Transcripts in. Structured minutes, action items, and accountability ledger out.
End-to-end meeting intelligence pipeline. Parses transcripts, segments topics, extracts action items with owners and deadlines, tracks accountability deltas across a meeting series, and generates executive summaries — with or without an LLM.
For non-technical readers: After every meeting, someone has to turn the recording or transcript into meeting notes, write up who agreed to do what, and track whether last week's action items actually got done. MeetMind automates all of this. Give it the transcript text, and it produces formatted meeting minutes, a list of follow-up tasks with the responsible person's name, and a running accountability log that tracks whether commitments from previous meetings were fulfilled.
MeetMind's core is the process_meeting function — a single orchestration call that runs the full pipeline from raw transcript text to every output artifact:
📝 Raw Transcript Text
│
▼
📖 Transcript Parser
Parses speaker turns, timestamps, utterances
│
▼
🗂️ Topic Segmentation
Detects meeting topic boundaries via
TextTiling-style lexical cohesion scoring
Outputs segment list with boundary indices
│
▼
🔍 Information Extraction
Per segment, extracts:
├── 📋 Action items (task + owner + deadline)
├── 📊 Decisions made
└── ❓ Open questions / parking lot items
│
▼
📈 Meeting Scorer
Engagement, clarity, and action-density scores
per segment and overall
│
▼
📚 Accountability Ledger (cross-meeting delta tracking)
Compares current meeting's action items against
prior ledger state: resolved / carried / new
│
▼
📝 Report Compiler
├── Executive summary (LLM or extractive fallback)
├── Formatted meeting minutes
├── Follow-up task list with owners
└── Accountability delta report
Topic Segmentation — Uses lexical cohesion scoring to detect topic boundary shifts in the utterance stream. The boundary_f1 function evaluates segmentation quality against ground-truth boundaries — ensuring the segmenter generalizes to different meeting styles (standups, design reviews, 1:1s).
Accountability Ledger — The Ledger class maintains a persistent record of action items across a meeting series. Each new meeting's Extraction is ingested via ledger.ingest(extraction) which computes a MeetingDelta: action items resolved since last meeting, items carried over unresolved, and newly committed items. This delta appears in the accountability report.
LLM-Optional Design — The executive summary step (executive_summary) uses an LLM when available but falls back to extractive summarization (selecting high-scoring sentences from the extraction) when no LLM is configured. The summary_mode field in MeetingResult records which path was taken. All other pipeline steps are fully deterministic and require no LLM.
Series Processing — run_series processes a complete synthetic meeting series from scratch: ingesting each transcript in order, updating the ledger state, and emitting all artifacts (minutes, follow-ups, ledger, accountability reports) to the output directory.
git clone https://github.com/nathaniel-gordon/meetmind
cd meetmind
pip install -e .# Process the bundled synthetic meeting series
python -m mag --output output/pytest tests/ -vmeetmind/
├── mag/
│ ├── pipeline.py # End-to-end orchestration: transcript → all artifacts
│ ├── parsing.py # Transcript parser (speaker turns, timestamps)
│ ├── segmentation.py # Topic boundary detection & boundary_f1 evaluation
│ ├── extract.py # Action item, decision & question extraction
│ ├── scoring.py # Per-segment engagement & action-density scoring
│ ├── ledger.py # Cross-meeting accountability ledger & delta tracking
│ ├── summarize.py # Executive summary (LLM or extractive fallback)
│ ├── report.py # Minutes, follow-up, ledger & accountability rendering
│ └── datagen.py # Synthetic meeting series generator
└── tests/
Built by Nathaniel Gordon