--- title: "Capabilities: four examples of open-weight models at work — CSIDIA" description: "Four CSIDIA capabilities: checking technical documents against standards, reading hard archives and documents, generating and measuring custom images of materials, an agent that prepares the work on data and mail. With open-weight AI models, on your machines or ours; the agent with the model best suited to the task, open-weight when the data cannot leave." url: https://csidia.com/en/capabilities/ lang: en publisher: CSIDIA S.r.l. (https://csidia.com/) --- Capabilities # Four capabilities already put to the test. Each starts from a problem many companies have. For each one we choose the model suited to the task, specialise or train it and run it on your machines or ours: open-weight when the data cannot leave, as in the first three examples. We measured the numbers ourselves, in our own trials. Documents and standardsArchives and documentsImages and materialsData and mail Documents and standards Open-weight modelSpecialised for the task ## Checking technical documents against standards, page by page The problem Manuals, technical files, instructions: to show that they comply with a directive, a regulation and the harmonised standards, hundreds or thousands of requirements have to be checked one by one, searching the document by hand. For machinery manufacturers, Regulation (EU) 2023/1230 applies from 20 January 2027. What we build A check built on the standards that apply to you. For each requirement the model searches the whole document, reads it more than once, decides and writes the outcome with the citation: the page and the passage it comes from. The model is open-weight, chosen for the task and specialised by the procedure of the check, and runs on your machines or ours. What you get The report requirement by requirement, with the outcome and the citation, which your expert checks and signs. When the document or the standard changes, the check is run again. - In one of our trials on four machinery manuals - 4,346 requirements checked, each with its outcome and the page of the manual - 0 requirements skipped: every item was read - 111 min to read the 10 standards of the check, 972 pages Archives and documents Open-weight modelsChosen and tuned for the task ## Reading hard archives and documents, and extracting the data The problem PDF manuals, scanned pages, technical drawings, tables running over several pages: they hold the codes, values and spare parts needed every day, but nobody can search them. Automatic reading tools, used on their own, lose exactly the codes. What we build A chain of open-weight models chosen for your documents: small models for text and tables, a large model for figures, drawings and scans. We put them together and tune them for the task; at the end a check makes sure no word of the PDF is lost and compares every number with the page. Everything happens on your machines or ours. What you get For each document a structured file, with the position and source of every piece, and text ready to search or to pass to your software. - In one of our trials on manuals never seen before - 98.4% of codes and values found, on 1,174 pages of technical manuals from different manufacturers (checked on 50 pages drawn at random) - 81.4% the best reading tool used on its own, on the same pages - 96.4% · 97.2% of codes on scanned pages and on CAD drawings - 4.8 s per page, on average - The measurements are on public manuals downloaded from the web, not on client documents. Images and materials Trained by usOne model per material ## Generating and measuring custom images of materials, pixel by pixel The problem Anyone who makes or sells a material recognised by sight, such as a stone, a wood or a fabric, needs new, credible images: at true scale, at the resolution of their best photographs, neither copies nor enlargements. What we build One model for each type of material, trained by us on your images: an open model fine-tuned on your files, or networks of our own trained from scratch. The detail is generated, not enlarged. First we stitch the photographs of real surfaces into a single image at true scale: it is the yardstick against which we measure the generated ones. Training and generation run on our machines or yours. What you get The generated images, at true scale, to look at enlarged down to the detail, and a separate model for each type of material. - Our measurements - 1 h 33 min to generate a full surface of about 3 by 2 metres, at 520 dpi - 0.17 pixels the median error on the seams, stitching 36 photographs of a real surface into a single image of about 2.9 by 2 metres - Always every generated image is measured automatically against the real ones: grain, direction and length of the veins, colour, copied pieces Data and mail The model suited to the taskA person decides ## An agent on data and mail, with a person who decides The problem Deadlines with several authorities, case files to follow, documents to keep in order, communications to prepare and mailboxes always full. The information is there, but scattered: whoever holds it together spends the day looking for it. What we build A shared model of the organisation, that is a single ontology, for deadlines, contacts, rules, documents and communications, and an AI agent working inside it: it reads the mail, follows the case files, prepares replies and texts. The agent works with the model best suited to the task, open-weight when the data cannot leave: in that case on your machines or ours. It changes nothing on its own: it proposes, and a person decides. What you get The work already prepared, one place to find deadlines, documents and communications, and the log of every step of the agent: what it read, what it proposed, who decided. - How the agent works - Reads the mail, the documents and the deadlines, where they already are - Prepares replies, texts and case files, ready to check - Proposes the changes, without making them on its own - You decide: no text goes out without a person’s approval How to read these numbers ## Measurements on a precise task, not general promises. The numbers on this page come from our own trials, one task at a time. We do not promise “better at everything”: in broad work, such as writing code or answering anything, large general-purpose models stay ahead. On your task, the first thing we discuss is how we would measure it. Is your problem a different one? ## The method is the same. We choose the right open model, specialise or train it on your data, measure it on your cases and put it to work where you decide. Tell us the task. [What we do, in three steps](https://csidia.com/en/what-we-do/) 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/) Page: https://csidia.com/en/capabilities/ · CSIDIA S.r.l., Vignola (MO), Italy · engagement@csidia.com · index for agents: https://csidia.com/llms.txt