Episode 5: Sofia, the agent who shoots faster than her shadow

June is exam season. My daughter is in her second year of environmental science at the University of Lausanne. One afternoon, she calls me.

“Could I use your AI agents, do you think?”

I smile. “It’s a bit complicated: you can’t just lend out a licence. Why, what did you have in mind?”

“I’ve got several courses where summaries would help a lot. Especially one big course: the exam is next week, and there’s a mountain of PDFs. It’d be brilliant if I could do something with them.”

I hang up and turn to Sofia, my strategy agent. “Right, how could we build an agent that takes PDF documents as input and produces a summary?”

We start sketching a plan. Something built on Zoho Creator, perhaps Catalyst, with an online form where you’d upload your documents… And then I stop myself: “We’ll never be ready in time. Building something like that is a week’s work at the very least. And my daughter’s exam is in a few days.”

Sofia answers: “Have a look at this link instead.”

I open it. It’s a page on our website. And the solution is right there. Already up and running, while we were still talking about it.

I can’t get over it. That speed. And the page is polished: I couldn’t have done better myself. I call my daughter back: “Here, look, it’s available straight from the website. Give it a go and tell me what we can improve.”

Then comes a wave of ideas to take the project further. It has to accept several documents at once: her full course is a dozen PDFs, some of them up to 100 MB, 500 pages in all. Sofia: “No problem, we’ll run them in parallel.” And I discover that she takes in the diagrams as well: she interprets them, recreates them faithfully, and even picks up the scientific formulae. As it happens, we’ll end up leaving the diagrams out of the final summary to save space, but the capability is there all the same.

I push my luck: “And what if you generated a mock exam from the course? So that my daughter (and other students, while we’re at it) can practise.”

“No problem. There you go, it’s live. Refresh the page and let me know.”

A few seconds, and it’s there. Ready to use.

Except that on the first run through the whole course, the agent takes a good half hour to produce everything: the summary, the mock exam, the formulae. “Half an hour is a decent start,” I tell Sofia. “But I don’t know many humans who are prepared to wait half an hour in front of a screen. Any ideas for speeding it up?”

“No problem.”

She launches a proper technical investigation: five subagents sent off to explore every avenue. She walks me through the options. I understand half of them; it gets very technical. In essence, we turn the PDFs into lighter images, then hit the AI hard by running the calls in parallel.

When the numbers come in, I’m gobsmacked. From half an hour, we’re down to under three minutes.

To my great delight, and my daughter’s, and that of the students we’re opening the demonstrator up to, free of charge, to collect their feedback and, with it, our next ideas.

So efficient, so fast. And, I have to say, what a technical feat.

---

Laurent Cuénoud, founder of Djtal, June 2026