Guardiangel

Inspiration

Most financial apps are reactive — they show you what you already did wrong. We wanted something that pushes back before you act: a guardian that knows you, not just your balance.

What We Built

Guardian Angel learns your goals through a short onboarding, then watches every payment, receipt, and question through two LLM agents that share state.

The advice agent answers with full context — profile, goals, recent transactions, prior decisions. Not generic advice, but grounded reasoning. In a dry-run with a freelance graphic designer carrying €1,200 in credit card debt, it returned:

"€500 camera: legitimate business investment, but fund it via an Equipment savings pot at €100–150 per invoice cycle."

The guardian agent runs silently after every event, deciding whether to warn and — crucially — what to learn. It writes back to the user profile: new pain points, behavioral patterns, spending flags. The profile genuinely evolves with use.

A Duolingo-style streak keeps users honest: stay true to your goals and the streak grows; slip up and it resets. Gamification turns abstract financial discipline into something you actually feel.

Challenges

Keeping two agents coherent without a database. We solved it with atomic profile patches — each guardian call returns a diff the advice agent reads on the next turn, so context accumulates without conflicts.

What We Learned

Structured outputs are essential for agentic loops. Once both agents spoke the same JSON schema, wiring them together was straightforward — the hard part was designing that schema to carry enough signal without bloating every prompt.