Maali Mentor

Maali Mentor is a gamified financial platform for Pakistan. It uses a 10-level path and a Gemini-powered Urdu AI Voice Coach to teach budgeting, Islamic finance, and tax filing with a wealth simulator

Comment 1

🏆 Maali Mentor (مالی مینٹر) — Your AI-Powered Roman Urdu Financial Guide

"Aapka Apna Maali Masheer" — Democratizing financial literacy for 240 million Pakistanis through voice-first AI, interactive learning maps, and real-time simulators.

💡 Inspiration

In Pakistan, formal financial literacy is critically low, hovering around 18%. While millions of citizens actively use digital wallets like Easypaisa, JazzCash, and Nayapay, a vast majority do not understand the silent erosion of their money due to inflation, the difference between simple saving and active investing, or the benefits of becoming a tax filer.

The primary barrier isn't a lack of interest—it's accessibility. Financial concepts are heavily gatekept behind dense English jargon or dry, text-heavy curricula that do not resonate.

We built Maali Mentor (مالی مینٹر) to shatter this barrier. We wanted to create an experience that felt like chatting with a knowledgeable friend over a cup of chai. By building a voice-first AI tutor that speaks and understands Roman Urdu (Urdu words written in English script) alongside an interactive financial life simulator, we are putting personalized, jargon-free financial education directly into everyone's pocket.

🚀 What It Does

Maali Mentor is an immersive personal finance companion that guides users from absolute financial novices to advanced planners through an adaptive, game-like experience:

  1. Gamified 10-Level Ascension Roadmap: Users explore a visually rich learning map, progressing from Level 1: Budgeting (the 50-30-20 rule) to Level 10: Tax Planning & Filing in Pakistan.
  2. Interactive Level Quizzes: At the end of each level, users test their knowledge with 20 comprehensive Roman Urdu MCQs. Passing grading ($score \ge 15$) unlocks the next rank and awards XP points.
  3. Interactive Investment & Inflation Simulator: A 10-turn financial life game where users balance their lifestyle spending and allocate savings across Cash, Bank Savings (8% yield), Mutual Funds (16% yield), or Shariah-compliant Islamic Funds (14% yield) while battling random life events (medical bills, weddings, salary raises) and a dynamic inflation rate.
  4. Roman Urdu Voice AI Tutor: A real-time voice chat companion. Users can talk or type in Roman Urdu, receive instant context-aware voice responses, and use intuitive floating controls to play, pause, or stop the audio stream.

🛠️ How We Built It

Maali Mentor is designed as a modular, decoupled full-stack application built using:

  • Frontend: Next.js 16 (using React hooks and Tailwind CSS) styled with a high-fidelity glassmorphic dark theme, Nastaliq typography rendering, smooth micro-animations, and dynamic canvas graphs.
  • Backend: FastAPI with SQLModel (SQLite) for high-throughput, structured storage of user profiles, learning progress, XP metrics, and simulator states.
  • AI Brain: Gemini 3.1 Flash Lite API. We selected this model for its sub-second latency, robust multilingual understanding, and cost-effective token parameters, making it ideal for real-time mobile chat.
  • Voice Pipeline:
    • Speech-to-Text (STT): Gemini's native multimodal audio framework directly processes raw recorded user voice messages (audio/webm).
    • Text-to-Speech (TTS): To maintain a native, warm Urdu accent rather than a robotic English voice pronouncing Urdu words, the LLM translates its Roman Urdu response to native Urdu script behind the scenes. This script is synthesized via Google Text-To-Speech (gTTS) in Urdu (lang="ur") and played back to the user at a natural cadence ($1.2\times$ human speech speed).

⚔️ Challenges We Faced & How We Overcame Them

1. The Accent Gap in Roman Urdu TTS

Using typical English speech engines to read Roman Urdu text (e.g., "mehangai" or "bachat") sounds robotic and incomprehensible. However, Urdu speech engines cannot read English letters.

  • The Solution: We built a fast, internal translation layer. When the AI tutor generates a response, it returns clean Roman Urdu text for the UI, but asynchronously triggers a Gemini task to output the equivalent text in native Nastaliq Urdu script. The backend feeds this Urdu script to gTTS, which pronounces it with a perfect native accent.

🏆 Accomplishments We're Proud Of

  • Sub-Millisecond Caching Engine: Built a smart cache mapper for common beginner questions (e.g., "mutual funds kiya hai", "saving vs investing"), bypassing LLM calls entirely to deliver instant, pre-loaded audio guides.
  • Dual-Mode Quiz Engine: Developed a robust grading system that automatically switches to a lightweight frontend local engine (/quiz_data.json and localStorage tracking) if the backend API is unreachable, ensuring zero learning interruptions.
  • Clean & Secure Input Sanitization: Implemented strict Pydantic regex sanitizers on usernames (^[a-zA-Z0-9_\s\-]+$), quiz options, and file upload size caps (10MB max limits on /voice endpoint) to guard against injection or DDoS.

🧠 What We Learned

  • Adaptive TTS Formatting: Synthesized audio engines will literally read out Markdown symbols like asterisks (*) or hash signs (#). We learned to create custom parser regexes to sanitize text streams before feeding them to speech synthesizers.
  • Multimodal Context Window Management: Running STT and conversation inside the same LLM window requires careful system prompting to ensure the model distinguishes between audio transcription instructions and chat history.
  • Shariah-compliant Retail Structures: Structuring Level 7 required deep research into Islamic banking contracts such as Mudarabah (profit-sharing) and Murabahah (cost-plus markup).

🔮 What's Next for Maali Mentor

  • WhatsApp Business Integration: Porting the voice agent directly to a WhatsApp chatbot, enabling users to send voice notes and receive audio messages on the platform they use daily.
  • SMS-based Offline Quizzes: Allowing users without internet connectivity to complete level quizzes and receive lessons via basic SMS strings.
  • Digital Wallet Integration: Partnering with local fintech applications to let users practice mock investments using their real-world wallets in a sandboxed learning environment.
Share this project:

Updates

Submission history