Case study
Tandem Social (AI Audio Network)
An AI-driven social network for live audio shows—Spaces-style sessions with Tandem-managed automation, transcripts, and instant post-show assets.
Problem
- Live audio requires low-latency rooms, presence, and reliable speaker management.
- Publishable artifacts are essential: transcripts, summaries, chapter markers, and highlights.
- Automation needs orchestration: schedules, retries, and run history across shows/debates.
- Operational resilience is required: fallbacks when orchestration is unavailable.
Solution
- Backend integrates the Tandem Engine through a dedicated wrapper around the tandem-client SDK.
- Tandem stores automation definitions, schedules, retries, run history, and cancellation state for managed content.
- For shows and debates, the backend supplies a spec that instructs Tandem to call internal callback endpoints.
- Callbacks land on internal run endpoints where dedup by tandem_run_id occurs, run state is recorded, and generation starts.
- Generation uses Tandem Social services: LLM calls through LiteLLM/OpenRouter, local Piper TTS, media writes, and DB updates.
- Admin APIs expose sync, run-now, list/inspect runs, and cancel flows over the Tandem-managed schedules.
- If Tandem is unavailable for manual runs, the backend can fall back to local generation for supported paths.
Architecture
Results
Orchestrated automation
Schedules, retries, and run history managed via Tandem with internal callbacks
Publishable artifacts
Transcripts, summaries, and chapters generated and stored with replays
Resilience
Fallback to local generation for manual runs when orchestration is unavailable
How this maps to client use cases
- Communities host live sessions and publish transcripts with chaptered summaries for async consumption.
- Newsrooms run live audio briefings and auto-publish concise recaps across channels.
- Creators run AMAs with AI-assisted production and instant post-show show notes/highlights.