PermitHero

Upload documents, distinguished by whether they are safe or contain sensitive information, and only metadata is gleaned from those.
Provide suggestions to improve phrasing and diction to better fit the needs of the application
Fill out information that is inconsistent between forms and needs refilling.

Inspiration

It is very difficult to apply for permits correctly, and often small issues with forms can have them delayed for months. There are also clues that aren't explicit in the way language can influence people who allow approval of documents. We realized the vague, non-traditional overlap of requirements and concerns that permeate between different departments in the city government can delay the process of permit application and slow down decision making, as well as present challenges to first generation immigrant business owners who have issues with writing in English.

What it does

Allows users to upload form application, identify discrepancies between individual forms and old permits, and provides contextual suggestions, aided by LLM integration that uses a vector database to identify common features among permit language that maximizes both the probability of approval and speed of approval.

The goal is to leverage existing forms, making sure data is secure (by distinguishing public and private documents), minimize the overhead of new forms filled out, automatically generate pdfs, and provide help with language and phrasing to maximize change of success and minimize difficulty and ambiguity for both those filling out forms and those tasked with reading hundreds per day.

How we built it

Written in React, using integrations with Chroma and OpenAI to aid in vector embeddings and semantic understanding of form details

Challenges we ran into

Connecting the vector database with true outcome based associations. Identifying whether or not something is relevant is not obvious by a large margin, but there are always contextual clues that have to be leveraged both on the data in the form and the relevant metadata of its use in the permit application funnel.

Accomplishments that we're proud of

Getting it to work for business permit applications, maximizing the degree of accuracy in ascertaining small differences in style based on the general intent of the application (via semantic analysis, different structures for embedding models)