Run Kashi on a real transcript

Paste a Zoom, Teams, or Meet transcript. Single-meeting, structural-only demo. The transcript text is never stored — only an optional metadata snapshot you can choose to save.

Demo analysis · no account required

Your transcript is processed in memory only and is not storedin Kashi's database.
This is not production tenant data and not a finding about any real person.
Do not paste confidential real-company data unless your organization's policy permits it.
Output is illustrative only — single-meeting demo cannot produce a longitudinal pattern.

Upload .vtt / .txt

Exploratory text pass (optional)
Separate from the structural result. If enabled, transcript text is sent to the Anthropic API in-memory for this request only — never written to the database, never logged, never returned in the response. Results from this pass are labelled as exploratory / lower-confidence and do not change the primary structural result. Off = structural-only (default).

Stateless — transcript is analyzed in a single request and discarded.

How this demo works

The same detectors that power the Personal Reflection + Executive Brief run server-side, return structural signals, and throw the transcript away. No authentication. The transcript text is never stored — you can optionally save a metadata-only snapshot (result state, counts, meeting type), never the transcript itself. No content sent to any third-party model unless you explicitly opt in to content analysis.

How production differs from this demo

This demo accepts a paste. In a real deployment, Zoom / Teams / Meet webhooks fire when the platform finishes transcribing; Kashi pulls the VTT via the platform API, runs the same detectors, and stores structural metrics only (no text body). Production also runs a rolling per-speaker baseline over 5+ comparable meetings, which enables longitudinal readings that a single paste cannot produce. See /technology.html §4 for the webhook + normaliser architecture.

What this demo computes (and what it doesn't)

Computed from this single transcript

Who got the floor and for how long — speaking-time balance across participants (Schmid Mast 2002)
Who got cut off — interruption timing by speaker pair (Anderson & Leaper 1998)
Questions that received no reply, and proposals that went unpicked up (Sacks / Schegloff / Jefferson 1974)
One-way agreement pressure — whether disagreement was distributed evenly (Franzen & Mader 2023)

Not computed from a single meeting

Per-speaker participation drop vs their own baseline, and pattern persistence across weeks. Those require 5+ meetings and a rolling baseline — the deployed pipeline supplies both via webhook ingest on a tenant-opt-in basis.