BUILD RECORD
Everything-Bot
Where the group-chat agent started. The prototype that proved people will use an AI inside their group chat if it respects the room.
PRODUCT BUILD
Everything-Bot
The message created the runtime. The reply ended it.
The first version of putting the agent inside the group chat rather than behind an app.
The hackathon build became the daily one: letmecheckbot, which has run every day since 2022 across 100,000+ messages.
One message to a media reply, in one short-lived function. 74 MB deployed, about five thousand lines.
One short-lived function per message, FFmpeg inside it. AWS Lambda API Gateway FFmpeg layer
Solo. Architecture, agent behavior, media pipeline, deployment.
Under 3s from cold start to a reply. An idle chat cost zero. Shipped for the AWS Lambda Hackathon.
hackathon telegram group chat retired
THE SUBMISSION
What it did in the thread
Screenshots and voice notes read, history searched in plain English.
THE METHOD
How it answered without leaving the chat
A message woke a function.
It picked fact-check, OCR, transcription, or history search.
FFmpeg rode along inside that function.
It replied where the question was asked, then ended.
Keeping a cheap model in character and on budget was harder than the scaling.
LIMITS
What the design cost
The cold start was the tax. Nothing stayed warm.
It spoke when spoken to, never on its own.
Sources or nothing. The point was the link, not the verdict.
My own group chats. Messages and media stayed private.
RECORD AND LINEAGE
Where the idea went next
letmecheckbot, which took the same job and runs every day.
AgentsThatEmail, the same idea in the inbox.
THE ARCHITECTURE
EXAMPLES
A RELATED PROBLEM
Does useful work already happen in a team chat?
This build is retired, but the interface still matters. If a team should be able to ask, search, and act without leaving the thread, write me.