Libreta

Inspiration

"Lo apunto en la libreta." Every cash business I know runs on a notebook behind the counter.

For the past five years I've worked as an ATM field technician, servicing and refilling my own fleet of machines in bodegas, corner shops, neighborhoods, and many other cash businesses. Before that, I did bilingual customer service at Oro Latino, a Chinese-Latino-owned jewelry store in Manhattan. I've watched these businesses up close: real customers, real revenue, almost all of it in cash.

The problem is that cash income is invisible. An owner can bring in tens of thousands of dollars a year and still get turned away for a small loan, because they have no statements, no books, and no paper trail a lender will accept. The money is real; the history just isn't written down in a way a bank can read. Most bookkeeping software assumes English, a bank account, and an accountant. Our community has none of those on the counter. It has a notebook and a phone.

What it does

Libreta is Spanish-first, voice-first bookkeeping for cash businesses, with the paperwork end done in English for lenders.

• Log it like you'd tell a friend. Type it, say it, or snap a supplier receipt. "Vendí 340 en efectivo hoy, le pagué 120 al de los refrescos" becomes two separate, categorized entries.
• One short question when something's missing. If an amount or detail is unclear, Libreta asks in Spanish with tap-able answers, then closes the entry.
• It talks back. Every entry is read back out loud, and Escuchar mi día speaks today's summary: what came in, what went out, what's left.
• Bring your history with you. Pasa tu cuaderno lets owners photograph the pages of the paper notebook they already keep (or upload a spreadsheet), review the extracted entries, and import months of history in minutes.
• The bank-ready report. One tap produces a clean English PDF with monthly revenue, expenses, and net cash flow, laid out the way a credit union, CDFI, or SBA microloan officer expects to see it.
• Taxes, roughly handled. A quarterly set-aside estimate with a rate the owner chooses:

$$\text{set-aside} \approx r \times (\text{income} - \text{expenses})$$

• Try before you sign up. Guest mode lets anyone use the full app first; signing in with Google or an email code keeps everything. Each business only ever sees its own records.

Everything the owner sees is in Spanish. Only the lender report is in English, because that's the language of the person on the other side of the loan desk.

How I built it

Base44 is the backbone. The data model (businesses, movements, categories, saved reports), authentication and guest mode, private file storage for receipt photos and voice notes, the AI calls that interpret entries and read handwritten notebook pages, the backend functions, and the entire UI all run on Base44.

ElevenLabs handles both directions of voice:

• Speech-to-text transcribes voice notes in Spanish and tolerates Spanglish, which is how people actually talk at the register.
• Text-to-speech (multilingual v2) reads back entries, follow-up questions, and the daily summary. Owners can pick a voice and preview it. Generated audio is cached per sentence and voice so repeat lines play instantly without regenerating. The API key stays server-side.

The flow: voice / text / photo → transcription or OCR → AI interpretation into one or more entries → follow-up question if needed → saved movement → spoken confirmation, with the reporting layer aggregating movements into monthly views and the lender PDF.

Challenges I ran into

• One sentence, many entries. Real speech mixes sales and expenses in one breath. Getting the interpreter to split a single message into separate, correctly signed entries took several iterations.
• Handwriting. Notebook pages are messy, abbreviated, and inconsistent. We added a review step so the owner confirms every imported line instead of trusting the AI blindly.
• Privacy vs. a demo anyone can try. Strict per-business data isolation meant a first-time visitor would see an empty ledger, so we had to design guest mode and sample data without weakening real accounts.
• Voices that sound like us. The default multilingual voices carry a soft English accent, which breaks trust with the exact people we're building for. We made the voice swappable so a native Latin American voice can be dropped in.

What I learned

• The hard part isn't the AI; it's the last mile of trust. Spanish, voice, and the owner's own notebook do more for adoption than any feature.
• Lenders don't need fancy dashboards. They need consistent monthly numbers they can verify. Designing backwards from the loan officer's checklist shaped the whole product.
• Building on Base44 let us spend the day on the problem instead of the plumbing.

What's next

• WhatsApp logging, so owners never have to open an app: si sabes mandar un WhatsApp, ya sabes usar Libreta.
• Pilots with real businesses and a partnership with a CDFI or credit union to accept Libreta reports as part of loan applications.
• Native regional voices and Spanish-language tax guidance for quarterly estimates.

Built at Latinos x AI Hackathon NYC by Adam Martinez.

Built With

• base44
• elevenlabs
• elevenlabs-speech-to-text
• elevenlabs-text-to-speech
• github
• javascript
• react