned on your task, with your data, and measured on your cases. **Where the models run** On our NVIDIA compute infrastructure (an NVIDIA GB300 server, a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q), or on-premise at your site, on infrastructure we design, install and manage. **For whom** Companies, professional firms and public bodies with information spread across many programs, confidential data, or a precise, recurring task. **How to start** A first meeting: 30 minutes, free, video call or in person. **Contact** [engagement@csidia.com](mailto:engagement@csidia.com) · [Book a meeting](https://csidia.com/en/book/) What we do ## Six things, always the same. 1. We bring people, documents, case files, processes and data into a single ontology, that is a shared model of the organisation, and have it governed and connected by AI agents that work for people: within their rules and permissions, with their confirmation before the actions that matter, and with a log of every step. 2. We choose the open-weight model that fits your task, from those we test and keep in a model register (what it does, where it runs, licence, test results, why it was chosen). 3. We train it on your data: by fine-tuning an existing model or, when no existing model is enough, with networks of our own trained from scratch. 4. We measure it on your cases, with numbers you can check, before putting it to work. 5. We run it privately: on our compute infrastructure (an NVIDIA GB300 server and a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q) or on-premise at your site. 6. We design, install and manage on-premise infrastructure: machines, network, power, cooling, security, updates, and new models verified before they are adopted. On broad tasks, such as writing code or answering anything, large general-purpose models stay ahead. We work on one task at a time, and we measure it. Who it is for ## If one of these is you, we should talk. - Companies, professional firms and public bodies whose people, documents, case files and data are spread across many programs, and who want a single system with AI agents that work for people, with rules, permissions and confirmations. - Companies and public bodies that want to use AI on confidential data (technical documents, archives, designs, case files) without sending it to a closed-model vendor. - Anyone with a precise, recurring task where a model trained on their own data does better than a general-purpose model: checking documents against standards, extracting data from hard archives, generating custom images, preparing work on mail and case files. - Anyone paying per token for a closed model today who wants their knowledge to stay in the company. - Anyone who wants on-premise AI, in house, but has nobody to design and run the infrastructure. - Machinery manufacturers and technical offices that must show their manuals and technical files comply: Regulation (EU) 2023/1230 applies from 20 January 2027. How we work ## Four stages. You can stop after any of them. 1. **A meeting** Video call or in person, 30 minutes, free. You tell us the purpose; we tell you whether it is a job for a trained model, what it would take (data, compute, time) and which route fits. If it is not a job for us, we say so. 2. **A trial on your documents** On a sample of your documents or data, under a non-disclosure agreement, on our infrastructure. We measure the result on your cases. The trial is agreed at the first meeting, tailored to you. 3. **Training** We choose the model, train or fine-tune it on our infrastructure and measure it again. Every choice is written down: which model, which licence, why. 4. **In service** The model runs on our infrastructure or on yours, which we install and manage. The AI prepares, a person decides. You can stop after any step. Services ## What you can ask us for. ### A system of AI agents on your company’s ontology People, documents, case files, processes and data in a single ontology, governed and connected by AI agents that work for people: within each person’s rules and permissions, with a person’s confirmation before the actions that matter. Each agent uses the model best suited to the task, open-weight when the data cannot leave. [Details](https://csidia.com/en/what-we-do/#system) ### A model trained for your purpose An open-weight model chosen for your task and trained or fine-tuned on your data, measured on your cases. On that task it works better than a general-purpose model; what it learns stays yours. [Details](https://csidia.com/en/what-we-do/) ### Your model, on our infrastructure or yours We train and run open-weight models on our compute (an NVIDIA GB300 server, a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q), or design, install and manage on-premise infrastructure at your site. [Details](https://csidia.com/en/infrastructure/) ### Checking technical documents against standards Every requirement of the standards that apply to you (for machinery: the Directive, Regulation (EU) 2023/1230, harmonised standards) searched for across the whole document, page by page, with the outcome and the citation. Your expert signs. [Details](https://csidia.com/en/capabilities/#standards) ### Reading hard archives and documents Manuals, scans, drawings and tables turned into searchable data, without losing the codes and without sending the pages outside: a chain of open-weight models chosen and tuned for the task. [Details](https://csidia.com/en/capabilities/#archives) ### Custom-generated images of materials One model per type of material, trained on your images, that generates new surfaces at true scale, pixel by pixel, and measures them against the real ones. [Details](https://csidia.com/en/capabilities/#images) ### An agent on your company’s data and mail It reads mail and documents, keeps track of deadlines and case files and prepares the work, with the model best suited to the task, open-weight when the data cannot leave. It proposes; a person decides. [Details](https://csidia.com/en/capabilities/#agent) Direct answers ## The questions people ask us, answered. ### What does CSIDIA do? It builds, for a company or public body, a system that brings people, documents, case files, processes and data into a single ontology, governed and connected by AI agents that work for people, within their rules and permissions and with their confirmation. Alongside, it trains and fine-tunes open-weight AI models on the client’s task, with its own data, and runs them privately: on CSIDIA’s NVIDIA compute infrastructure or on-premise, on infrastructure at the client’s site that CSIDIA designs, installs and manages. ### What does it mean to bring a company into a single ontology? It means building a shared model of the organisation, in which every person, document, case file, process and piece of data has one name, one place and its links to the rest: who handles which case file, which document is needed for which deadline, which process touches which data. On this model, people and AI agents find what they need without searching for it in many different programs. ### What do the AI agents do in the system? They work for people: they read mail and documents, keep track of case files and deadlines, prepare replies and texts and carry out the permitted actions in the company’s programs. Each agent has written rules and works within the permissions of the person or organisation it works for; the actions that matter start only with a person’s confirmation, and every step is logged. ### What examples of agents have you built? Three, each for a different job. A personal assistant that, on a workstation, coordinates a system of agents, one per project, and handles mail, calendar, WhatsApp, the work queue and the log; the actions that matter start with the person’s click. The agent of each person and client organisation on a professional firm’s platform: it works within that person’s boundaries and permissions and carries out the platform’s actions with their confirmation. An agent that runs an organisation’s day-to-day case files (compliance duties, mail, deadlines, documents) inside its platform: it proposes changes and does not make them on its own. ### Which model do the agents use? The model best suited to the task, chosen with you. For everyday work it can be a frontier model; when the data cannot leave, an open-weight model running on our infrastructure or yours. The agent for the firm’s clients, for example, uses a frontier model for normal work and an in-house open-weight model for sensitive data; in the personal assistant’s system, clients’ technical manuals are read only by an open-weight model on our machines, never by a cloud model. We always tell you which model reads what. ### What is an open-weight model? An AI model whose weights, the learned parameters, can be downloaded, installed and run on your own machines, instead of being used over the internet from a vendor and paid per use. It can be trained and fine-tuned on your own data. Open weights do not mean a free licence: every licence has to be read, and we check it before using a model. ### Do you fine-tune models? Do you train models from scratch? Yes. We fine-tune existing open-weight models on your data and, when no existing model is enough, we train networks of our own from scratch, for example to generate images of materials at true scale. Sometimes it is enough to put an open model to work inside a procedure built around your case: if so, we tell you. ### Does a model trained on our task do better than a large general-purpose model? On that task, yes: it knows your documents, your language and your rules, and that is what we measure on your cases before putting it into service. On broad tasks, such as writing code or answering anything, large general-purpose models stay ahead: we work on one task at a time. ### Where does the model run? Does our data leave the company? You decide: on our compute infrastructure, or on-premise, on infrastructure at your site that we design, install and manage. With your own infrastructure the data stays in the company. The model that works on your data is an open-weight model, and we always tell you which one it is and where it runs. ### Does what the model learns from our data stay ours? Yes. What the model learns from your data stays yours and does not become training material for someone else’s AI. ### What hardware do you use? Our infrastructure: an NVIDIA GB300 server and a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB of memory each. On the server the large model handles up to 40 requests at once, the usable capacity we measured. For infrastructure of your own, we size the machines on your workloads. ### Which models do you use? Open-weight models chosen for the task, each recorded in the project’s model register with its licence, test results and the reason it was chosen. Our large model is Qwen3.8-Flash-Next, open-weight, about 180 billion parameters in total; to read documents we also use small models that are already specialised. When a better open model comes out, we verify it on your cases before adopting it. ### Do we need to buy GPUs or servers? No. You can start on our infrastructure. If you later want the model at your site, we design, build, install and manage infrastructure of your own: network, power, cooling, security and updates. ### What do we need to get started? A purpose, your real documents (not ones rewritten for us), one person who decides and some time from someone who knows the process, to say whether a result is right. The data does not need to be in order: sorting it out is part of the work. ### What does it cost to start? The first meeting is free and lasts 30 minutes. The trial on your documents is agreed at the first meeting, tailored to you. A model in service has no per-use fees to pay to an AI vendor. ### Who decides what the AI does? A person. The AI prepares, a person checks and decides: no text goes out without a person’s approval, and agents carry out the actions that matter only with that person’s confirmation. ### What do you not do? We do not give legal opinions (we read the rules and build the checks; the opinion stays with your lawyer), we do not promise “better at everything”, we do not quote numbers we have not measured and we do not name our clients. ### Where are you and how do we get in touch? CSIDIA S.r.l., Via Giacomo Leopardi 20, 41058 Vignola (MO), Italy; VAT IT04190980369. Book a meeting at csidia.com/en/book/ or write to engagement@csidia.com: a person replies. Terms ## The words people use to find us, and what they mean here. **Company ontology** The shared model of an organisation: people, documents, case files, processes and data, each with one name, one place and its links to the rest. It is the basis our agents work on. **AI agent** A program that uses an AI model to do a job on a person’s behalf: it reads, prepares, proposes and carries out the permitted actions, within that person’s permissions and with their confirmation for the actions that matter. **Private AI** An AI model that runs on our infrastructure or yours, not at a closed-model vendor: you know which model reads your data and where it runs. **Open-weight models** Models you install and run on your own machines, and can train on your own data. They are the basis of every model we train. **Task-specific training and fine-tuning** Teaching a model a precise task with your data: by fine-tuning an existing model or training one from scratch. On that task the model does better than a general-purpose one. **On-premise AI** The infrastructure and the model at your site: we design, install and manage it, and the data stays in the company. **Per-token pricing** How large closed models are paid for: every piece of text read and written has a price. A model running on our infrastructure or yours has no such fee. The first step ## Book a meeting. Tell us the purpose. We tell you whether an agent or a trained model would do it better than you do today, what it would take (data, compute, time) and which route fits: our infrastructure or yours. If it is not a job for us, we say so. - video call or in person - 30 minutes - free [Book a meeting](https://csidia.com/en/book/?servizio=misura) [or write to engagement@csidia.com](mailto:engagement@csidia.com?subject=First%20meeting%3A%20A%20model%20trained%20for%20your%20purpose) The service A model trained for your purpose An open-weight model trained on your data, on our infrastructure or yours. What it learns stays yours. [What we do](https://csidia.com/en/what-we-do/) [Capabilities](https://csidia.com/en/capabilities/) Page: https://csidia.com/en/at-a-glance/ · CSIDIA S.r.l., Vignola (MO), Italy · engagement@csidia.com · index for agents: https://csidia.com/llms.txt