[Name TBD] — Enterprise AI Agent via WhatsApp

An AI agent that silently reads all your business email and Slack, but only interrupts you in WhatsApp when something actually needs your attention. You approve or edit with a tap or a voice note. Every action is Ed25519 signed for compliance.

Built at YHack 2026 | Harper "Personal AI Agents in Enterprises" Track

---

The Problem

Insurance brokers get 50+ emails and Slack messages a day. They don't have time to read them all — but they can't afford to miss the urgent ones. Current solutions either require another dashboard (that nobody checks) or let AI act autonomously (which goes wrong — ask the Meta researcher whose inbox got wrecked by an AI agent).

The Solution

An AI agent that lives where you already are — WhatsApp. It:

1. Silently ingests all email and Slack in the background
2. Categorizes and ranks every message by urgency (urgent / action-required / FYI / low-priority)
3. Only interrupts you for what matters — via WhatsApp interactive buttons
4. Drafts responses you can approve with one tap, edit via voice note, or skip
5. Signs every action with Ed25519 for a tamper-proof audit trail

The agent's job is to NOT show you things. 50 messages come in → it surfaces 2 that matter. The value is the 48 context switches you didn't make.

---

Architecture

  Raspberry Pi (on the table)              Railway (cloud)
  ============================             ==================

  Gmail (IMAP) ──► NATS pub ─┐
                              │
  Slack (Bot)  ──► NATS pub ─┤
                              │
                    NATS Server ◄────────► WhatsApp Bridge
                    (JetStream)            (Node.js)
                              │            - Meta Cloud API
                    messages.inbound.*     - Interactive buttons
                              │            - Voice note handling
                              ▼
                    OpenFang Agent
                    - Categorize + Score
                    - Draft response
                    - Ed25519 sign
                              │
                    messages.ranked ──────► WhatsApp alert
                                           [Send] [Details] [Skip]
                              │
                    messages.approved ◄──── User taps / voice note
                              │
                    Agent sends reply via
                    original channel (SMTP/Slack)
                              │
                    Ed25519 signed receipt → SQLite

Why This Architecture?

• NATS wildcard subscriptions — Agent subscribes to messages.inbound.*. Adding a new channel = one 10-line publisher. Zero agent code changes.
• Raspberry Pi — The entire agent OS runs on a $35 board. Edge computing for enterprises that can't send data to the cloud.
• Railway — Only the WhatsApp bridge needs a public URL (for Meta webhook callbacks). Everything else runs on the Pi.
• Progressive disclosure — WhatsApp buttons let you go from 1-line alert → full gist + draft → complete email body, without ever leaving the chat.

---

WhatsApp UX

Alert (only for urgent/action-required)

🔴 John Chen — COI for 42 Main St
Needs cert by EOD (2nd follow-up)

[✅ Send]  [📋 Details]  [⏭ Skip]

Details (tap to expand)

From: john.chen@acmecorp.com
Subject: RE: COI for 42 Main St
Via: Email │ 10:23 AM │ Thread: 3 msgs

Gist: Following up 2nd time since Monday.
Contractor starts Thursday.

Draft: "Hi John, pulling the cert now.
You'll have it within the hour."

[✅ Send] [✏️ Edit] [🔗 Full]

Voice

🎤 "anything urgent?"
→ "Nothing urgent. 3 action items — biggest
   is Acme renewal expiring Friday."

🎤 "edit the reply to mention Thursday deadline"
→ Draft updated. [✅ Send] [✏️ Edit More]

---

Tech Stack

Component	Tech	Where
Agent OS	OpenFang (Rust, 32MB)	Raspberry Pi
Message bus	NATS + JetStream	Raspberry Pi
WhatsApp	Meta Cloud API	Railway
Email	IMAP/SMTP adapter	Raspberry Pi
Slack	Socket Mode adapter	Raspberry Pi
LLM	Groq (speed) + Gemini 2.5 Flash (quality)	API
Voice STT	Groq Whisper	API
Audit store	SQLite + Ed25519 signatures	Raspberry Pi
Backend	TypeScript / Node.js	Both

Security

• Ed25519 signatures on every categorization, approval, and send action
• WASM sandbox — OpenFang runs agent code in isolated WASM
• Merkle hash-chain audit log — immutable, verifiable
• NATS NKeys — Ed25519 client-to-broker authentication
• Channel isolation — throwaway accounts, scoped bot tokens, test numbers

Cost: $0

Everything runs on free tiers: Railway ($5 trial), Gemini (free), Groq (free), Meta WhatsApp (test number), MLH Raspberry Pi.

---

Setup

# Clone
git clone https://github.com/<username>/yhack-agent.git
cd yhack-agent
npm install

# Pi: Start NATS + OpenFang + adapters
npm run pi:start

# Railway: Deploy WhatsApp bridge
npm run bridge:deploy

See architecture.md for detailed setup, NATS subjects, and message schemas.

---

Built with OpenFang, NATS, Gemini, and Groq at YHack 2026.