Penny

Inspiration The average millionaire started investing at 14. The average American starts at 33. That's 19 years of compound interest — gone. Most kids learn about money too late, and by the time they get their first job, the habits are already set. Existing tools like Greenlight and BusyKid show dashboards, but nobody actually talks to the kid. We wanted to build an AI agent that doesn't just display numbers — it teaches, adapts, and acts like a real financial mentor that grows with your child.

What it does Penny is an AI-powered financial literacy agent for kids. Kids complete chores, parents approve them, and they earn virtual coins. Once they hit a $50 threshold, Penny — a video avatar powered by Tavus — appears and has a real conversation with the kid. She detects their interests, suggests 3 personalized investment options, and walks them through the decision. Parents get a phone call to approve or reject, and if they reject, a Negotiation Agent analyzes the data and turns it into a learning moment for the kid. A Risk Monitor watches their portfolio 24/7 and alerts both parent and child on major news. Every Sunday, parents receive a weekly report card via email.

How we built it We built a React + TypeScript frontend with Vite and Tailwind CSS, using Auth0 for parent/child role-based authentication. The backend is FastAPI (Python) with Ghost as the data layer for storing chores, portfolio data, and investment records. We integrated Tavus CVI for Penny's video avatar, Claude (Anthropic) as the AI brain powering all 5 agents, and Ghost's email capabilities for automated weekly parent report cards. The frontend and backend communicate through REST APIs with JWT-based auth.

Challenges we ran into Coordinating five autonomous agents that need to hand off context to each other — Penny detecting interests, the Portfolio Builder generating options, the Approval Agent calling parents, the Negotiation Agent analyzing rejections, and the Risk Monitor scanning news — was the hardest part. Making sure each agent had the right context at the right time without losing the conversational thread required careful prompt engineering and state management. Getting Auth0 RBAC to cleanly separate parent and child permissions while keeping the UX simple for kids was another challenge.

Accomplishments that we're proud of Penny is a self-improving system — after every conversation, she stores what topics were covered, whether the kid understood, which analogies worked, and the kid's attention span. The next conversation reads this full history and adapts. Penny never repeats a lesson a kid already mastered. We're also proud of the rejection flow: instead of just saying "no," the Negotiation Agent shows parents real data (historical performance, diversification scores, confidence percentages) before their final decision, and then Penny turns that rejection into a genuine teaching moment for the kid.

What we learned We learned that the most impactful AI agents aren't the ones that do the most — they're the ones that know when to act and when to hand off. Building a multi-agent system taught us that agent orchestration is harder than any single agent's logic. We also learned that designing for kids means every interaction has to be simple, friendly, and rewarding — complexity has to live entirely behind the scenes.

What's next for Penny Real brokerage integration so kids can make actual micro-investments with parental approval. Expanding Penny's teaching curriculum beyond stocks to savings goals, budgeting, and entrepreneurship. Adding multiplayer features where siblings or friends can compare portfolios and learn together. And building out the self-improving layer further so Penny can adapt not just to individual kids, but across cohorts — learning which teaching strategies work best for different age groups and interest profiles.

Built With

• auth0
• ghost
• javascript
• livekit
• python