Fulton Ring

Honored to be featured in Fast Company for our work rebuilding FEMA's Climate-Informed Future Risk Index! A huge thank you to Kristin Toussaint spotlighting our work and the mission to preserving these datasets. We plan on continuing to maintain this tool and will ensure that all of this taxpayer funded data remains publicly accessible, free of charge. Check out the full article here: https://lnkd.in/etJufERS

We’re excited to announce that Fulton Ring has created a restoration of FEMA’s “Future Risk Index” tool, which was taken down last week. You can access our version here: https://lnkd.in/ebvyQFxT The Future Risk Index was originally launched in December of 2024. It shows climate informed economic loss at the county level for different emission scenarios and natural disasters. In recent weeks, many tools similar to this have been taken down by the Federal Government, leaving many researchers without access to state sponsored data. If you’re interested in preserving government datasets, we recommend: ?? Downloading and screenshotting any tools you rely on that haven’t been taken down yet ??Forking replicas like ours: https://lnkd.in/eM8X_E67 ???? Developing or getting in touch with developers like ourselves who can help recreate tools that have been removed or are in danger of being removed very soon

Last week, we participated in the 1st annual Contech Alliance Hackathon. This was a great opportunity for us to demo our skills in building gen AI applications. We tackled Gryps' challenge and built a chat application to answer construction and civil engineering questions from structured and unstructured data. Here’s some key takeaways we’ve found true in this project and others we’ve done before: ?? PDFs are not a solved problem: Our unstructured data came in the form of Certificates of Occupancy from the NYC Department of Buildings. Think blurry, scanned PDFs from the 60s and 70s with names like ‘M00009315.pdf’. Traditional OCR techniques, even advanced services like unstructured.io and Reducto tend to struggle with extracting context from these documents. We were able to work around this and extract structured information from these documents using a vision model, in this case #GPT-4o from OpenAI. ?? Vector Search alone is not enough: Because we were working with a mixture of structured and unstructured data, it made sense to build an agent routing architecture to select which datasets to query. We used #LangGraph from LangChain to build this, which made it easy for us to add a feature where we could ask the user for additional clarification on questions that the bot struggled with. ?? Agentic search tools need fine tuning to be especially useful: While we utilized GPT-4V to build queries against our datasets, its lack of construction-specific context led to challenges. Questions like "What are the applicable fire codes?" or "What is the rough building footprint?" often resulted in misdirected searches. Moreover, the limited training data on architectural drawings means base models struggle to interpret drawings and their relationships within documents. ??? The Path Forward: We've found that many unstructured data querying challenges can be addressed through a combination of structured databases and improved data quality. While tools like Vanna AI are useful in situations where your data is entered perfectly (like DoB's complaints database), they struggle with fuzzy information like what we extracted from COs. For instance, handling an address ('123 West 57th street' vs '123 W. 57th') or inconsistent column names ('post_code' vs 'ZipCode') can result in incorrect data being returned. At Fulton Ring, developing solutions for these issues is a key focus on our 2025 roadmap. Overall, there is still a lot of progress left to be made when it comes to AI search and tooling. To see how we can build similar solutions for your business, definitely get in touch with us here: https://lnkd.in/gGdpzqNH Link to full deck: https://lnkd.in/gpH8EvaW