ARVENT
I connected AI to building models before it was a trend.
- Python
- FastAPI
- Speckle (GraphQL, webhooks, specklepy)
- Ollama
- Llama 3.1
- SQLite
Most people on a building project never open the 3D model, but they still have questions about it. ARVENT lets them ask in a comment, for example how many doors have no fire rating, and replies with the answer, the exact data behind it and a link that shows those parts in the model. It also checks each new version of the model against the project's written requirements.
The problem
Most people on a project never open the model, but they still have questions about it: what this panel's insulation value is, how many doors have no fire rating, whether the latest version still meets the specification.
What it does
In 2023 I connected Speckle, an open platform for sharing 3D building models, to a large language model, so a project team could ask a model questions in plain language. ARVENT (Assistant for Requirements, Verification and Element Notes in Threads) is the 2026 rebuild. It runs on a local Llama model through Ollama, so model data and project documents never leave the machine.
It lives in Speckle's comment threads. Select an element and ask about it, and it answers from the element's properties and any attached specification PDF, with page references. Ask about the whole model and it answers that too. Every time a new version comes in, it checks it against rules taken from the specification. Each failing rule gets its own thread that highlights the elements, and ARVENT replies there once a later version fixes them.
The language model never does the maths and never writes the database query. It only fills in a structured query plan that's limited to real field names. Code checks and fixes that plan, runs it, and shows the exact query and result in every reply. The model can suggest rules from the specification, but nothing runs until a person approves it.
Along the way I also built a custom connector between Speckle and Revit for another client.
Features
- Answers questions about a selected element from its properties and any attached specification PDF, with page references
- Answers questions about the whole model, such as counts, totals and breakdowns per floor
- Keeps every element of the latest version of each model in a local SQLite database
- The language model only fills in a structured query plan; code checks and corrects it and runs it as a safe database query
- Each reply shows the exact query and result, plus a link that opens exactly those elements in the Speckle viewer
- Suggests rules from specification PDFs; a person approves them before they run
- Checks every new model version and opens one pinned thread per failing rule, highlighting the elements in the viewer
- Replies "Resolved" when a later version fixes the problem
- Missing values are reported as missing, never invented, and page references are checked against the documents
- Runs on a local Llama model through Ollama, so model data and documents never leave the machine
- Works with both Speckle data formats (version 2 Revit parameters and version 3 properties)
- Webhook server with signature checks and protection against duplicate retries; a command line for indexing, questions and checks; Docker setup with GPU support
Terms on this page
- Revit
- Autodesk's main software for designing buildings as BIM models. Most architects and engineers use it.
- Speckle
- An open platform for sharing and viewing 3D building models online.
- Language model
- An AI model that reads and writes text, like the one behind ChatGPT. Llama is a free one that can run on your own computer.
See it
Screenshots & demos
