AspAIra

Problem & Motivation

Millions of workers are hired by employment agencies to migrate to the Middle East and take on low-income service jobs. Due to lack of financial opportunity in their home countries, these workers migrate to the Middle East and often send large portions of their earnings back to their home country to provide financial support to their families. Unfortunately, they face significant vulnerabilities including debt traps, recruitment expenses, lack of financial literacy, and limited knowledge of safe banking options. Many domestic workers—particularly live-in maids—lack formal financial education, resulting in instability both for themselves and their dependents back home. Returning to their home country, as many aim today, requires building sustainable independence.

We strongly believe this is a problem with opportunity for widespread impact both across and beyond the Middle East. For our capstone project, we are focused on the population of ~750,000 domestic workers in the UAE, specifically those who migrated from the Philippines and are working as live-in maids.

Our Solution

Our platform vision is to provide personalized tools that build trust, foster growth, enable financial empowerment, and strengthen community connections.

• Digital Identity & Credit Score: Build trust with verified digital profiles to support safe banking and employment opportunities.
• Coaching & Education: Personalized financial literacy and savings guidance tailored to migrant workers' unique needs.
• Microfinance Matching: Connect workers with donors and microfinance opportunities to build stability and independence.
• Community Engagement: Foster mentorship, peer support, and connections through community-driven programs.

For our capstone project, we took a focused approach and developed this Coaching & Education element of the platform as an MVP.

Data Science Approach

FIELD RESEARCH

With the help of maids.cc employment agency, we surveyed 26 Filipino domestic workers in the UAE to collect field insights about financial behaviors and needs.

Summary of key user profile findings:

• 85% don’t have a bank account
• 88% have children back at home
• 70% are at a high school education level
• 22% have debt in their home country

Summary of key user financial needs:

• 70% are interested in budgeting and saving
• 37% are interested in how to open bank accounts
• Nearly all who replied to open ended feedback mentioned that providing for their children’s education as a significant challenge

AI AGENT DEVELOPMENT

Financial Coach AI Agent

Based on insights from our field research, we leveraged Dify (an open-source platform that enables fast development of LLM applications) to build an AI agent financial coach. The financial coach takes user profile inputs—both demographic information (home country, number of children, years in UAE, etc.) and financial behavior information (bank accounts, debt, remittance methods & amounts)—to personalize dynamic education content. We thoughtfully iterated on the agent’s prompt, instructing the LLM to walk users through a step by step education session using simple, accessible language and ending each learning session in a short quiz to monitor comprehension.

With Generative AI projects, a thoughtful evaluation framework is key to meeting users’ needs. Our team selected five key metrics for selecting the Agent LLM: personalization, language simplicity, response length, content relevance, and content difficulty. Reviewing industry standard for similar LLM applications, we performed conversational evaluation (using a defined rubric based on the defined metrics) of both GPT-4o and Claude 3.5 Sonnet in our Dify agent framework, ultimately selecting Claude 3.5 Sonnet to power the final financial coach AI agent.

Multi-Judge AI Evaluation Framework

Again leveraging Dify, the team built a Multi-Judge LLM Evaluation Agent for conversation evaluations

Based on these findings, we designed AspAIra, a mobile-first AI-driven financial literacy coach. Our system leverages user profile inputs, personalized prompting, and real-time conversations with a financial coach agent, enhanced by multi-judge LLM evaluation (Claude, GPT-4, Gemini) to ensure fairness and content quality.

DEPLOYMENT ARCHITECTURE

Aspaira is a full end to end deployed web application, deployed with the help of Amazon Cloudfront to be available globally for testing in the UAE. The technical stack for the web application includes Streamlit frontend, FastAPI backend, Dify orchestration, and AWS DynamoDB storage. Additionally, our custom Evaluation Service for running the Multi-Judge LLM Evaluation Agent is deployed via AWS Fargate.

Evaluation

INTERNAL EVALUATION

As a team, we performed 60+ conversations with the financial coach agent, taking on both “typical” and “edge case” personas defined by insights from our field research survey. These conversations were leveraged as input to the Multi-Judge Evaluation Framework.

Additionally, our DynamoDB database was designed to capture other key user interaction and platform metrics like latency.

EXTERNAL EVALUATION

We shared our web application link with maids in the UAE for real user feedback. Users completed their profiles in the app, education sessions with the financial coach, and filled out a brief survey.

Quotes from users: ”I love it 😍😍 Very simple and intelligent.”

“The bot’s customization and personalization options were tailored to my personal needs and were assistive.”

Key Learnings & Impact

• Advanced prompt engineering and structured user profiling improved educational outcomes for low-literacy users.
• Multi-judge evaluation methods increased fairness and robustness in AI model assessments.
• Field testing validated the demand for personalized, accessible financial education among migrant workers.

AspAIra demonstrates the potential of responsible AI to foster financial inclusion and sustainable impact for vulnerable global populations. Our platform is the first of its kind aimed toward financial empowerment for migrant workers in the UAE and we hope to see this type of work continue!

Acknowledgements

We thank maids.cc agency for initial feedback, the domestic workers who participated in our field research, Berkeley MIDS faculty and mentors, and Dify and AWS for enabling our rapid prototyping and deployment.