-
-
Complete call workflow, end to end
-
Feature pool
-
/Chain feature, call clinic, ask for appointment if not available, check for alternatives, disconnect, ask approval on telegram call again.
Inspiration
I wanted to create something that can be used by as major number of people as possible. What I can created can be used by scam alert accouchements, appointment bookings, hotels and any client based calling agencies. Also, normal people can use it for their daily purposes.
What it does
You can prompt a telegram bot and it can make calls in behalf of you. It can call and book appointments, make accouchements to hundreds of people, find earliest appointments using chain calling feature and more.
How we built it
Python + python-telegram-bot for the chat interface, Groq for fast natural-language extraction (turning a free-text message into structured clinic/date/patient data), and the CALL-E CLI for the actual phone calls via plan_call → run_call → get_call_run. SQLite + APScheduler handle persistence and scheduling. We built it in phases — starting with a single /call command, then layering on scheduling, multi-clinic calling, fallback chains, and a callback-confirmation flow — refactoring the decision logic into a channel-agnostic core along the way so the same backend can later power a web dashboard, not just Telegram.
Challenges we ran into
Real phone calls surface real-world messiness: a UTC-vs-local-time bug made our "stuck call" safety net fire after seconds instead of minutes; a rogue auto-restart script caused multiple bot instances to silently fight over the same Telegram connection for hours; and one coding-agent mistake nearly rewrote two of our core files from memory instead of making a small edit — recovered only because we'd started committing to git. We also spent real effort ruling out whether occasional instant call failures.
Accomplishments that we're proud of
Real evidence beats confident guesses — nearly every "obvious" bug turned out to be something else once we actually looked at logs, timestamps, and raw API responses. We learned to distrust "tests pass" as a claim and demand real terminal output, to commit early and often, and to never let an AI coding agent rewrite a whole file when a targeted edit would do.
What we learned
Learnt a lot about AI calling agents, automations on Telegram and more.
What's next for ATC (Automated Telegram Caller)
A web dashboard sharing the same backend as the Telegram bot, for enterprise/multi-channel use. Recurring daily calls (checking on report status, doctor availability) and sequential broadcast calling (invitations, announcements) for more general-purpose use cases beyond appointment booking.
Log in or sign up for Devpost to join the conversation.