Pulse: AI community manager built on Swytchcode
Link to open source: https://github.com/RudraO2/pulse
Built for Track 3 (AI Community Agent).
Members ask on Telegram, Slack or WhatsApp. Pulse answers from the community's Notion knowledge base in about a second. When it doesn't know, it asks the organizers in Slack #mods, relays their answer, and saves it to Notion, so the next person gets an instant answer. It reads each member's mood, follows upset members until they're sorted, and brings in a human (email + phone push) when needed.
It's a real agent, not a pipeline: every message runs an AI tool loop where the model chooses to search, reply, ask organizers, flag a member, propose a member's answer for the FAQ, or stay silent.
Swytchcode is the core. Every Telegram, Slack, Notion and Resend call goes through it: 34 allow-listed methods, 9 policies enforced before execution (no leaked secrets, no @channel pings, no phishing links, email only to organizers), a guardrail self-test, dry-run previews before any broadcast, and a full audit log.
Organizers get an Agent Console ("Lunch moved to 1:30, tell everyone, update the FAQ"): Pulse plans, shows the exact Swytchcode dry-run request, waits for approval, then executes. There's also a phone app for one-tap approvals.
Live today: Pulse is in this event's WhatsApp group. It answers from the Participant Guide, knows what's happening right now from the schedule and the clock, and escalates what it doesn't know to the organizers.
Stack: TypeScript, Vercel AI SDK, Groq / Gemini / Claude, Swytchcode CLI + policies, Hono, React, SQLite, Baileys.
Image Gallery: upload from C:\Users\rpxi1\OneDrive\Documents\Desktop\Swytchcode
