Immigrant First

Immigrant searching for home
gpt4
claude2
value

Inspiration

Immigrant First is built by immigrants who faced challenges navigating the complexity and challenges of the immigration process. It's driven by a desire to use safe Generative AI to streamline immigration petition generation for immigrants.

What it does

Our current product generates a complete EB1A immigration petition using GPT 4 and Claude 2 with the following:

Cover Letter
Table of Contents
Complete Petition
Final Merits Determination

How we built it

The product works in the following ways:

Collect the resume in the form of an input file from the immigrant
Store the resume information in the form of a vector database
Collect information from Perplexity to understand the importance of petitioners' endeavors
Use GPT 4 and Claude 2 to generate the immigration petition, cover letter, table of contents, and final merits determination

Challenges we ran into

Business Challenges:

We took a while to figure out how to measure the value generated by Claude 2 in comparison to GPT-4

Technical Challenges:

prompt engineering: generating a long and formal petition letter and ensuring in-text citation and footnotes are included in the output
evaluating the output from GPT4 and Claude2

Accomplishments that we're proud of

Able to use Perplexity to browse the web and obtain internet sources for adding credibility to the petition
Able to get a working demo that compares the current two most popular LLM models: GPT4 and Claude2.

What we learned

Improvements in prompt for Claude2 improve the performance of LLM output significantly
Though vector databases can solve some of the context size challenges, larger context size helps in creating more coherent stories
As we increased the quantity of content generated, the quality of content got impacted. A combination of multiple LLMs provides high-quality content output