Build Logs

Stop renting your video pipeline

Per-minute AI clip tools make sense until you publish weekly. Then the math flips, and owning the pipeline wins on cost, quality, and control. Here's the math from my own show.

Every week I turn one long livestream into a week of short-form video: transcribed, segmented, clipped, cropped vertical, captioned, thumbnailed, and staged for posting. For a while, tools that do pieces of this for a per-minute fee were the obvious answer. They are fast to start with and genuinely decent.

Then you publish every single week, and three problems compound.

The three problems

The meter never stops. Per-minute pricing on a weekly show is a subscription that scales with your success. And the pricing page you signed up under is not a contract; it quietly changes.
Generic tools don't know your content. Cloud clip pickers optimize for generic virality. They do not know which local names matter, which segment your audience actually shows up for, or that the school board discussion is the story this week even though nobody yelled.
You own nothing. Cancel, and your workflow evaporates. Your transcripts, your corrections, your archive of processed episodes: theirs, or gone.

What owning it looks like

My pipeline runs on hardware I already own. Transcription runs locally on a GPU, so an episode costs electricity, not an API bill. A corrections dictionary fixes the local names the model gets wrong, once, and remembers forever. An LLM reads the transcript and proposes clips; I approve them. The pipeline crops, captions, renders branded thumbnails, and stages everything for posting. Roughly a hundred episodes have been through it, and every one made the archive searchable instead of just published.

The stack is Python, FFmpeg, a local Whisper model, and an LLM for clip discovery. Nothing exotic. The value is not in any component; it is in the plumbing between them, and plumbing is a one-time cost.

When renting is still right

If you post once a month, rent. The setup time will not pay back. The crossover is frequency: the moment content is a weekly obligation, every per-unit cost becomes a tax on consistency, and consistency is the entire game.

I wrote up the whole pipeline, including where it breaks and what I still do by hand, as a free guide. Get it here.

Six stacked translucent network layers with a single light beam passing through all of them

Six vendors, one question: tracing a call across systems that don't know each other

A phone problem can live in any of six layers owned by five different vendors. This is how the ops stack from the earlier case study actually works: a diagnostic ladder that walks the whole call path from one conversation.

Investigation wall with a county map and glowing threads connecting pins between towns

The monitoring engine where a new market is a config file

A watch-everything intelligence engine for one geographic market: public sources in, deduplicated and relevance-scored items out, a cited morning digest by 7 a.m. Adding a second market took a JSON file and zero code changes.

Macro of a RAM stick with one chip glowing amber among teal ones

My backup system found a bad stick of RAM

A one-person operation cannot afford a two-week rebuild after a dead machine. So both machines rebuild from one encrypted off-site copy, one script, one passphrase. Then the backups started failing on exactly one machine, and that was the point.

A desk clock dissolving into streams of data that reassemble into an orderly ledger

I stopped logging my time. The evidence does it now.

Every consultant hates timesheets, and every consultant leaks billable hours to bad memory. This system reconstructs the workday from evidence the tools already emit, and a human approves every entry before it becomes money.