Why I built Diao Tham

In Thai, เดี๋ยวทำ means “I’ll do it in a bit.” It’s an easy promise to make in a group chat. The hard part comes tomorrow, when I remember promising something but have to scroll back to find out what it was and when it’s due. I built Diao Tham so I could capture a message while I was busy and still have its exact wording when I had time to plan.

One message, two deadlines

Organize the project files tomorrow at 10 am and archive the screenshots tomorrow at 11 am.

I used this message to test the hosted app on November 18 in Bangkok. Firebase AI Logic returned two tasks for November 19:

  • 10:00 AM — Organize the project files
  • 11:00 AM — Archive the screenshots

Both appeared in My tasks. I could edit either one, check the original message, and choose Add to Calendar to put it on the app’s calendar. Google Calendar sync is a separate action with its own permission. That distinction matters when a message contains someone else’s suggested time and I haven’t agreed to it yet.

What makes an extracted task trustworthy?

A time can look precise and still be wrong. If I review a message on Thursday, “tomorrow” in a message from Monday still means Tuesday. The app passes the message’s reference date and time zone to the model. It asks for a title, date, time, and an exact quote supporting each action. The TypeScript validator rejects a result if its quoted evidence doesn’t appear in the original. The model can leave a date or time empty and flag the ambiguity for review.

Where the AI runs

Diao Tham saves the original in the browser before calling a model. For on-device extraction, it checks Chrome’s model readiness and the message’s language before requesting Gemini Nano 4 through Firebase AI Logic. When local AI is unavailable, the saved message can wait. The user can also approve sending that specific message to cloud Gemini. Firebase App Check protects the hosted AI requests.

I verified cloud extraction with Thai and English examples and verified the in-app calendar on the deployed site. The on-device path is implemented; a successful Nano 4 run on compatible hardware and a live Google Calendar write remain to be verified.

Try it as a judge

Open the demo in desktop Chrome. For the account-free route, choose Set up on this device → Basic helper (no AI) and capture the example above. Choose Review later, reopen it from My tasks with Review message, then choose Try cloud AI → Send this text to AI. Compare the two times with the source message, edit one, and add it to the in-app calendar. That last step is why I wanted this tool: the plan reflects the time I actually intend to keep.

Built With

Share this project:

Updates

Submission history