Dainty Trading
Home / Projects / Whisp

Whisp

A Telegram reminder bot that listens to plain English. Tell Whisp “remind me to call mum every Sunday at 6pm but not while I’m on holiday,” and it figures out what you meant.

Category  AI · Bots Stack  Python · Postgres LLM  Anthropic Claude Status  In final testing

The problem

Existing reminder apps make you fight a date picker. Existing voice assistants make you fight their grammar. Most of the time the right interface is the chat window already open on your phone — you say what you mean, the bot does the rest.

What we built

Whisp is a Telegram bot that accepts natural-language scheduling and turns it into structured reminders. It supports recurring tasks (“every weekday”, “the last Friday of each month”), per-user timezones, quiet hours, vacation mode, and one-shot reminders with relative dates (“in two hours”, “next Tuesday morning”). Tasks live in Postgres; a worker fires them on time and handles silent windows.

The AI angle

Claude does the parsing. The prompt is small, deterministic, and asks for a structured JSON object. We constrain it with a schema and a tight set of examples; the bot only proceeds if the parse round-trips back into a sentence the user agrees with. When the user says “yes that’s right”, both the input and the parse get logged to a regression set we use to vet new prompt versions.

How it’s used

  • Busy parents who want to capture a reminder without unlocking another app.
  • Founders who already live in Telegram for ops chats.
  • Teams using a shared Whisp bot as a lightweight scheduler for routine pings.

What it taught us

That natural-language input only works when the bot reflects back what it understood, every time. The single biggest reduction in user errors came from making Whisp say “OK — reminder set for Tuesday 9pm, repeats weekly. Reply ‘no’ to fix.” before committing. The undo is more important than the parser.