Trinity

Your Triadic Financial Guardian

Comment

Inspiration

I'm a broke college student. Last week I had $25 to my name and rent was due. I was bleeding $75.98/month on subscriptions I forgot about. Hulu was only used twice a month, Adobe's "free trial" quietly charging $69.99/month for three months. The money I needed was already there, just hidden in my inbox. I needed AI agents that actually work while I sleep, study, and party. By scanning my email 24/7, the agents find every forgotten subscription and hold me accountable to review them before I waste another dollar. With Fetch.ai's Agentverse, I built three coordinated agents that give me complete financial visibility and make sure I never forget what I'm paying for.

What it does

Trinity scans your Gmail inbox and shows you every subscription you're paying for.

  • Scout Agent: Scans thousands of emails, finds every subscription receipt, renewal notice, and billing confirmation.
  • Strategist Agent: Analyzes for duplicates (Spotify + YouTube Music), forgotten trials that became paid, price increases, and calculates your total monthly burn.
  • Executor Agent: Sends check-in reminders at intervals YOU choose (1/3/6/12 months), shows exactly how to cancel, tracks money saved.

How we built it

Backend: Three-Agent Architecture on Fetch.ai Agentverse. We built three specialized AI agents, each using LLMs for intelligent decision-making:

🔍 Scout Agent - Scans email JSON data from Gmail API, uses LLM to categorize emails (active subscriptions, free trials, promos, price increases), extracts key data (price, date, sender), sends to Strategist

🧠 Strategist Agent - Analyzes all subscriptions using LLM to find overlaps, downgrade opportunities, available discounts, and free trials about to charge. Prioritizes by savings potential and sends recommendations to Executor

⚡ Executor Agent - Generates check-in notifications, uses LLM to draft cancellation emails when users decide to cancel, coordinates with Gmail API to send emails (with user approval)

Frontend: React Native + Expo

  • Tinder-style swipe interface using react-native-deck-swiper (swipe right = keep, left = cancel, down = snooze)
  • Gmail OAuth for secure read-only email access
  • Real-time agent activity dashboard
  • WebSocket for agent-to-frontend communication

Tech Stack: Fetch.ai uAgents SDK (Python), Gmail API, React Native, Expo, WebSocket

Challenges we ran into

  • LLMs in multi-agent workflows: Simple LLM API calls are easy. Embedding LLMs within agent logic with conditional workflows and state management is entirely different. Each agent's LLM output needs to feed into the next agent's logic perfectly—any formatting inconsistency breaks the chain.

  • Prompt engineering = code - In multi-agent systems, prompts ARE the implementation. Vague prompts fail. We needed strict, direct prompts with rigid formatting ("Extract price as XX.XX, date as YYYY-MM-DD") so outputs could be reliably consumed by the next agent. Prompt quality directly determines if agents can coordinate.

  • Orchestrating three volatile pieces: LLMs are probabilistic, so coordinating three of them felt like conducting an orchestra where everyone improvises. We needed fallback logic when Scout misses data, retry mechanisms when Strategist gets confused, and validation before Executor sends emails.

  • 24-hour learning curve: None of us had built multi-agent LLM systems before. Learning Fetch.ai's uAgents SDK, agent communication protocols, LLM orchestration, and React Native simultaneously meant constant debugging and simplifying until we had something that worked end-to-end.

Accomplishments that we're proud of

  • Technicalities: Built a polished React Native app and developed multiple AI agents by developers who'd never touched it before this weekend.
  • Real multi-agent coordination running on Fetch.ai Agentverse, not mock-ups: Three agents actually communicating via message passing, making autonomous decisions, and working 24/7.
  • Pioneering collaborative AI architecture: Instead of one monolithic bot, we built agents that specialize and work as a team. Scout finds, Strategist analyzes, Executor coordinates. Our approach is how AI should work, mimicking human collaboration, not replacing it with a single oracle.
  • Solving prevention, not just detection: Most apps show you what you spent. We prevent future waste through scheduled accountability.

What we learned

  • Multi-agent systems shine with specialization: Giving each agent a distinct role (Scout finds, Strategist analyzes, Executor coordinates) made coordination purposeful and reduced complexity.

  • Prompts are code: In multi-agent systems, prompt quality directly determines success. Strict, well-formatted prompts with clear instructions are what make LLMs reliable in production, not just demos.

  • LLM integration requires architecture: LLMs are probabilistic, but agent workflows need deterministic outputs. We learned to build validation layers, fallback logic, and structured formats so one agent's output reliably feeds the next.

  • Users want control, not full automation: People don't want AI logging into accounts and canceling things. Our check-in system works because it provides awareness and accountability while users stay in control.

  • The problem is bigger than expected: We thought college students waste $200-300/year on subscriptions. Reality: $500-800+. Even financially aware people have forgotten subscriptions. The market validated itself.

  • Fetch.ai removes infrastructure complexity: We didn't build distributed systems or manage servers from scratch. Fetch.ai gave us agent orchestration so we could focus on logic and UX—critical for a 24-hour sprint.

What's next for Trinity

In the long run, we hope to expand beyond college students and explore features like bill negotiation, cashback tracking, and investment insights. Our vision is to grow into a more comprehensive financial assistant that helps users save money while reducing the mental overhead of managing finances. We're starting with subscriptions because it's the problem we understand best as broke college students. If we can nail this one feature really well, the multi-agent architecture we've built could potentially scale to other areas of personal finance down the road.

Built With

Share this project:

Updates