A tradesperson describes a job in a voice note and gets a quote with real prices back. It all runs locally: Whisper for speech, Gemma 3 4B through Ollama, and a small FastAPI app on top.
Demo :
I upload a voice note, check the transcript and the lines, and download the PDF. You can speak in any language, and the quote comes out in English, Hindi or Spanish, so a Hinglish note can become a Hindi quote.
Who'd pay
Solo electricians, plumbers and handymen, people who price a job by talking through it.
The one design rule
The model reads, the code does the money. Gemma never sees the prices in my price book. It gets item ids, names and aliases, and returns which item the speaker meant and how many. Python looks up the price and does the math in Decimal. A 4B model is fine with "two valves, no, three" and shouldn't be trusted with arithmetic.
Top comments (0)