CoHost AI Studio

The problem

Independent podcast producers run every episode through five to eight disconnected tools, with a manual handoff at each step and no signal when a bad master or a thin transcript slips through to publish.

How it’s built

1. A DAG pipeline with idempotency and fault isolation

   Twenty-three steps run as a topologically sorted DAG on a thread pool. Each step has an idempotency predicate, so partial runs resume cleanly, and per-step cascade behavior isolates a failure (a stitch timeout does not block show-notes generation). The graph is validated for cycles at build time, before the first episode runs.

2. An instrumented quality gate, not a vibe check

   Eleven steps are scored against programmatic metrics: LUFS delta for audio mastering, word confidence and speaker resolution for transcripts, artifact completeness for video. A weighted composite plus a per-step hard-fail threshold feeds straight into the publish decision; below the bar, the pipeline halts before distribution.

3. LLM evals layered over deterministic scores

   Five content-heavy steps add opt-in LLM evaluation via OpenRouter, each returning a criterion, score, and reasoning that merge into the scorecard alongside the programmatic metrics. Fast deterministic checks always run; the slower qualitative ones are opt-in.

By the numbers