Client build · Jack Hayford Ministries
A 1,500-item media backlog, processed hands-off
A ministry had ~1,500 archived recordings to process, at about 75 minutes of hands-on work each.
Where the 75 minutes went
The client had years of archived audio they wanted searchable and republished. The process was entirely manual: listen, transcribe, summarize, note the passages referenced, normalize the audio, add an intro, update the spreadsheet, upload. Roughly 75 minutes per recording, with about 1,500 waiting.
At that rate the backlog wasn't a project, it was a full-time job for a year. The constraint wasn't cost. It was that a person had to sit with each file start to finish.
Turning review decisions into a queue
Most steps only needed a person because nobody had automated them. The genuine judgment calls were review decisions, and reviewing is fast when the draft already exists.
- Timed each stage of the manual process to find where the 75 minutes went.
- Chained a transcription API into the Claude API for summaries and passage tagging, so the two slowest steps became a draft to check rather than a page to write.
- Automated the work around them: file intake, audio post-processing, spreadsheet updates, output handling.
- Added batch support, so the operator's job became queueing and reviewing.
- Containerized it and deployed to Railway so it runs as a service, not on someone's laptop.
Outcome
Per-recording time dropped from about 75 minutes of hands-on work to roughly 10 minutes unattended. The pipeline is deployed and maintained in production.
The number I care about isn't the 65 minutes saved. It's that a task needing someone's full attention became a queue someone checks — the difference between a backlog that gets worked and one that gets abandoned.
The harder part was not automating the review step as well. The client's credibility depends on the summaries being right, so a person still reads them. They're just reading instead of typing.