This one covers October 1 through 8.

An AI video editor that runs on my own machine

This is the one I’m proudest of. We built an AI video editor that takes a raw podcast recording and hands back a finished episode: the long-form cut, four short clips, captions, episode metadata, a QA report and a thumbnail. No editing app. No cloud service. The recordings never leave my computer.

It shipped three full episodes in its first two days, end to end. Watching a raw file come out the other side as something I’d actually post made me grin.

If you run a podcast for your business, you know editing one episode means a freelancer or a few hours of somebody’s week.

It’s a v1, and it acts like one. A human still picks the titles and listens to the final cut. One caption file came out about half a second early. The sync check caught it and I re-timed it. From here on, every correction I make becomes a permanent QA rule, so the same mistake can’t ship twice.

Behind it sit 253 automated tests. The QA gates catch things like black frames, frozen video, caption drift and graphics carrying the wrong brand.

Friday builds software while I sleep

Each night Friday (my chief of staff) picks one well-defined task off its own to-do list, builds it in an isolated sandbox copy of the code, and leaves a reviewed result for me in the morning. Nothing merges until I approve it. My morning job is a yes or a no.

Small fixes used to pile up behind me. They get done overnight now.

Friday only touches tasks I’ve tagged and scoped ahead of time, in a short list of my own repos. Every run has a hard budget, and a watchdog kills any run that goes over it. Picking the night’s task used to cost an AI call every time. It’s plain code now, zero tokens.

The first night didn’t go clean. The morning review got rejected on a formatting technicality, so we fixed that. The first build that counted, a podcast-thumbnail improvement, made it to release a day later. Friday’s test suite ended the week passing about 1,800 tests.

One scoreboard for every agent

Once you’re running more than two or three agents, you can’t tell which ones are earning their keep.

So every agent I run reports the same standard numbers: what it did, what it cost in tokens, and what broke. Those roll up into one “fleet output” table in my Monday brief. My assistant and sales agents also run a nightly self-audit and send the findings to Friday.

We built it in eight test-driven slices in one day. All eight merged.

Two agents on my other machine show up as “not collected” until they send the new report format. The brief says so instead of hiding the gap.

My marketing engine shipped eight releases in a week

This is the AI system that plans, writes, checks and publishes weekly content and client reports for the agency I run and for my own businesses. Eight releases went out October 1 through 7. The ones worth mentioning:

• An install checker that tells a fresh install exactly what’s missing and how to fix it.
• Fact-grounded client reviews. Every number in a talking track has to trace back to actual data, and anything we couldn’t collect gets labeled that way instead of showing up as a zero.
• Monthly reports that classify every call from its transcript. A silent call is spam. A call that connected with no recording counts as missed.
• A test bench that scores how well a local model tells leads from spam before we trust it.

Most of the week went into making the numbers provable. A report that’s confidently wrong is worse than no report at all. I’d rather hand someone a gap marked “we couldn’t get this” than a zero that looks like a fact.

A new site for AI Builders Labs

Smaller one. The AI Builders Labs site got a rebuilt version 3, and its blog now runs on the same content engine as my other sites. Four seed posts are up. My part each week is one approval.

What broke: two sites went quiet and nobody noticed

One of my own business sites went about ten days without a new post. The content was written. It was approved. It never published. A directory site I run has gone dark the same way twice in a month.

No alarm went off, because nothing had failed. Things just stopped.

That’s the failure I worry about most with automation. A crash gets your attention. Silence doesn’t.

So we built a daily “gone-dark” alarm that covers every site the engine publishes to. If a site’s queue runs dry or a scheduled post doesn’t go live, I hear about it that morning instead of weeks later.

Next week I’ll write up a blind test where the prompt I thought was better lost, 4 to 2.