AutoPilot for your Wallet

Managing finances saving, investing, debt and taxes can be overwhelming and advice is often biased. We built an LLM that taps your data via APIs to offer unbiased personal guidance at near zero cost.

Comment

Inspiration

I realized that most people struggle with basic financial decisions—how much to save, where to invest, and how to manage debt and taxes—yet traditional advice is often biased or prohibitively expensive. Seeing friends hop between Google searches, well-meaning forums, and pricey advisors inspired me to build a more accessible solution.

What I Learned

  • LLM Alignment: How to fine-tune large language models so their recommendations stay accurate and unbiased.
  • API Integration: Best practices for securely connecting banking and investment APIs to fetch real-time data.
  • Privacy & Compliance: The importance of handling OAuth flows, encryption, and user consent in financial applications.
  • UX for Trust: Designing prompts and interface flows that build confidence in automated advice.

How I Built It

  1. Data Layer:
    • Integrated Plaid and Yodlee APIs for transaction history, account balances, and investment portfolios.
    • Stored metadata in an encrypted PostgreSQL instance.
  2. Modeling Layer:
    • Fine-tuned an open-source LLM using domain-specific financial corpora.
    • Employed LangChain to orchestrate retrieval-augmented generation (RAG) for up-to-date figures.
  3. API Server:
    • Developed a FastAPI backend with OAuth2 authentication, rate limiting, and caching.
    • Exposed endpoints for “getRecommendation,” “projectSavings,” and “simulateInvestment.”
  4. Frontend Prototype:
    • Built a React dashboard that visualizes your net worth trajectory and recommended actions.

Challenges Faced

  • Data Quality: Inconsistent transaction categorization forced me to implement custom normalization rules.
  • Model Hallucinations: Initial prompts produced confident but incorrect advice—iterative prompt engineering and feedback loops were key.
  • Performance: Balancing real-time API calls with model latency required aggressive caching and async processing.
  • User Trust: Ensuring transparency (e.g., “Here’s why I recommend this…”) to prevent users from blindly following automated suggestions.

––
This journey taught me not only the technical stack behind a personalized finance assistant but also the human factors—trust, clarity, and privacy—that make it truly useful.

Built With

Share this project:

Updates