Every small business has that task — the repetitive, time-eating chore that has nothing to do with why you started, but somehow swallows an afternoon every week.
For a volleyball coach I work with — Kha, who runs an online coaching business — it was video. Players send him film, he records drills and breakdowns, and every single clip had to be hand-renamed, dropped into the right Google Drive folder, linked, and logged into a tracking sheet for his editing team. Eight to ten videos a week, every week. Pure file admin standing between him and actually coaching.
So I built him a tool that does all of it the moment he drops a video into a folder. Here's how it works.
The Manual Loop
Kha's old process was five steps, every clip: connect the drive, rename the file, create a dated subfolder in Drive, upload the clip and copy the link, then paste that link and a title into the sheet his editors watch. None of it is hard. All of it is friction — and friction multiplied by ten videos a week is a part-time job you never agreed to.
The goal wasn't to replace his system. His editors already knew it. The goal was to make the system run itself.
The Architecture
The tool is a quiet background watcher on his Mac. Each stage has one job and hands off to the next through a clean interface — which means I can swap the AI model or the storage layer without touching anything else.
Watching the folder
Kha drops a clip into a single inbox folder — that's the only thing he ever touches. The watcher notices the new file, but it waits until the file has finished writing before doing anything (a half-uploaded recording is worse than no recording). It also remembers what it has already processed, so a reconnected drive never doubles up.
Understanding the video
This is where it earns its keep. Instead of a generic model, I taught it volleyball — positions, skills, shot types, set tempo, even how to read a scoreboard. It watches the actual footage (and listens to Kha's voiceover when he narrates) and returns a clean, structured read of the clip: what skill, what situation, and a short, human title.
The single most important design decision was teaching it when to say "I don't know." On a focused coaching clip where Kha is talking through one player, it can confidently name the position. On a wide game clip with two full teams and no narration, even the best AI genuinely can't tell which player is which — so instead of guessing wrong, it steps back and describes only what it can actually see.
if scene == "single_focus": # one player, usually narrated title = f"{position} {skill} {situation}" # "Outside hitter block, set point" else: # wide game clip — too ambiguous title = f"{skill} {situation}" # "Game point block" — no guessed position
Accuracy over impressiveness. A confidently wrong label is worse than an honest, general one — and it's the difference between a flashy demo and a tool someone trusts every single day.
Filing it into his system — not a new one
Here's the part that made the whole thing land: the tool doesn't invent its own folders or its own sheet. It does exactly what Kha did by hand. It creates the same dated subfolder under his existing Drive folder, uploads the clip inside, grabs the folder link, and appends a row to the same tracking sheet his editors already open every day.
His editors didn't have to learn a thing. From their side, the rows just started appearing on their own.
What Makes It Actually Work
The flow above is the easy 80%. The last 20% — the part that decides whether a tool gets trusted or quietly abandoned — is in the details:
- It fits the existing workflow. No new app for the editors, no migration, no retraining. The tool adapted to the humans, not the other way around.
- He owns his own data. It signs in as Kha through Google, so every file lands in his Drive, under his account — not some service we control.
- Nothing is ever lost. Originals are kept, every clip is backed up to Drive, and a file the AI can't process is quarantined and flagged — never silently dropped.
- It runs itself. It starts in the background every time he opens his laptop. He never launches it. He just drops clips.
The Result
A weekly chunk of file admin disappeared. His editors kept the exact workflow they already knew. And Kha got his time back for the work that actually grows the business — coaching and content.
The hardest part of building this wasn't the AI. It was two things that have nothing to do with models: making the tool humble enough to admit what it can't see, and respectful enough to slot into a system the humans already trusted. Get those right, and the automation just... disappears into the background. Which is exactly where good automation belongs.