Think Twice

Inspiration

We've all been there. It's midnight, you're on your phone, and you're about to buy something you don't need. There's nobody to stop you. Your budgeting app quietly logs it after the fact. We are the most financially self-aware generation ever. We know we overspend. The problem is that emotion always beats logic at checkout, in the moment, when it matters most. We wanted to build the friend that actually tells you no. Not a notification. Not a spending tracker. A real-time debate that happens before you spend, one that knows your balance, your history, and your habits.

What it does

Think Twice is a Chrome extension and web app that intercepts your purchase decisions on Amazon before you check out. When you click "Ask The Board," four AI agents with distinct financial philosophies debate whether you should buy the item in real time. After multiple rounds of debate, the board delivers a verdict: Buy, Hold, or Pass, with full reasoning attached. A built-in AI chatbot lets you ask follow-up questions about any decision.

How we built it

Frontend: Next.js + TypeScript + Tailwind CSS
Backend: Python + FastAPI
AI Layer: Claude claude-sonnet-4-6 with multi-agent orchestration — each agent runs as an independent reasoning instance with shared purchase context
Database: MongoDB Atlas for storing user profiles, debate history, and spending patterns
Banking Data: Capital One Nessie API for real account balance and transaction context
Chrome Extension: Manifest V3, injecting into Amazon product pages to detect item name and price

Challenges we ran into

Getting four agents to debate each other coherently was harder than expected. Each agent needed to be aware of what the others had argued in prior rounds without losing its own distinct voice or reasoning style. Too much context and the responses became generic; too little and the debate felt disconnected. Designing a robust NoSQL schema in MongoDB Atlas was another significant challenge. Debate state, agent memory, user financial profiles, and transaction history all had to be modeled without relational joins, requiring careful thought about document structure and embedding strategy. We also hit Anthropic API rate limits during heavy multi-agent calls, which pushed us to optimize how we batched and sequenced agent turns.

Accomplishments that we're proud of

We built a working multi-agent debate system where four AI personas genuinely disagree with each other and update their arguments across rounds — and it runs end-to-end from a Chrome extension click on an Amazon page to a final verdict in under 10 seconds.

We're also proud of how cohesive the full product feels: extension, web dashboard, debate UI, and chatbot all work together as a single experience rather than disconnected pieces.

What we learned

We learned how agentic AI development actually works in practice, not just calling an API, but orchestrating multiple reasoning agents with shared context, round-based memory, and distinct personalities. We also learned how impactful a spending approval system grounded in real reasoning can be compared to passive budgeting tools.

On the technical side, we deepened our understanding of NoSQL schema design under real constraints and got hands-on experience integrating live banking data through the Nessie API.

What's next for Think Twice

Auto-scanning the open browser tab so no manual product input is needed: the extension reads product info directly from the page and submits it automatically A deal-finder agent that surfaces cheaper alternatives, discounts, and coupon codes whenever you search for something A deeper financial profile per user so agents can use your history, values, and goals to deliver more personalized verdicts over time