studio mava

Happy week, everyone! We’re excited to announce the beta launch of our latest development – flinq!, a startup@us project. Over the past few months, we have worked hard on our web-app, which now supports the complex review of construction planning through intuitive chat input, powered by AI-driven queries. flinq serves at the critical junction of planning and regulation, aiding all professionals in the construction industry to ensure they meet the necessary standards, laws, and guidelines. We’re thrilled—and a bit astounded—that as architects, we've successfully brought this project to life. Currently focused on Germany, we have numerous innovative ideas in the pipeline and, admittedly, only so much time! 😊 While we love coding and evolving our web-app, we are now on the lookout for a CTO. If you share our vision and have a technical background, we’d love to connect with you! For more information, or to join us on this exciting journey, visit our website: https://flinq.ai

Imagine: a future where you can simply ask your building model whether it complies with state laws. Believe it or not, we're closer to making this a reality than you might think. We want to share a little snipped of our proof of concept testing out the Retrieval-Augmented Generation (RAG) with the langchain-neo4j integration. As an input we created a pseudo building representation of for example an Ifc-file in neo4j. The multi-housing building has two storeys and 5 flats with different amount of rooms. We left out some kitchens in the flats "04" and "05" on purpose in order to test the capability of the RAG. Then we embedded the State building regulations for Baden-Württemberg in Germany (LBOBW) as a text and as a vector representation of the text into the neo4j database and tested several questions using gpt-4 and gpt-3.5-turbo as a LLM (Large Language Model). We are impressed with the results! All our questions where answered correctly by extracting the relevant information from the knowledge graph and condensing it into an answer based on the provided document (law) and the building structure. This is the specific article of the the State building regulations for Baden-Württemberg (LBOBW) that the RAG was retrieving over the cypher: "§ 35 Flats (1) In buildings with more than two flats, the flats on one floor must be barrier-free accessible; this obligation can also be fulfilled by barrier-free accessible flats on several floors if the total floor area of these flats is not less than the floor area of the utilisation units on the ground floor. In these flats, the living rooms and bedrooms, a toilet, a bathroom and the kitchen or kitchenette must be barrier-free and accessible by wheelchair. [...] (2) Each flat must have a kitchen or kitchenette. Kitchens or kitchenettes without windows are permitted if they can be ventilated on their own. (3) Each flat must have its own water meter. This does not apply to changes of use if the requirement in sentence 1 can only be met with disproportionate effort. (4) In buildings with more than two flats, areas for parking prams and walking aids must be available for communal use that are as accessible as possible at ground level or easily accessible by ramps or lifts. (5) A storage room must be available for each flat" Original in German (https://lnkd.in/e8QDpwad) #KnowledgeGraph #RAG #AI #LangChain #neo4j #LLM

Who owns the database? A question we have been discussing a lot lately. With the current attention that AI is getting, we have been confronted several times with questions around databases. After all, it is no secret that big companies have access to a lot of information and have the resources to develop powerful tools around AI. But what about transparency? What if we want to have the control over the inputs? And what if we want to experiment with different models ourselves? At the end of the day, it comes down to who owns the database. We see this as an opportunity to level the playing field by creating our own qualitative datasets. We have started to systematically collect our own datasets from the internet, respecting copyright of course, or by creating our own content. In the long term, we want to make these available as open source datasets for everyone. If you would like to contribute, please contact us!