Originally, all of our datalogging was a one-way transmission from the roaster to the computer. This allowed us to produce a graph. Later, we established two-way communication with the PLC in a way that allowed full control of the roaster. We could set temperature and had full control over an electronic modulating gas valve and the VFDs that controlled drum and fan speed. This led to operator assist functions like cooling tray stirring and cooling fan triggers as well as retained charge temp and automated fan, fuel and drum speed settings for automated between batch recovery.

Roast Replay was originally conceived as a more conventional PID algorithm-based system that would automatically modulate settings relative to established time target events of green/yellow, first crack and discharge.

The challenge with that type of automation is the equation it is based on can't account for coffee taking heat differently all throughout the roast. It off-gasses enough moisture from charge to green/yellow to severely retard heat absorption and the backwards looking modulation equation is very much different relative to the ROR data during the mid-phase of the roast when the Maillard reactions in the outer seed off-gas a tremendous amount of CO2. The also reduces heat absorption, but to a lesser extent than the initial H2O.

This is to say that the ROR feedback loop that the PID algorithm generates during the "drying" phase of the roast causes the system to over correct temperature readings during the mid-phase. Worse, you have doubtlessly witnessed many times the roast going "exothermic" about 10-12F prior to first crack. What's really happening is the outer seed temp is high enough that it no longer supports Maillard reaction, off-gassing effectively ceases, and the coffee takes on heat much more efficiently.

Conventional PID automation invariably overcorrects leading to an excess of drunk driving along the roast profile.

FYI: if you don't want an operator to see this in their profile, you use a big fat sloppy thermocouple and position it closer to the centerline of the drum. The graph looks smooth, but the coffee is still being thermally whiplashed all the way to finish temp with a poorer cup as a result. At least until you get used to it and it becomes your "normal."

As it turned out, our roasters are mechanically stable enough to simply emulate the operator. The system does require more precise pre-heat and charge habits. It's also a good idea to try to ensure that charge weight and initial temp are as close to the original roast as possible. Beyond that, when necessary, manual override results in a new profile that can be replayed for the rest of the production day too to account for variations in the coffee or the environment.

In a sense, our system relies more on real intelligence. In its current form, the system is significantly more precise and marginally more useful than alternatives from other manufacturers.

We are not Luddites. We are looking at ways to improve Roast Replay and to automate manual roasters and roasters from other manufacturers. AI may end up being a part of that.

It is worth noting too that for as useful as current AI tools have become, much of the "buzz" around AI is a product of a gargantuan tax dodge. Companies are copy and pasting AI into every description possible and reaping a 25% Federal AI R&D tax credit. When you see Tesla claiming a $10 billion AI R&D charge, it's marginally less about actual money spent on AI more about labeling everything possible "AI" to book the $2.5 billion tax credit.

That's just the K Street lobbyists in DC. Lobbyists have also taken the AI tax dodge roadshow to states too. As near as I can tell, 36 states have also carved out AI related tax credits.

No joke: I literally sent a note to our project manager last week to see what we can do to legally squeeze into a place at the AI tax trough.

This is embarrassing enough to make me a little queasy, but true.

We have a very ambitious development schedule planned for this year. Stay tuned.

Steve