data_enrichment.txt

Engineering AI agent: property data enrichment
The data that can be extracted from broker submissions has three major issues: (1) it is
unstandardized, (2) unvetted, and (3) sometimes incorrect. Thus, property data
enrichment is a critical value add that ResiQuant provides to its customers. To enrich the
property data, ResiQuant is developing a multimodal AI model that should be very
accurate at classification tasks that require structural engineering expertise, knowledge of
constructing practices, and building code regulations. The classifications should be
compatible with the inputs of catastrophe models (see "Required for CAT model" column
in this list here).
This multimodal AI model processes information about the specific building considering
engineering context. The building-specific data is scraped from the internet and/or
provided by the broker; this data usually includes street view images, satellite images,
interior images, building permits, real estate listings, tax data, and other documents. The
structural engineering context that the model shall account for includes:
1. Historical building codes and timeline of applicability in each
jurisdiction/municipality in the US.
2. Documentation in the forms of papers and interviews with experienced structural
engineers and contractors of the construction practices in the areas we operate.
3. Manually labeled images of the building.
4. Building drawings and associated exterior images.
The short videos in this folder show a handful of cases where licensed engineers in
California manually review building data and use their domain expertise to perform a
series of classification tasks for each building.

Questionnaire for property data enrichment
5. Describe an AI pipeline that you will use to conduct the classification task for
determining the construction class of a building considering structural
engineering expertise. Feel free to respond with as much detail possible, include
diagrams, or any other means to support your answer. Be specific with the model
types and algorithms you will use. Document your design decisions including
comments about other alternatives that you considered and how you would
evaluate your design decisions during development.
6. Describe all the data preparation work necessary for the structural engineering
context data to use effectively in a multimodal AI model. What model
architecture would you use?
7. What data do you need for improving the accuracy of the classification
algorithm?
8. If you have another AI engineer supporting you in addition to building domain
experts, how would you approach the implementation with this pipeline
assuming you are the team leader?