master summarizer

Caffeine-powered, vibe-coded multi-agent document summarization pipeline initially developed for processing Colombian municipal planning documents into a styled executive summary and 80-100 PowerPoint slide outlines. The entire pipeline coordinates multiple LLM models through an OpenCode server.

Pipeline Stages

1. Parse & Preprocess - DOCX/PDF parsing + Gemini image descriptions.
2. Chunk - Deterministic semantic chunking (heading-aware, token-budgeted).
3. Summarize Chunks - Per-chunk faithful summaries (Claude Sonnet).
4. Learn Style - Infer style guide from example PDFs + manual guidelines (Claude Opus).
5. Central Synthesis - Master draft from all summaries following style guide (Claude Opus).
6. Review - Systematic verification + refinement loop (GPT-5.2).
7. Slide Generation - Structured slide outlines from final draft (GPT-5.2).

Stages 1-2 are deterministic. Stages 1b and 3-7 use LLM agents. The pipeline pauses between stages for manual confirmation and saves state for resumability. All LLM stages are resilient: each item (image, chunk, draft section, slide section) is saved atomically to disk immediately after completion. On restart, already-completed items are skipped automatically—no tokens are wasted.

Stack

• Python (>= 3.11)
• uv for fast package management
• OpenCode CLI orchestrator