confIA

Inspiration

Personal finance apps assume people have time and energy to manually log every purchase. In Peru, that assumption breaks down fast: workers lose significant hours a year stuck in traffic, and small, unplanned "micro-spending" — a coffee here, a snack there — quietly eats up a meaningful share of monthly income. It's no surprise that most people abandon budgeting apps within days of downloading them.

We built confIA around one person who represents this reality: María, a tired professional who wants to manage her money better but simply doesn't have the bandwidth to type every receipt into a form at the end of an exhausting day. The problem was never a lack of financial knowledge — it was friction.

What it does

confIA removes the manual work from expense tracking. A user takes a photo of a receipt — no forms, no manual entry. GPT-5.6's vision capabilities read the receipt and extract the date, amount, and merchant, and automatically infer the spending category. The structured data is shown back to the user for a one-tap confirmation, instantly updating their daily spending limit.

Every logged action — capturing a receipt, staying under budget, spending less than usual — also contributes to a personal Trust Score: not a credit bureau or a score shared with banks, but the user's own financial health indicator, designed to build a positive habit loop rather than external judgment.

How we built it

Codex was used throughout the hackathon to build and iterate on the receipt-parsing pipeline, backend logic, and the Trust Score calculation — accelerating development and letting us test multiple approaches quickly. GPT-5.6's vision API powers the core runtime functionality: reading and structuring receipt data in real time. [PLACEHOLDER — confirmar stack exacto: framework de backend, frontend, base de datos. Ejemplo tentativo: Python backend, PostgreSQL for transaction and score data, [frontend framework] for the chat-based UI.]

Challenges we ran into

Getting reliable, structured extraction from real-world receipts — crumpled paper, faded thermal print, and inconsistent formats — required careful prompt design around GPT-5.6's vision capabilities. Scoping the project tightly enough to ship a working, coherent flow within the hackathon timeframe, instead of spreading effort across too many features. Designing the Trust Score responsibly: we deliberately avoided framing it as an alternative credit bureau or a score shared with financial institutions, to keep the concept honest, user-owned, and compliant with data protection principles.

Accomplishments that we're proud of

A working end-to-end flow — photo → AI extraction → confirmation → score update — built and demoed with real screen recordings, not mockups. Grounding the product in real behavioral data (app abandonment rates, time poverty, micro-spending) rather than assumptions. A Trust Score design that motivates the user directly, without overselling a financial product we couldn't build or validate in a hackathon timeframe.

What we learned

Practical lessons on prompting GPT-5.6's vision capabilities for noisy, real-world documents rather than clean sample data. How to use Codex effectively to move fast without sacrificing code quality, by making clear product decisions ourselves and letting Codex accelerate implementation. The value of narrowing scope aggressively under time pressure — a focused, working MVP beats a broad, half-finished one.

What's next for confIA

Explore potential partnerships with financial institutions or retailers to offer real benefits (discounts, credit products) tied to a user's Trust Score — always opt-in and compliant with personal data protection regulations. Extend automation beyond expense logging into automatic payment reminders, closing the loop between spending awareness and bill management. Expand the analytics dashboard with deeper spending pattern insights. Validate and refine the Trust Score formula with real user data over time.

Built With

• codex
• computer-vision
• gpt-5.6
• javascript
• json
• node.js
• ocr
• openai-api
• postgresql
• python
• react
• tailwindcss