AI for architects and engineers, inside your firm
An AI for architecture and engineering firms that reads your reports, specifications, bills of quantities and drawings, answers with the exact code clause, and drafts technical reports from your own templates. It runs on a server in your office: client projects never leave it.
AI for architects and engineers, not for renderings
When people talk about AI for architects they usually mean renderings from a prompt. This is a different problem: in a design office, time goes into searching. Which section of the building code applies, what the official commentary says, how we solved the same detail on a project three years ago, where the right version of the specification is.
The answer is an assistant that knows your documents: 10, 50 or 200 GB of PDFs across reports, specifications, bills of quantities and drawings, plus the codes and standards you use every day. You ask a question in plain language and get an answer with the exact reference and a link to the source. All on the firm's own hardware, without sending a single file to ChatGPT or any other cloud service.
What AI does for an engineering or architecture firm
Search building codes and standards
National codes and their commentaries, the standards the firm already licenses, local building regulations. Ask "what checks does an existing timber floor need" and get the answer with code, chapter and clause to open and verify.
Search past projects
"Have we done a masonry vault retrofit before?" The archive becomes searchable by project, client, year and type of work. You find the report, the drawing and the cost item you need, not thirty PDFs to scroll through.
Read drawings and tables
Title blocks, legends, material schedules, cost summaries, bill-of-quantities items: a model that reads images turns them into structured text, so even a scanned PDF or a drawing exported as an image becomes searchable.
Draft technical reports
Design statements, material reports, the descriptive parts of calculation reports: the AI starts from the templates your firm already uses and the project's own documents, and hands you a draft to review and sign. Responsibility stays with the professional; the copy-paste work doesn't.
From PDFs to an answer that cites the clause
Reading the documents (OCR and vision)
Every PDF is read page by page. Text is extracted directly; drawings, tables and scans go through a model that reads images and returns text and tables with their structure intact. A bill of quantities stays a bill of quantities, with codes, descriptions and quantities.
An index by project and by code
Content is split into passages and indexed with its references: project, document and page for the archive; code, chapter, section and clause for standards. That is what makes it possible to cite the source instead of summarising from memory.
Answers that cite code and clause
When you ask a question, the system retrieves the relevant passages and the model answers from those alone, showing where each statement comes from. If the answer isn't in the documents, it says so.
Drafts from your firm's templates
For reports, the AI uses your templates as structure and the project documents as content. The result has your firm's format and tone, ready for the engineer's review.
RAG, not training: why it matters
To make the AI "know" your documents you don't train the model: you use RAG (retrieval-augmented generation), retrieving the right passages and handing them to the model with the question. A code update or a new project enters the index in minutes, and every answer has a source you can check. Fine-tuning only makes sense for style, report format or document classification, not for teaching content.
On-premise AI: your projects stay in the office
Projects, land registry data, geotechnical reports, client details: this is data you don't want on a third-party service, and often your client contract doesn't allow it. That is why the whole system runs locally: PDF reading, index and the model that answers. No external API calls, no document leaving the firm's network.
An open-weight model of around 30 billion parameters is enough for the answers: on a workstation or small server with a suitable GPU it is fast enough for a firm's daily use and capable enough to reason over technical text. Hardware is sized on document volume and number of users, and we work it out together before you buy anything.
Private AI vs ChatGPT vs industry software
| Private AI in your office | ChatGPT and generic chatbots | Industry software with AI | |
|---|---|---|---|
| Where documents end up | On the firm's server | On the vendor's servers | Depends on the vendor |
| Knows your past projects | Yes, the whole indexed archive | Only files uploaded to that chat | Usually only that software's files |
| Cites code and clause | Yes, with a link to the source | Often invents or approximates references | Varies |
| Drafts in your firm's format | Yes, from your templates | Generic format | The software's format |
| Best when | Large archive, confidential data, daily use | General questions, no confidential data | You already run everything in that software |
An engineer who builds AI systems
I'm Stefano Salvucci, a registered engineer (Order of Engineers of Rome) with 12+ years across software development, systems architecture and automation. I know from the inside how a design office produces and searches documents, and I work every day with language models, RAG and open-weight models. I design the system, install it in your office and hand it over to the people who use it.
Questions about AI for architects and engineers
Can I just use ChatGPT for building codes?
For a general explanation, yes; for a reference that goes into a report, no. A generic chatbot doesn't work from the text of the code, so it can cite wrong or outdated sections. A RAG system answers only from indexed documents and shows the passage it used, so you can check it.
Do client documents leave the office?
No. PDF reading, the index and the model all run on the firm's hardware, with no calls to cloud services. That is the reason to choose on-premise AI over a chatbot subscription.
Does it work with scanned drawings and tables?
Yes: pages without selectable text, tables and drawings go through a model that reads images and extracts text and structure. A very dense drawing is harder than a report, and we test on your real documents before indexing the whole archive.
How many documents can it handle?
The typical case is an archive of 10 to 200 GB of PDFs. The practical limit is not storage but the time of the first pass over drawings and scans, which happens once; after that, new documents are added as they come.
Does it write technical reports for me?
It writes drafts: structure from your templates, content from the project documents, code references cited. Checks, design decisions and the signature stay with the professional.
Do we need to train a model on our documents?
No. RAG is how the system knows your documents: it updates in minutes and cites sources. Fine-tuning is only worth it to imitate a very specific style or format, or to classify documents automatically.
Is this an AI for architects that also makes renderings?
No. There are plenty of dedicated tools for renderings and concepts. This covers the document side: codes, archive, reports. It is where a firm loses the most hours and where data confidentiality matters most.
How do we start?
With a free 30-minute intro call: tell me how your firm works and where time gets lost. If it makes sense, we start with a prototype on a sample of your documents, so you see the answers before deciding to install everything.
Let's see if it makes sense for your firm
30 minutes, free: we assess where AI can save your studio time.