Inspiration
Millions of Indonesian warung owners and ojol (ride-hailing) drivers keep their books in their heads or on scraps of paper. Typing numbers into an app mid-shift just doesn't happen, but saying one sentence does: "Bu Sri ngutang lima belas ribu" (Bu Sri owes fifteen thousand), "isi bensin goceng" (fuel, five thousand).
What it does
Buku Bon is a voice-first ledger skeleton with two very different apps built on it:
- Kasbon Warung — a corner shop's credit book. Say who owes what; balances per customer update instantly. "Pak Budi lunas" settles a customer's whole balance to zero.
- Dompet Ojol — a driver's daily net. "dapat orderan dua puluh tiga ribu", "dapat tip goceng", "isi bensin dua puluh ribu". Set "target hari ini 200 ribu" and see progress plus the last 7 days.
Both apps understand spoken numbers in words, fast speech like "satu juta dua ratus ribu", digit formats (15rb, Rp 15.000, 1,5 juta) and street slang (cepek, goceng, ceban). Every entry is shown for confirmation before saving. No account, no server: data stays in the browser, works offline (service worker), installs as a PWA, and exports to CSV. If the browser has no speech recognition, typing the same sentence works.
How we built it
Plain ES modules, HTML and CSS — no framework, no build step, no runtime dependencies. The skeleton (skeleton/) holds the parser, store, Web Speech wrapper (id-ID) and the UI mount. Each app folder (app-warung/, app-ojol/) contains only an index.html and a config.js that declares intents, keywords, slang tags and the summary view. A third app is a copy and an edit away.
It was built spec-driven with Kiro: steering docs, an EARS spec with design properties and a task list, agent hooks, and property-based tests. Tasks 10–14 were implemented by Kiro CLI in real sessions (condensed logs in docs/kiro-sessions/). Details are in the Kiro section of this submission.
Challenges we ran into
- Indonesian spoken numbers stack scales (juta, ribu, ratus) and contractions (sejuta, seratus, setengah); speech recognizers mix words and digits. The parser uses a three-accumulator algorithm and is checked by a round-trip property test on every amount from 1 to 9,999,999.
- Keeping the skeleton truly generic: "lunas" (settle) and "target" needed new behaviour, which Kiro added as generic config flags (
settle,meta) instead of app special cases. - A weekly chart that drew no bar for today: a smoke test caught it, and Kiro traced it to a percentage height inside a flex column without a definite height.
Accomplishments that we're proud of
Two genuinely different apps from one tiny, dependency-free skeleton that works offline on a cheap phone, in the language people actually speak.
What we learned
Spec + steering made the agent's changes predictable: it extended the skeleton generically because the steering said so, and it updated the spec (R7–R9) as it went.
What's next for Buku Bon
A third app (market stall or arisan), voice read-back of balances, and WhatsApp-friendly reminders for customers who owe.
Built With
- css3
- github
- html5
- javascript
- kiro
- pwa
- service-worker
- web-speech-api