ZBot

ZBot is an agentic Discord AI that autonomously prioritizes conversations, sends urgent alerts, and summarizes messages. It learns from your corrections to improve future classifications.

Comment

Inspiration:

We often return to Discord after being offline to find 50+ unread messages. We start reading but eventually give up, risking missing an important message buried among dozens of conversations. As our servers and group chats grow, keeping up with everything becomes increasingly difficult. This inspired us to create ZBot, an AI agent that summarizes conversations and prioritizes important information so we can catch up without reading every message.

What it does:

  • Autonomous Message Prioritization: Groups related messages into conversations and categorizes them into four levels:

🔴 VVIM: Very Very Important Messages requiring urgent attention.

👀 Take a Look: Relevant conversations worth reviewing.

😂 Fun Stuff: Memes, jokes, and entertaining conversations.

🗑️ Ignore: Conversations unlikely to need attention.

  • Urgent Alerts: Sends direct messages when important conversations require attention.

  • Customizable Recaps: Lets users request summaries for the last hour, last three hours, yesterday, or today, and ask “What did I miss?”

  • Interactive Discord Dashboard: Provides an in-Discord interface to explore summaries and navigate between categories, with a refresh option for new messages.

  • Feedback-Based Learning: Lets users correct misclassified conversations. ZBot uses these corrections to improve future classifications of similar conversations.

  • Reset Learning: Allows users to forget previously taught corrections and start fresh, giving them control over what ZBot has learned from their feedback.

  • Custom Interests: Users lists topics they care about. Matching conversations get a ⭐️ and move to the top of their section, and borderline ones are lifted into 👀 Take a Look. Interests never raise anything to 🔴 VVIM.

ZBot goes beyond basic summarization by continuously monitoring conversations, making explainable prioritization decisions, and adapting to each user's preferences.

How it does:


   Discord channels (only ones a server admin allowed)
        │  every message is stored (text, author, time; images described once; GIFs recognized)
        ▼
 ┌──────────────────────── every minute, on its own ────────────────────────┐
 │ 1. Group      messages → conversations (time gaps, replies, shared topic) │
 │ 2. Retrieve   RAG: similar labeled examples, your ✏️ corrections,         │
 │               related earlier conversations in the channel                │
 │ 3. Understand LLM writes a one-line summary + 4 scores (0-10):            │
 │               relevance · urgency · actionability · entertainment         │
 │ 4. Decide     rules.json (explicit criteria, first match wins) picks      │
 │               🔴 VVIM · 👀 Take a Look · 😂 Fun Stuff · 🗑️ Ignore           │
 │ 5. Act        urgent → DM alert · you come online → welcome-back summary │
 │               · recap times · "What did I miss?" · /z dashboard           │
 └───────────────────────────────────────────────────────────────────────────┘
        ▲
        └── learns from you: move a conversation with ✏️ and look-alikes follow (just for you)

Design choices worth knowing:

  • Context, not keywords: A conversation is judged with its earlier messages and related channel history, so "ok let's do render tonight" counts as a decision, while a bare "lol" stays noise.

  • Privacy by design: Z only reads channels a server's admins allowed, each person only ever sees channels they can see in Discord (checked live), images are described once and never stored, documents are never opened, and /z forget_me deletes everything about you.

  • Doesn't invent or judge: Summaries use only what the messages say, and Z describes what happened, never the people.

  • Free to run: Groq's free AI tier and Render's free hosting, with encrypted backups to Discord because the free host wipes its disk on restart.

Challenges:

  • Implementing RAG: Initially, ZBot categorized conversations mainly using numerical scores, often missing context. Implementing RAG allowed it to retrieve similar examples, previous corrections, and conversation history, improving contextual understanding and reducing misclassifications.

  • Cloud Deployment: Since this was our first experience with cloud hosting, deploying ZBot to run continuously without a local computer took additional time and troubleshooting.

  • Discord Permissions: Deploying ZBot across multiple servers requires authorization and appropriate permissions from each server's administrators, which adds a setup requirement.

Accomplishments + Learnings:

  • Built our first agentic AI within a limited timeframe, capable of continuously monitoring authorized Discord channels, prioritizing conversations, and sending urgent alerts.
  • Integrated Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) to improve contextual understanding, reduce misclassifications, and personalize future decisions using user feedback.
  • Gained hands-on experience with cloud deployment, learning how to host and maintain an AI agent beyond our local machines.
  • Learned how agentic AI combines contextual analysis, rule-based decisions, and autonomous actions to prioritize information without requiring a user prompt for every task.
  • Recognized the importance of teamwork, clear communication, and meeting deadlines while developing a complex project under time constraints.

What's next for ZBot?

We plan to continue improving ZBot's reliability, intelligence, and user experience by focusing on the following areas:

  • Improve Reliability: Test ZBot across more Discord servers, identify recurring classification errors, and use the findings to reduce hallucinations and improve accuracy.
  • Enhance the Dashboard: Build a more interactive dashboard that makes browsing summaries and navigating between categories easier.
  • Smarter Summaries: Improve the summarization process to make summaries more concise, accurate, and context-aware.
  • Improve Presentation: Use emojis more effectively to make summaries easier to scan and more visually engaging.
  • Refine Our Branding: Explore improvements to ZBot's name, logo, and overall visual identity as the project evolves.

Built With

Share this project:

Updates

Submission history