Benefits Navigator

The first Agentic Financial Strategist powered by Gemini 3 Flash. Maximize rewards, discover hidden perks, and automate your financial goals.

Comment

💡 Inspiration

We all love credit card points, but maximizing them is a full-time job. Billions of dollars in rewards are left on the table every year because the "rules" of credit cards—multipliers, rotating categories, insurance policies, and credits—are too complex for any human to track manually. We realized that while humans struggle with this complexity, it’s the perfect problem for a high-reasoning AI model. We built Benefits Navigator to turn confusing financial fine print into a simple, automated strategy that lives in your pocket.

📱 What it does

Benefits Navigator is an AI-powered iOS application that serves as your personal financial agent.

  • Smart Wallet: Instantly adds cards by name and pulls their official benefit data—uncovering both popular rewards and hidden perks—using Gemini grounded with Google Search.
  • Sign-up Bonus Tracker: Automatically tracks your spending progress toward massive sign-up bonuses (e.g. "Spend $4,000 to earn $750") by analyzing your transactions. It provides a clear progress dashboard, ensuring you never leave free money on the table.
  • Context-Aware Recommendations: At checkout? Tell the app where you are (e.g., "Starbucks" or "Japan Airlines"). It instantly recommends the single best card in your wallet—optimizing for maximum rewards or specific protections like Extended Warranty, Flight Insurance, or Car Rental Coverage—while factoring in your personal financial situation and goals.
  • Agentic Roadmaps: Long-term planning. Tell the Agent "I want a free flight to Tokyo," and it builds a step-by-step roadmap (e.g., "Open Card X," "Spend $3k in 3 months")—tailored to your overall financial profile. It tracks your progress and automatically adapts the strategy in real-time based on your spending and feedback.
  • Action Center: A comprehensive hub for managing your card benefits. Users can create records for specific perks—like Car Rental Insurance, Airport Benefits, Extended Warranty, and Price Protection—and get detailed, grounded instructions from Gemini on exactly how to use them based on their card's fine print. It also automatically monitors your recent purchases for price drops, alerting you if an item you bought is now cheaper so you can claim the difference.

🛠️ How we built it

We built a Client-Server Architecture with an Agentic Loop.

  • Backend: Python & FastAPI running on Google Cloud Run.
  • Database: Firebase Firestore for real-time user state and authentication.
  • AI: Google Gemini 3 Flash. We chose the Flash model specifically for its ultra-low latency, which is critical for a mobile app experience where users expect instant answers at the register.
  • Frontend: Native iOS app built with SwiftUI for a premium, responsive feel.
  • Grounding: We utilize Gemini's built-in Google Search tool to verify credit card offers and retrieve real-time data on retailers.

System Architecture

Benefits Navigator Architecture

🧠 Gemini Integration

Benefits Navigator is built entirely around the reasoning capabilities of Gemini 3 Flash. While features like the Smart Wallet use Gemini for instant, one-off analysis, our flagship "Marathon Agent" introduces a persistent, autonomous "Agentic Loop" that manages your long-term financial strategy.

  1. Marathon Agent (Agentic Loop):
    • Architecture: The agent runs on a "Wake Up → Think → Act → Sleep" cycle. It doesn't just respond to a prompt; it maintains a persistent life cycle.
    • State Management: We use Firebase Firestore to store the agent's memory in two distinct states:
      • Private State ("The Brain"): Stores the "Thought Signature"—a rolling summary of the agent's past reasoning and logic. This is fed to the agent every time it updates the roadmap, allowing it to remember context from months ago without rereading every transaction.
      • Public State ("The UI"): Stores the structured Roadmap and Side Quests that the iOS app listens to in real-time.
    • Long-Context Reasoning: By feeding your entire financial profile and spending habits into Gemini 3 Flash, the agent generates and updates a multi-step Roadmap (e.g., "Open Card X", then "Spend $3k") that adapts automatically as you spend.
  2. Real-Time Reasoning:
    • For immediate decisions (like "Which card of mine do I use at Starbucks?"), Gemini compares complex variables—multipliers, rotating categories, and insurance perks—against your current cards, financial situation, benefit priority, and goals in seconds.
  3. Grounding:
    • We use Gemini's Google Search Tool to fetch the latest "Guide to Benefits" PDFs and retailer MCC codes in real-time, ensuring our data never goes stale.
  4. Self-Updating Knowledge Base:
    • To keep benefits accurate without redundant API calls, we built a Global Cards Architecture. A scheduled agentic job autonomously browses the web once a month to verify the latest "Terms and Conditions" for every supported card. If Gemini detects a change (e.g., a reward rate shift), it updates the Global Collection, instantly propagating the new correct data to every user's wallet.

Gemini 3 Flash's speed allows us to run these complex reasoning chains in seconds, making the "Agent" feel alive and responsive.

🚧 Challenges we ran into

  • Hallucinations vs. Accuracy: Financial advice needs to be precise. Early iterations would sometimes invent card benefits. We solved this by implementing strict "Grounding" prompts that force Gemini to cite official sources via Google Search before making a claim.
  • Latency: Creating a full financial roadmap takes time. We implemented a "Thinking..." state in the UI and used background workers (FastAPI BackgroundTasks) combined with Firestore listeners to push updates to the phone asynchronously, keeping the app snappy.

🏆 Accomplishments that we're proud of

  • Building a fully functional Agentic Loop for the goal roadmap that autonomously updates itself based on user feedback.
  • Achieving a seamless integration between a Python backend and a native Swift UI.
  • Building a Self-Healing Data Layer where autonomous agents actively browse the web to keep credit card benefit terms up-to-date without human intervention.

📚 What we learned

We learned that Context is King. The difference between a generic recommendation and a truly helpful one lies in feeding the model the right details—the user's specific goals, their current progress, their financial situation, benefit priority, and the exact merchant details. Gemini 3 proved exceptionally good at handling this large context window without losing focus on the primary task.

🔮 What's next for Benefits Navigator

  • Receipt Scanning: Using Gemini's multimodal capabilities to scan physical receipts for missed points.
  • Automated Redemption: Having the Agent not just tell you how to use points, but actually finding award flight availability for you.
  • Android Support: Bringing the experience to the Google Pixel ecosystem.

Built With

Share this project:

Updates