Inspiration

What it does

How we built it## Inspiration

Every day we get messages from email, Telegram and class groups. Important ones, like a professor's deadline, get buried under noise. Filters like "mute this sender" can't understand context, and we still have to read, decide, set reminders and write replies ourselves. We wanted an assistant that does that follow-up work for us.

What it does

InboxPilot is an autonomous AI agent that reads incoming messages and completes the follow-up work:

Understands the sender, intent, deadline and priority
Makes a step-by-step plan before acting
Uses tools to create tasks, add calendar events and draft replies
Asks for our approval before risky actions like sending a reply
Learns from our approvals and rejections over time

Example: A message from "Prof. Rao: Submit the lab record by tomorrow 10 AM" is marked urgent. A task and calendar reminder are created, and a reply is drafted and held for approval.

How we will build it

Inputs: Telegram bot (live), Gmail API (optional) and a simulated feed as backup
Backend: Python and FastAPI with an event queue and an enrichment step
Agent: An LLM with tool calling (planner, calendar, task, reply and memory tools) and confidence scores
Memory: SQLite or Supabase for contacts, threads, approvals and decisions
Dashboard: Streamlit or React showing the agent's plan and tool calls live

Challenges we expect

Deciding when the agent should act alone and when it should ask a human
Keeping the agent reliable and fast within 24 hours
Handling privacy by filtering sensitive messages (like OTPs) and never sending anything without approval