PRISM Agentic AI
They hired me for a landing page, then handed me the product: a private video library, 50-odd customers, that they wanted to turn into something that answers.
Confidential · industrial software

What a factory worker opens. Before this project it was a video library — you searched, you pressed play, you hoped.
A private YouTube, and not much more
A pump fails. A technician finds the video and presses play. Four minutes in, a tool they don't have, so they walk off to fetch it. At minute nine, a warning that needed reading first.
Stop making people watch a video to get an answer. Let them ask, and land the answer on the step it came from.
Two modes, two kinds of people
Editor Mode makes a procedure; Viewer Mode uses it. The person who knows the job records it once; everyone else asks later, in another language.
Four conversations changed the brief
I talked to four existing users. Nobody asked for better search. They wanted tools, materials and warnings — none of it treated as information.
What it looks like when it works
Ask what's actually happening, not what it's called
Nobody knows the document's title; they know the machine is making a noise. Matches come back with a score and a line on why.

The answer lands on a step, with a timestamp
It explains the bubbling, then points at Step 3 · 3:20. Go back to Step 2 is a button.

Tools and materials before you start, not at minute four
Their own tab. What do I need in my hand?, answered without playing anything.

Where a procedure comes from
An expert uploads a recording, AI drafts it, a human confirms every field, a named approver signs off. The old product had none of this.


Dubbing, translation and captions sit in an optional phase before approval.
Released to a test group first
Not all 50 customers at once. Their feedback drove a final round of changes before it replaced the old product.
The product is real; every name and figure here is demo content.
Want the reasoning behind each decision, and what I turned down?