I used to think that too (I am also a computer guy). But it is very difficult to have every layer of the stack including hardware be verifiable. And while paper is a pain in the ass it's also just a lot more difficult to hack. Many of Jovan's methods rely on analyzing the unintentional corroborating redundancies like paper folds or scribble styles, that are virtually impossible to make line up when you're faking it. Digital techniques are always about eliminating unintentional redundancy, by representing artifacts in a more succinct digital format. Sure you can add in digital signatures etc but there are always ways around these things through the "real world" - like how social engineering, not cryptography, is usually the weak point in a digital security architecture.

I'm thinking hand marked paper ballots, kept in a sealed box with multiple observers. Why try to optimize this process further?

For the counting, I don't see why hand counting is out of the question, if it's done in a distributed fashion. An initial hand count at the precinct level followed by a machine scan and publish the scans and tallies.

All candidates can bring in their own machines to examine the paper, or download and analyze the scans, and reconcile differences.

I don't have it all worked out, but any black box is a huge risk and keeping the paper is the only reason we have any hope of sorting out 2020. (Even so in 2020 there are a lot of "ballot marking devices" in use, and who knows if they did what the voter intended. Those have got to go except for people who legit need assistance)