AI for Financial Literacy to Rural People

Inspiration

The gap in financial understanding in rural India is structural: limited access to reliable information, dependence on verbal communication, and inconsistent internet connectivity. Most financial education tools assume literacy, stable connectivity, and familiarity with digital interfaces. The mismatch between these assumptions and rural realities defined the problem and shaped the solution direction: a voice-first, multilingual, context-aware AI system that teaches fundamental financial concepts without relying on high-end devices or continuous internet.

What I Learned

1. User constraints dictate system architecture.
   Building for low literacy and low connectivity forces a shift toward offline-first design, simplified NLU, and localized examples.

2. Financial education requires progressive scaffolding.
   Concepts like interest, EMI, or insurance must be sequenced. The structure resembles a curriculum more than a chatbot.
   Example:
   Understanding of EMI → f(Simple Interest, Loan Terms)

3. Local-language ASR/NLU is non-trivial.
   Dialectal variations and code-mixing—e.g., Hindi–Punjabi and Hindi–Bhojpuri blends—create ambiguity that a generic model cannot handle.

4. Cultural grounding changes learning outcomes.
   A farming-based metaphor increases comprehension more than abstract definitions.

How I Built the Project

1. Requirements and User Modeling

Mapped target users: small farmers, shopkeepers, SHG members, daily-wage earners.
Constraints identified: low literacy, voice preference, intermittent connectivity, variable dialects.

2. Content Framework

Created a modular financial curriculum:

• Savings and budgeting
  
• Simple and compound interest
  
• Loans and credit
  
• Insurance
  
• Digital payments and safety
  
• Government schemes
  

Each module structured as:
Concept → Example → Activity → Verification question

3. Dialogue System

Implemented intent–response templates:

• ask_definition
  
• ask_example
  
• ask_steps
  
• ask_warning
  

Built a small knowledge graph:

[ Node: Concept → (Definition, Example, Risk, Next Step) ]

4. Local-Language Pipeline

Used lightweight ASR + TTS models optimized for Hindi and Punjabi.
Compression + quantization for offline deployment.

5. Interaction Design

Voice-first navigation.
Short, contextual explanations using scenarios familiar to rural users.
Example: interest explained through crop-cycle investment instead of abstract equations.

6. Offline Deployment

All modules packaged locally.
No dependency on continuous internet except for periodic content updates.

Challenges Faced

1. Dialect diversity
   Continuous code-mixing reduced ASR reliability. Required constrained language models and domain-specific vocabulary.

2. Designing non-literate UX
   Screen-based text flows failed. Voice-led flow with confirmation prompts became mandatory.

3. Balancing accuracy with simplicity
   Precise financial definitions conflict with cognitive load constraints. Required iterative simplification.

4. Offline-first engineering
   Keeping model size small without degrading quality was a persistent constraint.

5. Cultural adaptation
   Imported examples didn’t work. Every explanation needed localization to farming, livestock, SHGs, or local markets.

Summary

The project is a structured attempt to build an accessible, multilingual, offline-capable AI tutor that teaches rural users fundamental financial concepts through voice-based, culturally grounded interactions. The constraints of the target audience directly shaped the technical and instructional design approach.

Built With

• django
• flutter
• numpy
• pandas
• python
• react-native