Technical Documentation

About the Buffy Bot System

The Buffy Bot is an AI-powered conversation analysis and academic paper generation system. It extracts discussions about Buffy the Vampire Slayer from Zulip chat logs and generates academic papers using a multi-agent AI workflow.

System Architecture

1. Conversation Extraction

The system extracts private conversations from Zulip and optionally breaks them into weekly chunks for detailed analysis.

uv run main.py extract-private --limit 10000 --weekly-chunks

2. Graduate Bot Analysis

Graduate-level AI agents analyze the weekly conversation chunks, identifying key themes, philosophical concepts, and noteworthy observations.

uv run main.py run-grad-bots --weekly-dir "conversations/weekly"

3. Research Bot

The researcher bot synthesizes the graduate notes to propose 5-8 compelling paper topics, each with: - A detailed abstract - Evidence assessment (0-100 rating) - Key episode references - Novelty evaluation

uv run main.py researcher-bot

4. Conference Simulation

The run-conference command fully automates the academic lifecycle:

uv run main.py run-conference

This command: - Generates papers for all highly-rated abstracts - Reviews each paper with multiple AI peer reviewers - Rewrites papers based on feedback - Continues revision cycles until all reviews are ACCEPT - Features realistic academic filename chaos (paper_copy1_FINAL2.md, etc.) - Includes acceptance bias that increases with longer filenames (simulating academic reality)

Paper Generation Pipeline

Postdoc Bot

Rates each weekly note for relevance to the paper topic (0-100 scale). Only weeks meeting the threshold are included.

Research Assistant

Selects the most relevant episode scripts based on the topic and conversation analysis.

Professor Bot

Writes the academic paper using: - Filtered weekly notes (or verbatim transcripts for highly relevant weeks) - Selected episode scripts - Theoretical frameworks appropriate to the topic

Output Structure

Each paper folder contains: - paper.md / paper_final1.md - The paper in Markdown format - paper.pdf / paper_final1.pdf - PDF version - postdoc_ratings.json - Relevance ratings for each week - reviews/ - Peer review JSON files - scripts/ - Episode script files used

Revised papers get version suffixes (_v2, _v2_v2) and accumulate increasingly chaotic filenames.

Configuration

Key parameters are configurable in prompts.toml: - max_weeks_for_paper - Maximum weeks to include (default: 8) - Agent prompts and system messages - Review criteria and standards

Technologies Used

  • Claude Sonnet 4: All AI agents use Claude for text generation
  • Python: Core system implementation
  • uv: Dependency management
  • Rich: Terminal UI
  • Markdown & PDF: Paper output formats
  • Zulip API: Conversation extraction

Static Site Generation

This website was generated using static_site_generator.py, which: - Scans all paper folders in the papers/ directory - Identifies the final accepted version of each paper series - Generates responsive HTML pages with embedded CSS - Creates individual pages for papers and reviews - Copies PDFs for download

uv run static_site_generator.py --papers-dir papers --output-dir conference_site

Source Code

The complete source code is available in the project repository. Key files: - main.py - Core CLI and agent orchestration - prompts.toml - Agent prompts and configuration - static_site_generator.py - Website generation - grad_bot_logger.py - Logging utilities

← Back to Conference