Empowering Social Entities: Building an AI Grant Writing Assistant with Vertex AI & RAG

 Introduction

Small non-profits and social entities are the backbone of many communities, yet they often face a significant hurdle: securing funding through complex grant applications. Official call documents, like those from the EU's Daphne Programme, can be particularly demanding for resource-constrained organizations.

To help bridge this gap, I developed inSocialTrust (conceptualized at insocialtrust.com), an AI-powered assistant designed as an "on-demand grant writing advisor". This project, created for the Google & Kaggle Gen AI Intensive Course Capstone (find the public notebook [inSocialTrust: AI-Powered Grant Proposal Assistant]), leverages Retrieval-Augmented Generation (RAG) and Google Cloud's Vertex AI to empower these vital organizations. This post explores the challenge, the inSocialTrust solution, and the technology behind it.

The Challenge: Decoding Complex Grant Requirements

Applying for public grants often presents formidable barriers for smaller social organizations:

• Information Overload: Grant calls are frequently lengthy, complex documents filled with jargon and specific, non-negotiable criteria. Thorough analysis is time-consuming.
• Resource Constraints: These organizations rarely have dedicated grant writers or the spare staff capacity needed to meticulously craft tailored proposals.
• Compliance Risks: Misinterpreting or failing to address specific requirements can lead to immediate disqualification, wasting precious time and effort.

This bottleneck hinders many potentially impactful projects, representing a loss for both the organizations and the communities they benefit.

Our Solution: inSocialTrust - An AI Grant Advisor Built on RAG

inSocialTrust (insocialtrust.com) aims to level the playing field by providing an intelligent assistant grounded directly in the official grant documentation. It serves as a real-time guide, utilizing Retrieval-Augmented Generation (RAG) powered by Google's Vertex AI.

Here’s the core workflow:

1. Ingest & Understand: The system reads the official grant call PDF. For this project, we used the EU Daphne Programme - Line 7 document, originally published in Spanish. Vertex AI's multilingual text embeddings (text-multilingual-embedding-002) are used to understand and vectorize the content.
2. Index: These vector embeddings are stored in a searchable ChromaDB vector database, creating a knowledge base specific to the grant call.
3. Assist & Evaluate: When a user needs help with a specific proposal section or wants their draft text evaluated:
   • The system retrieves the most relevant passages from the indexed grant document based on the user's query or topic.
   • It feeds this retrieved context along with the user's draft text into Vertex AI's Gemini Pro LLM.
   • Guided by a carefully crafted prompt, the LLM generates constructive feedback in English, evaluating the user's text against the official criteria present in the retrieved context.

This ensures that the advice and evaluations provided by inSocialTrust are directly based on the source material, reducing compliance risks and the analytical burden on the organization.