Using AI for Bean Selection and Roast Graph Analysis

Roasting on the Bullet
1

Over time, I’ve developed a fairly systematic way of using AI as part of my roasting workflow, and it has significantly improved my results. My own focus is the Vienna profile, so that is the framework these tools are built around, but the same basic approach could be adapted to whatever roast profile you prefer.

I’m using two separate Grok projects. One is dedicated to bean selection, and the other is dedicated to roast analysis. The bean-selection project is built to research a coffee lot deeply before I ever buy it. It looks at origin, altitude, variety, processing method, density, moisture, chaff tendencies, flavor evolution from lighter roasts into Vienna, freshness window, and any existing Bullet profile experience it can find. It also weighs how different processing methods tend to hold up at darker roast levels. Each bean is then rated on a 1 to 7 scale for its suitability for Vienna.

The second project is focused on the roast graph itself. That one is built around my Vienna workflow and evaluates the curve in a structured way: charge and early phases, mid-roast behavior, first-crack phase, post-first-crack development, likely strengths, likely defects, and what to change next time. It also helps me stay disciplined about preheat choice, momentum, fan usage, DTR, and second-crack approach instead of just roasting by feel and hoping the cup turns out. In addition, all of my final, tweaked recipes are stored in the source files and analyzed for possible application to a new bean recipe, or for an AI-produced hybrid of several recipes. That gives me a very strong starting point when I am approaching a bean I’ve never roasted before.

What has impressed me most is not that AI “knows coffee,” but that it helps organize and apply a lot of roasting logic consistently. It forces me to think more clearly about bean suitability before I buy, and it gives me a more objective post-roast review than I would get from memory alone. In my case, that has translated into better bean choices, fewer wasted roasts, cleaner Vienna development, and more consistent results in the cup.

I’m not using AI as a substitute for experience or tasting. The cup still has the final word. But I am using it as a research assistant and roast analyst, and for me it has been remarkably effective. It has made my decision-making more systematic and has improved both my confidence and my outcomes.

For anyone roasting on the Bullet, especially if you keep detailed notes and roast for a specific target like Vienna, AI can be a very useful tool. The key is giving it strong instructions, good roast data, and enough context to evaluate the bean and the graph in a way that matches your own roasting goals.

4 Likes
2

"A.i."llio Bullet uses some more primitive or machine learning tools. do you use them much? how does Grok help over what Aillio provides? Aillio is collecting a lot of data and should be able to provide some intereating insights.

3

That’s true, and Aillio does provide some useful data-driven and limited machine-learning tools inside RoasTime, especially things like Smart Prediction, logging, playback, and recipe automation. But it’s still basically a closed system operating within the Aillio/RoasTime/Roast.World environment. It is not doing open-ended research across the internet the way a true AI tool like Grok can. That is a major difference. What I’m using Grok for goes beyond curve prediction or playback. I’m using it to research beans before purchase, compare prior successful recipes, generate hybrid starting-point recipes for new beans, and analyze roast graphs at a broader reasoning level.

4

@carldebar.JZiZ
Why Grok and not other AI model? Have you tried using other AI model?

5

I’ve used both Grok and ChatGPT, but Grok has proven to be the most reliable and accurate - so that’s where my projects reside.

2 Likes
6

You piqued my interest, Carl ! I have been enjoying using Anthropic’ AI for a couple of months now. It is an incredible office assistant. I find myself thanking and saying please to it!! I’m afraid one day I will thank our toaster for doing its job!! :scream:
Could you be explicit how you feed data to Grok? Does it get the graph output and go from there? Thanks for the idea. JP

7

@patrick9724, on the roast-analysis side I keep it very simple: I just upload a screenshot of the roast graph into Grok and let it work from that. I’m not feeding it raw exported data or doing anything complicated.

What makes that useful is that the project already has a fairly detailed instruction set behind it. So when it sees the graph, it is not just giving generic comments. It is looking at things like charge recovery, whether the roast carried momentum properly through the middle, where first crack landed relative to time and temperature, how the post-first-crack section behaved, whether the curve looks like it is drifting toward stall, bake, or sprint, and whether the finish was appropriate for the Vienna target. It is also working from specific Bullet rules on preheat branches, charge approach, power and fan strategy, DTR, and expected second-crack window.

I also use a simple stethoscope mod on the Bullet that lets me hear crack activity much more clearly than I could otherwise. That gives me a high level of confidence in when I mark FC, FCE, and SC, which in turn makes DTR measurement much more accurate. Since DTR is such a crucial factor in the final outcome, that matters a great deal. Check out my mod at Stethoscope Mod for Clear First & Second Crack Detection.

The other piece is that my final, tweaked recipes are already stored in the source material. So Grok is not only reading the graph in front of it, but also comparing it against roast patterns that have already worked well for me.

The project has been a lot of fun and has added a whole new dimension to roasting. But more than that, it has made a marked improvement in the cup. I get a lot of repeat customers, and the unsolicited comments have been that my coffee is the best they can get. :grin:

2 Likes