VoxLog

The problem

A site supervisor spends the day on site, not at a keyboard. The log gets written in the evening, from memory, and whatever was not noted disappears. Reports produced that way are incomplete and impossible to defend in a dispute.

How it works

You press a button and speak, from ten seconds to ten minutes. Audio is saved second by second on the phone, transcribed, then split into events: crew, work, delivery, issue, safety, weather, follow-up. At the end of the day the report is assembled by code, from verified facts only.

Core capabilities

• Installable mobile app, usable offline on site
• Dictated time wins over the note time ("quarter to four" = 3:45 p.m.)
• Photos attached automatically to the nearest fact, correctable in one tap
• Follow-up items carried day to day until a dictated closure
• PDF and Markdown reports shareable from the phone, locked once final
• Search across history with a supporting quote, or a "not verifiable" answer

Stated limits

• Version 0.2: usable end to end, but not yet through a full construction season
• A fact whose quote cannot be found in the transcript is excluded until a human confirms it
• Transcription quality depends on ambient noise and diction
• The local stack needs a decent machine; the cloud option moves data outside Quebec

Technologies

• Progressive web app, TypeScript, API and interface monorepo
• Whisper transcription running locally
• Qwen3 language model served by Ollama, on the machine
• Deterministic code checks between the model and the report
• Open-Meteo weather, PDF and Markdown generation