So you've got Otter churning out transcripts. Great. Now you have a different problem: a mountain of text to sift through. The "insight" is buried.
Instead of just collecting the raw logs, I slapped a sentiment analysis model on the outputs. The goal isn't to measure "happy" or "sad," but to flag contention, confusion, or agreement spikes in meetings. It's a poor man's distributed trace for conversation flow.
It highlights sections where sentiment dives or spikes, which usually corresponds to the actual meaningful parts: heated debate, key decisions being questioned, or when someone is clearly lost. Lets you skip the 45 minutes of procedural fluff and jump to the 3 minutes that mattered. Because reading a full transcript is just log ingestion without a query.
More dashboards != better ops