Building LogiqMoM: Lessons From Summarizing Multi-Hour Client Meetings

When we set out to build LogiqMoM, our internal meeting intelligence tool for consultants, we thought transcription and summarization would be trivial. We quickly hit severe production limitations with off-the-shelf tools.

The Problem We Hit

In 60-to-90 minute consulting workshops, critical commitments and decisions are often decided in the first twenty minutes. Standard LLM context window summaries consistently diluted early decisions, over-indexing on whatever was spoken during the closing ten minutes.

What We Tried

We initially experimented with feeding the entire raw transcript into large context models. While the text fit inside token limits, hallucinated action items occurred whenever two speakers discussed hypothetical ideas that were rejected later in the call.

What Worked

We redesigned the pipeline into three distinct stages:

1. Acoustic Diarization & Topic Splitting: We segment the transcript into distinct agenda blocks based on speaker cadence and topic shifts.
2. Entity-Action Extraction: Each block is evaluated individually for explicit commitments, extracting the task, assigned owner, and target deadline into structured JSON.
3. The Edit-Before-Share Gate: Rather than emailing minutes automatically, we stage the synthesized output in a quick 60-second review screen where the meeting lead confirms or edits owners before dispatch.

What We Would Do Differently

We would have built the evaluation dataset on audio containing cross-talk and background coffee shop noise from day one. Acoustic diarization degrades rapidly with room reverb unless audio pre-filtering is tuned specifically for laptop microphones.