Inspiration

My inspiration is to review or reshape how the workers of FNB took the orders from the customers

What it does

CakapBayar is a voice-powered POS system for Malaysian food stalls. What it does:

Customer speaks order in Malay/English: "Saya nak 2 teh tarik" AI transcribes & processes order automatically Worker just confirms payment System tracks sales, manages shifts, sends reports

Replaces: $300/month English-only POS systems with a FREE bilingual solution that understands how Malaysians actually speak.

How we built it

Built with: Next.js + React - Frontend Groq Whisper - Transcribes Malay audio to text Groq Llama 3.3 - Parses orders with Malaysian context Convex - Real-time database for transactions/menu ElevenLabs - Voice announcements in Malay Cursor AI - Coded entire app in 12 hours

Flow: Voice → Transcribe → AI Parse → Display Order → Confirm Payment → Save to Database

Challenges we ran into

Challenges: OpenAI API failed 2 hours in - Switched to Groq (saved the project) Malay-English code-switching hard for AI - "Saya nak tiga roti" parsed "tiga" as item name, not quantity Mobile mic permissions broken - Needed HTTPS on localhost Git authentication 403 errors - Username mismatch (mirulhaziq vs mirulhziq) UI too complex for workers - Redesigned with bigger buttons, Malay labels, simpler flow Convex schema changes broke old data - Had to write migration script 8-10 second delays felt slow - Optimized to 2-3 seconds with parallel processing Claude is AI and can make mistakes. Please double-check responses.

Accomplishments that we're proud of

Solved real problem - Built a $300/month POS alternative that's FREE and works for Malaysian food stalls True bilingual AI - First system that understands "Saya nak 2 teh tarik and 1 roti" (Malay-English code-switching) Built in 12 hours - Fully functional app with voice ordering, payments, shift management, email reports using Cursor AI

What we learned

Groq is a game-changer - 10x faster than OpenAI, free tier, saved our project when OpenAI failed Code-switching is linguistically complex - Malaysian speech patterns need specialized prompting, not just multilingual models Cursor AI = massive productivity boost - What takes 4 hours normally took 30 minutes. 80% correct code first try Small businesses want different features - Scrapped fancy ML analytics. They want: fast, simple, cheap, offline

What's next for CakapnBayar

stay tuned

Built With

  • anthropic
  • convex
  • cursor
  • elevenlabs
  • groq
  • next.js
  • react
  • whisper
Share this project:

Updates