Skip to content
Back to the portfolio

Tucano: Finances and Shopping with a Telegram Agent

  • React
  • TypeScript
  • Convex
  • Vite
  • TailwindCSS
  • Vercel
  • OpenRouter
  • Telegram
  • Cursor

My role

Personal project: I made the whole product, from design and visual identity to the Telegram agent.

Budgeting dies at the moment of writing things down. Tucano brings finances and the shopping list into one app and removes the friction of logging: you can record an expense by sending a message, or a voice note, to a Telegram bot.

Telegram agent: "spent 50 at the grocery store on my Nubank card in 3 installments" becomes a card transaction, split into installments and already categorized. It understands voice notes, keeps the conversation context and answers questions such as the monthly summary.
Finances: income and expenses by billing month, pending or paid, recurring subscriptions and credit cards.
Shopping: monthly lists with a budget, and a product catalog with default brands and unit prices.
Dashboard: the month's analytics on a single screen.
Mobile-first: bottom navigation, a floating action button and bottom-sheet forms, so logging takes seconds.

Under the hood: React 19 + Vite + TypeScript on the front end, Convex as the backend (database, auth, storage and HTTP actions for the Telegram webhook), OpenRouter for the agent, OpenAI for voice transcription and continuous deployment on Vercel.

Before

The first version, on Firebase, relied on forms for everything, and the dashboard query made about 91 database reads.

The problem

An expense spreadsheet lasts about two weeks. The hard part was always the logging: opening the right app, finding the category, remembering which card you used and how many installments. Every extra step is an expense that never gets recorded, and tracking with gaps is useless for making decisions.

The first version of Tucano, built on Firebase, already combined finances and the shopping list, but everything still went through forms. What was missing was a way to log things on the spot, the way people actually talk.

The solution

I rewrote Tucano from scratch with React 19, Vite and TypeScript on the front end and Convex as the backend: database, auth, storage and real-time queries, with continuous deployment on Vercel shipping Convex alongside it.

The centerpiece is a Telegram agent. A webhook (a Convex HTTP action) receives the message; if it is a voice note, OpenAI transcribes it; a model on OpenRouter interprets the request and returns one of 13 structured actions (log an expense, an installment purchase on a card, a subscription, a shopping list item, the monthly summary…). Each action is validated before it becomes a use case, the category is picked from the user's own categories, and the conversation context is stored for multi-turn requests.

In the app, the interface is mobile-first: bottom navigation, a floating action button and bottom sheets. The dashboard brings together the month's analytics; its query started at about 91 database reads and dropped to 3 with batched fetching. The visual identity uses a single color, green on near-black, and the logo grows out of the monthly view's ring.

Decisions and trade-offs

The model returns one of 13 structured actions, validated before becoming a use case, instead of acting on free text. The category is picked from the user's own categories. The dashboard switched to batched reads, going from about 91 reads to 3.

The results

Tucano is live, with the Telegram agent working end to end. Logging an expense now means sending a message or a voice note; the entry lands with the amount, card, installments and category already filled in, and shows up in the app in real time. The rewrite took 40 commits, with Vitest and convex-test suites covering the agent, the transcription and the webhook.

Results by the numbers

  • 91 → 3Dashboard reads
  • 13Agent actions
  • 10Convex tables
  • 40Rewrite commits
  • 10Test files