Open-source deposition transcript summarization for litigation. Generates four summary formats from a single transcript PDF using local LLMs via Ollama.
- Page-Line — Key testimony indexed by page:line reference
- Narrative — Flowing prose summary of the deposition
- Topical — Organized by subject matter (custody, finances, conduct, etc.)
- Q&A Extract — Critical questions and answers pulled verbatim
Transcript PDF → OCR (if scanned) → Q&A block segmentation → LLM summarization → .docx output
Designed for local GPU inference. Runs on Ollama with models like qwen3.5:35b (32K+ context). A 200-page deposition processes in ~5 minutes on V100 hardware.
Scaffolded — pipeline architecture defined, implementation in progress.
MIT