A sarcastic Slack bot that corrects people when they're wrong. Tag @Samaraj on a message and it will deliver a factual correction with dry wit and a subtle burn.

Powered by a self-hosted Llama 3.2 3B model — no third-party AI APIs involved.

1. Someone posts something wrong in a channel
2. You reply to that message with @Samaraj
3. Samaraj reads the message, generates a snarky correction, and replies in the thread

You can also DM Samaraj directly — just send a message and it'll correct you.

User: Python was created in 1995

Samaraj: Python was actually created in 1991 by Guido van Rossum. I'm not sure what decade they were counting on, but it's not 1995.

• Python with slack_bolt (Socket Mode)
• Llama 3.2 3B via llama-cpp-python (CPU inference)
• GGUF quantized model (Q4_K_M, ~2 GB) from HuggingFace
• Docker container deployed on Railway

• Python 3.11+
• A HuggingFace account (free)
• A Slack workspace where you can install apps
• A Railway account (for deployment)

git clone <your-repo-url>
cd samaraj
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Go to meta-llama/Llama-3.2-3B-Instruct on HuggingFace and accept Meta's license agreement.

Go to huggingface.co/settings/tokens and create a Read token.

1. Go to api.slack.com/apps > Create New App > From scratch
2. Name it "Samaraj", select your workspace
3. Socket Mode > Enable > Create app-level token (name: samaraj-socket, scope: connections:write) > Save the token as SLACK_APP_TOKEN
4. OAuth & Permissions > Add Bot Token Scopes:
   • app_mentions:read
   • chat:write
   • channels:history
   • im:history
5. Event Subscriptions > Enable > Subscribe to bot events: app_mention, message.im
6. App Home > Enable Messages Tab > Check "Allow users to send Slash commands and messages from the messages tab"
7. Install to Workspace > Copy the Bot User OAuth Token as SLACK_BOT_TOKEN
8. Invite @Samaraj to channels: /invite @Samaraj

cp .env.example .env

Fill in your .env:

SLACK_BOT_TOKEN=xoxb-your-bot-token
SLACK_APP_TOKEN=xapp-your-app-token
HF_TOKEN=hf_your_huggingface_token
MODEL_PATH=models/Llama-3.2-3B-Instruct-Q4_K_M.gguf

python scripts/download_model.py

This downloads ~2 GB.

python -m src.app

1. Push the repo to GitHub
2. Go to railway.app > New Project > Deploy from GitHub Repo
3. Add environment variables in the Railway dashboard:
   • SLACK_BOT_TOKEN
   • SLACK_APP_TOKEN
   • HF_TOKEN
4. Deploy — Railway builds the Docker image and downloads the model during build

Test Samaraj's responses without the Slack connection:

python scripts/test_locally.py "The Great Wall of China is visible from space"

├── src/
│   ├── app.py           # Slack event handler (main entrypoint)
│   ├── model.py         # Llama model wrapper
│   └── prompts.py       # Samaraj's personality prompt
├── tests/               # Unit tests
├── scripts/
│   ├── download_model.py  # Download model from HuggingFace
│   └── test_locally.py    # Test without Slack
├── Dockerfile           # Container with model baked in
└── railway.toml         # Railway deployment config

python -m pytest tests/ -v