Calendar orchestration in the backend
Dispatch was moved out of the recorder and into the service, so one central component decides which meetings are captured across the organisation. Capture no longer depends on any individual remembering to start it.
Speaker-attributed transcription
Diarization separates speakers on GPU, turning a continuous transcript into an attributable record that can be skimmed by who spoke rather than read end to end.
Semantic search across the archive
A multilingual embedding model indexes every record into pgvector, so the archive is searched by meaning rather than keyword — including across meetings held in different languages.
Conversational retrieval over meetings
Questions are answered from the record set itself, so a decision taken three months ago can be recovered without knowing which meeting it was taken in.
Corporate single sign-on
Authentication runs through Microsoft Entra with authorisation code and PKCE, restricted to the organisation's tenant. Signing in is itself the proof of membership; there is no separate user directory to maintain.
Isolated from the capture infrastructure
The platform runs on its own host, separate from the media-capture bot, reaching the recorder and the inference services over the internal network. Fragile capture infrastructure stays isolated from the service people actually use.
Self-contained deployment
Database, API, web client, reverse proxy and embedding service ship as a Docker Compose stack, so the whole platform runs inside the organisation's own network.