Inspiration

The name Chalkup comes from the old phrase “chalk it up”—the classic tradition of writing a customer's debt on a slate at the corner shop.

We noticed that while fintech has advanced immensely, millions of micro-entrepreneurs—home bakers, local tutors, market vendors, mobile mechanics, and small contractors—still run their businesses on informal credit. They track who owes what in dog-eared notebooks, memory, or chaotic chat threads. Existing financial apps are bogged down by heavy forms, required logins, and rigid typing requirements that are nearly impossible to use when your hands are full. Furthermore, blunt payment reminders often strain relationships with loyal repeat customers. We wanted to build something as fast and friction-free as a traditional slate, but supercharged by modern local-first tech and AI.

What it does

Chalkup is a voice-first, local-first informal credit ledger designed for solo and micro-business owners.

  • Natural Voice or Text Entry: Tap the button and speak naturally—for example, "Maria took two dozen cupcakes, owes me forty, pays Friday."
  • Smart Parsing & Validation: An LLM extracts structured fields, strictly validated by a Zod schema.
  • Confirm-Before-Save Safeguard: Nothing is saved automatically. The app presents a confirmation card to review details, using fuzzy matching (Fuse.js) to map names to existing customers or prompt a picker for ambiguous ones.
  • The Ledger: A clean, lightning-fast dashboard tracking customers, running balances, and per-customer transaction histories.
  • Tone-Aware Reminders: Generates friendly, neutral, or firm reminder drafts that can be instantly shared via WhatsApp (wa.me), SMS, or email links without burning bridges.
  • 100% Local-First: All data lives securely right on your device using IndexedDB (Dexie). No accounts, no sign-ups, and no cloud database required.

How we built it

We engineered Chalkup for speed, reliability, and privacy:

  • Frontend & UI: Built with React Router (Framework Mode), TypeScript, Tailwind CSS, and ShadCN UI components to deliver a modern, accessible interface with large touch targets and high-contrast elements.
  • Local Storage: Dexie.js wrapping IndexedDB acts as our sole database, ensuring instant reads/writes and complete offline functionality for viewing the ledger.
  • Parsing & Logic Pipeline: LLM structured outputs coupled with Zod validation. Relative dates (like "Friday") are resolved deterministically in code rather than hallucinated by the model. A robust regex fallback handles basic phrasing offline.
  • Entity Matching: Fuse.js performs fuzzy matching against existing customer records to prevent duplicate profiles when names are spelled slightly differently.
  • Automated Evaluation Harness: Built an automated testing script (pnpm eval) running multi-intent test phrases to measure extraction accuracy across all key fields.
Chalkup parse eval — 12 phrases

  debt-01… OK      debt-05… OK      debt-06… OK
  debt-02… OK      pay-01… OK       debt-07… OK
  debt-03… OK      pay-02… OK       debt-08… OK
  debt-04… OK      pay-03… OK       pay-04… OK

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Field accuracy
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Field        | Correct | Total | Accuracy |
|--------------|---------|-------|----------|
| intent       |      12 |    12 |   100.0% |
| customerName |      12 |    12 |   100.0% |
| amount       |      12 |    12 |   100.0% |
| direction    |      12 |    12 |   100.0% |
| description  |      12 |    12 |   100.0% |
| dueDate      |      12 |    12 |   100.0% |

Exact match (all scored fields): 12/12 (100.0%)

Challenges we ran into

  • Handling Conversational Variance: Spoken language is messy. Users say "put forty bucks on John's tab" or "John owes 40 for the fix". Structuring these diverse inputs reliably required tightening our Zod schemas and separating relative date computation from the model.
  • Balancing Offline Limitations: While browsing the ledger and viewing history work completely offline, voice transcription and LLM parsing require network connectivity. Designing graceful degradation—such as a manual entry fallback and regex parsing—was crucial.
  • Frictionless Confirmation: We had to ensure the "confirm card" didn't add unnecessary clicks while still protecting users from erroneous data entry caused by transcription typos.

Accomplishments that we're proud of

  • 100% Eval Accuracy Score: Achieved 100.0% exact match accuracy across all 12 test phrases covering intents, customer names, amounts, directions, descriptions, and due dates in our automated evaluation harness.
  • Zero-Friction Privacy Architecture: Achieving a seamless user experience without forcing anyone to create an account, hand over email credentials, or sync data to an insecure cloud database.
  • Relationship-Preserving Design: The tone-aware reminder system solves a genuine emotional and social pain point for micro-business owners who dread asking for money.

What we learned

  • Local-First is Liberating: Relying solely on IndexedDB via Dexie made state management lightning fast and completely eliminated server-side auth friction.
  • Code-Side Determinism: Offloading relative date math and calculations to code instead of trusting the LLM completely eliminated date-drift bugs and secured our 100% score on dueDate.

What's next for Chalkup

  • Optional Encrypted Peer-to-Peer Backup: Allowing users to export/import backups securely or sync across their own personal devices without a centralized cloud server.
  • Expanded Currency & Localization: Supporting multi-currency formats and regional payment link integrations.
  • Enhanced Offline Voice Fallback: Integrating browser-native speech recognition APIs for basic offline audio-to-text fallback when internet connectivity is spotty.

Built With

  • cloudflare-worker-ai
  • dexie
  • fuse.js
  • react
  • react-router-framework
  • shadcn-ui
  • tailwind-css
  • zod
Share this project:

Updates

Submission history