Robert Söesemann, who joins to share about Salesforce & AI from the trenches, is a Salesforce principle architect, and Salesforce MVP.
Main Points
- What happened since our earlier podcast episodes
- Robert’s extensive AI experimentation journey
- Lessons Learned and Next Steps
- Robert’s take on Native AI features by Salesforce
- Thoughts on AI copilots for programming
- How Robert leverage AI in his job and private life
- What’s new and noteworthy in PMD
Links
- Robert’s LinkedIn
- Aquiva Labs
- Robert’s Salesforceben post: Building 3 Salesforce AI Apps
- Robert’s video talk on Custom PMD rules
- Robert’s own GPT: Robert on Software Craftsmanship
- Robert’s previous episodes: Clean code and PMD
Video Teaser
Contact Me
If you enjoy this podcast, please reach me out to let me know your thoughts or simply say Hello. My (Xi’s) contact info:
Take care, until next time :)!
Transcript
AI-generated, unedited.
[00:13] Hey everybody! This is Xi Xiao. This is a new episode of Salesforce Way podcast. It's been a long while, and today I have a returning guest coming back with me. His name is Robert Söesemann. Hello, Robert. Hey Xi, thanks for having me again. It's been a while. It's been a long while. Back to the podcast. Do you remember when was the last time? I think it was one of your first podcasts. I remember you and I—we immediately fell in love with each other because you cared so much about good code quality, clean code, and automatic code checking. Yeah, we had two episodes: the Clean Coder and the PMD. Right. That was okay. Cool. And eight. Cool. But it's been couple of years already. So what are you up to nowadays? Probably boring answer, like everybody else. Like AI? No. Yeah. So okay, I'm still in the space where I was when we when we talked. I'm still doing Salesforce. I'm still. Doing, AppExchange development most of my time. So I'm not. I have changed jobs at the moment. I'm a principal architect at a at a huge Salesforce partner, Akiva Labs. As I partner, PDO partner, but I but I still do what I do, like hearing about product development on the Salesforce platform. So that's that's my main
[01:43] job. And as a Kiva is really a great and innovative company, I I spend like the last year or a bit more. When when was GPT four when when did it came out? Like over a year ago, I'm doing a lot of like. AI experimentation. You know, how can we use it internally? How can we make make use of AI for our customers? And yeah, and when I say experimentation, I really mean like internal projects with people who are not in projects. So that's that's. That's really like the the the cool thing of my last year that I had the chance to actually lead such such projects. We we open source stuff and yeah. I will tell you. So I guess you changed the company. The company now is Aquiva Labs, right? Yes. The company wants to invest a bit in the AI side to do some experimentations, and you also personally are passionate about this. Exactly. So our like management and our founders. Immediately, you know, you you probably remember when GPT four came out. It was absolutely crazy. Like everybody was in the middle of, ah, this is shit. It doesn't work, or it's it will like. Kill our jobs, destroy humankind, and we were in
[03:13] the middle of this. And thought, is there will it like destroy the company or help us a lot? And so, we thought to actually find this out, we we have to make a concerted effort. So we got together like a bunch of people, and we we we created a temporary AI task force. And we thought, you know, we have to do many things. We we need to think about how we can use it in the company, you know, like ethics guidelines, guidelines for people on how to to use those new tools and when to use them. But we also said we need to actually build something, you know, build small apps. We need to use it in the space we actually play all the time, like. Creating source apps or helping customers, and yeah, and then then we basically created multiple apps. We even released them on the App Exchange or made them open source, and yeah, that was very cool because it it was a very different kind of. How you make projects? You know, no normal projects. They need a longer planning phase. They have user teams with Q and A and product owners, and you know, many of our projects are really long and. And heavy when it comes to processes, and here we really just had like like me as an architect, two developers, sometimes maybe a quality engineer, but we were we we saw the experimentation, you know, small projects where you just
[04:44] basically do a proof of concept or an MVP. Is so great for learning because you know you stop when when you have enough. You just take the people who are currently we we say on the bench. You know, not an billable project. And I I really fall in love into this this way of like working. I mean it's luxury because it's often not paid projects. But I I think we also can learn from from those projects for maybe customer projects. You know sometimes customers have to pay a lot and define a lot. And maybe when you do really small and lean and pragmatic experimentation, you can. Yeah, good, good, good stuff done much faster. Yeah, that sounds exactly the same as programming. We learn programming from non-commercial product, open source, similar things. That this is why people do open source because you can just do it. You know, you you don't get approval for. You do it. You have time. You don't get money for it. So you you, it it it feels very different. And I really. You know, and this year, many many developers in Akiva and many people in Akiva worked in those internal projects, and especially developers said, "I loved it. I loved this way of working." You know, I spent ninety percent of my time programming. Communication was very lean. There was no meetings. There was no politics of bullshit and. You know, I'm I'm not
[06:14] naive. I know that those things need to be there, but you really see people smiling and being more happy when they do something and. You know, and my hope is that this is not like playing versus real world, but maybe we can we can actually learn on how to get more and more experimentation into. Definitely, I want to hear what you have learned. You know, pros and cons or different stuff. So those are really like the meat. But before going into those details, did you focus more on the Salesforce AI, or is it like? They are in the industry. Yeah, that's that's a very good question because you know when when we today say everybody's doing AI, know we are we're standing on the shoulders of giants. You know what we use in AI, even when we say we do development, it's just the good stuff which smart people, Capellan and stuff have done. You know, even my dad in the seventies has like his. Doctor Theitis or how is it called? He he even worked on neural networks. So I know that this is really going far back. But what what we did was AI for. Salesforce. I mean, we did AI. You know, we looked at AI that we can use for our own work. You know, copilot for development, for writing documents and specifications. But we also looked into how to put it into apps, into Salesforce
[07:44] apps. But the problem was, you know, a year ago there was nothing. There was no Einstein Studio, no model builder, no copilot. Salesforce had a lot of Nice announcement as always, but there was nothing. So what we had to do was basically build our own AI. On top of OpenAI, AWS, but make make use of it in Apex and Flow, you know. As I said, so we really build AppExchange packages. We really build Salesforce apps. So we did not do arbitrary Python development on Hugging Face, or you know, it was always how can we actually build something for Salesforce customers? Right. Well, we have those AI tools. How can we combine with the Salesforce business scenarios to help us? Right. Yeah. Exactly. And. And you know what I'm going to tell you. I also I'm also preparing, or I nearly have prepared my my conference talk in Berlin in May. There is the first Salesforce Community event, you know, and I I have a. As a session there, and I will talk in detail about what we built, what we learned with demos. So I, I even recently, not sure if you saw that on on Salesforce Bend, there was a quite lengthy article by me, which is basically also covering that, like
[09:14] what we. We learned. Yeah, I will put that article in our show notes for sure. Cool. Yeah. And now, if you may teach me what you have learned.
[09:28] Where to start? You know, there's there's many details. There's many details which which I cannot even summarize in a in a podcast. But the the bigger answer is. AI is hard, you know. AI is hard. We we Salesforce people, even people like you and me who do this for a while, our architect, we are in a quite easy state. You know, Salesforce is not the the big computer science stuff. Salesforce is really easy with low code, and a lot of pains are taken from us by the platform. And this is why we are mostly on the platform because we are more interested in actually getting business value than nerding around. And when you when you look at AI, especially custom AI, it's really hard. You have to look at all those places that we. Have not looked for for so many years as Salesforce developers. You know, I can just tell you that, you know, we we built one app which was really just native Salesforce app using, using ChatGPT API. And then we we struggled with the the size of the context window, you know, because we wanted to work on more text. And then we we looked into vector databases and rag retrieval augmented generation and then we had to learn about AWS and Google Cloud and so much stuff and this was so complicated and it was so different from from normal Salesforce development
[10:58] which is very restricted and expensive but you know how it works. And where it doesn't work, and and with, with all those more complicated platforms of platform technologies, it's really you feel like you are a student in the first grade, you know. So so this is my learning. Don't underestimate AI in general. It's it's it's still complicated. It's still like a, a ground where where the really smart people play. It's it's very different at Salesforce, and the second learning is, you know, in this year we we have built apps and most of those apps can actually be replaced by native Salesforce stuff now. You know, we you will see it in my article. I think at the end I'm talking about this. Salesforce has over the month announced things which will basically replace all this custom code. You know, the coolest latest app we have built is like it's called My Org Butler. It's a small utility bar. Lightning component in in in your Salesforce or, and there's a little Butler and you can ask questions. You know, what should I work on today? Which opportunity is is burning? Or you can say, yeah, create me a task or remind me in two weeks to do X Y Z. So really cool. And currently
[12:29] there's it's not replaceable totally by Salesforce, but Salesforce you know at TDX announced Copilots and Copilot Actions. And it's it's it's basically exactly that. So if if you if you are a customer who buys Copilot and has it in our org, I would recommend to not use custom stuff because it's so much simpler. And there are other things, you know, and one app. Ask your document. You know you can basically put in a five hundred page contract or specification in your Salesforce org, and then ask questions about this document, and do that. We really had to work on an AWS driven backend with a vector database and had to learn a lot. And Salesforce just said, "Oh yeah, Data Cloud will also have a vector database, and you will you can do semantic search and documents." You know, all of that will will have a price tag at Salesforce, but. One day it will just be there. Hmm. Robert, from what you just mentioned, I heard that your experiments, the three apps you built in AppExchange, those things were really concrete business cases because all other industries they they're using this, right? You upload the document you want, just. Get the AI to support you to answer questions around that document. It's been there already.
[13:59] We that that was that was the goal that we. What you know? In the beginning, we also put a lot of effort and people in it, and we thought if we build something, it must be usable for us in the company. There must be a chance that maybe customers might buy it, or at least we we can put it on the App Exchange for free and get some some marketing benefits. And yes, that that was the case. So when when we thought about what we built, we looked at. Business needs in a key way, and but it's also risk because all the other big companies are building the same thing. Salesforce also, sooner or later, want to build the same. Yes, right. And but it's happening. You are right, but we knew that, and we we revisited our decision to do that again and again. You know. The difference is, you can wait for Salesforce to just bring you, and then you use it and don't understand it. But we understand all that stuff, you know. I can like in Akiva. There's nobody who will say, "Oh, how does this copilot work?" or "What is a vector database?" or "Why doesn't work as, as I thought?" Because we we now have really a lot of people with. Quite deep understanding of of AI concepts. You know, to do that, we really had to watch online courses about all the stuff, and. And we think that that is that is a benefit
[15:29] in today's world. We are not only users; we actually understand what it is. You know, when you when you get a data lake from Salesforce and don't understand what a data lake is in comparison to a relational database, you might struggle with it. But and that's even more true with AI concepts. Right, right. Yeah, less sentence. Maybe you know, as a as a big company, when when customers come to Akiva and say we want some AI in our org or in our app, we probably recommend them to use native stuff. But if they say I don't want to do that because that's too expensive, we need to build it. And and and to do that, we actually had to do it for ourselves. Right, right. Well, it's really good here. I mean, as a software engineer, we always we know by heart only by hands-on to trial and error, see the things that are compiling, are config successfully. Then you know you really understand how things work down. Otherwise, if you just call in the API and get the answers, then it's a black box for you. Yeah, and and this is even more true in a in a hype technology, you know, because internet is full of oh, this will change everything. This is so disruptive, and you know, you have seen it. All those tools which replace us coders, replace the developer. It's it's. Most of that stuff is way too much. You know, it's just not true. And if you believe that, and customers
[16:59] will believe that. You know, business people they will just take it for granted. And if you, as a provider of Solutions or of software cannot really argument and say no. This actually doesn't doesn't work, or at least not as as good as you think. Then you are. Yeah, yeah. We are in Salesforce ecosystem. We know what marketing slogan means.
[17:25] Exactly. But I think it's everywhere. You know. Right, right, right. But on top of that, what is the current status of AI in Salesforce? Actually, I didn't really follow that closely. I know there were a lot of announcements. Is it already coming? Is it released? Yes. I. You know, Salesforce was always early jumping onto things. So many people who see those AI announcement, they say, "Hey, Robert, this is like always with Salesforce." You know, they announce something and then it takes hundred years to actually get. But my impression was Salesforce really. Change the way they work. You know, AI isn't new for them. You know, they in in 2016 I think they hired this German wonderkid, Richard Zohar, who who created those first Einstein things. You know, next best actions and sentiment AI. So I think they they have watched the space, and when GPT four came out, they really immediately announced stuff. And then also immediately delivered. You know, I remember like last Dreamforce, we immediately got got our hands dirty with this this trust layer. You know, you could basically talk to GPT, but in a safe way. And we did pilots for customers who presented at Dreamforce. Yeah, and and since then they they really
[18:55] bring things out. You know, most of the stuff it's it's in beta, and then it's there. You can buy it. So I really believe that Salesforce has has has speed up and. And they, I think, they understand that people will move away from Salesforce if if they don't deliver, you know, because Salesforce highly competes with everything AI because Salesforce is about. Do your business. The rest is simple. You know, no software. So, low code. You know, this—that's all the promise of. You can actually focus on what's relevant to customer, not on the the tech stuff. So, I believe that Salesforce really put all the energy in in making. Those AI things work, and when you look at the the release notes, you see things like Prompt Builder GA, Model Builder, Bring Your Own Model. You know, it it all feels a little bit crafty. I'm not sure if that's the right word. You know, when you look at it, it works. But some polishing, but at least it's there. So I think the the the scope, the features they give, it's it's very wide. So it's it's really a lot of stuff. Okay, it sounds to me is that definitely AI is in the high priority list item. Yes, absolutely. And also, it's wide. You mentioned maybe there are a lot of announcements.
[20:26] Maybe in the end, only a couple of them will survive. It's everywhere, you know. AI is important for all the apps, you know, for Salesforce. Yes. And what what I think Salesforce is in the low code market, which means. More people can build software. More people can do stuff, and I think AI is the new big low-code thing. You know, forget flow. Why should people even draw stuff? You know, we all know that flow is often worse or not much better than actually coding, but AI is the real low-code. So honestly, I think. Flow will actually go away as the number one concept that people have in mind when they think of Salesforce low code, and all those assistants, you know, all these generate me here, generate me there, you know, prompts, co-pilots, assistants. That's the new thing. In the future, you will not draw your business process. You will just, you know, dictate it. You will talk to your assistant and it will make it work. And it will be maybe as as bad as flow, but you know I'm I'm an old style programmer here. I think before ChatGPT was announced, I spent some time to learn the AI stuff. The The Python programming to build the we also got the Torch PyTorch or those. This is the book I bought like three years, three four years Python
[21:56] learning Python machine learning, yeah okay, yeah. Yeah, but for me, AI is like not in my personal habit. For example, when I'm programming, I know there is the copilots that can help you to build some boilerplate code, but I really never really have that habit of doing that. So, on your end, what do you think on those AI assistants for programmers? Okay, yeah. If I'm not jumping too fast to a different topic, no, absolutely. Um, I I was thinking, what should I tell you? Um, the first thing I wanted to tell you, please do it. Please don't think that you can get over it. You know. Like month ago, when everybody was saying we're we will all be replaced by AI, there was this wonderful saying: "You will not be replaced in your job by AI, but by somebody who's using AI." I really feel it. You will not be able to do your job, especially in Salesforce, when you say I don't care about this AI stuff. You know, you should not like a lambing just follow the copilot and let all your code generate for you. That's also dangerous because you will have more bad code and more maintainability problems. But if you're not able to like. Save a little bit of work here, delegate a little bit of work there. You will not you will not be able to compete
[23:26] with developers who do it. So, and that leads me to the other thing. I'm very sceptic. I don't see that those copilots generate good enough code. And I also don't see that it's fun to work with them. It's often very risky. You know, they give you wrong answers in a very non-obvious way, so it it will take you longer to actually find the problems. But nonetheless, you need to train it. And I can really tell you that, especially in the last month. You know, in the month before, I was just skeptical and basically ignored it or. Played with GPT, GBT, or Copilot, and was was very disappointed. But now, I see myself. Use it as like a pairing partner more and more. You know, when I write a blog article, I I give it material and let it draft for me, and then I see, oh, this is so terrible, or I let it write some code, and or I show it my code and ask what's the problem. So I would say that I use forms of generative AI nearly daily. In private life, but also in work life, and I would say per week it saves me between half a day or a full day, so a lot. Three months ago, I would have said I don't need it. You mentioned writing document like a blog or some technical spec, and you also mentioned programming. These two things
[24:57] I definitely can see the benefits. Do you have some other examples that you definitely used? Yeah, I mean it's understanding stuff, you know, like taking documents and finding stuff. You know, for example, in the just it's it's a it's stupid example, but in the past I really like scrolled through the full Salesforce release notes of all the five hundred pages. Okay, I just like load it into our SQ document and say, what's about scratch orgs or something. So you know, understanding. I also use it for analyzing data. You know, sometimes, you know, from customers we get like Excel sheets with things and with GPT four plus. You know, when you pay for it, you can really analyze. Data, you know, or you can put doc files in, or yeah. So I I also use it to make sense out of a lot of information. So I even have like a local LLM vector database. You know, when I get stuff from customers, I just put all the documents we get in there, and then I say, is there anything in it about X Y Z? And then I get like an answer, and I I use it. I do less Google searches. I have basically the ChatGPT open all day. Is it ChatGPT four or is it the free version? Okay. Okay. Before with web crawling
[26:27] and. And I say, yeah. What do you think? Even when I yesterday I finished a book about like Python learning, and I said, hey, make me make me a table with all the difference between Apex and Python. You know, okay. And it was it was good. It was really like having a person next to you, with whom you can just make arbitrary discussions about anything. So and yeah, and I. I I know that I was very skeptical myself, and I want to tell you that the more you say you don't need it, the less you will do it. But but actually should constantly like play with it and think about it and give it as a second chance a third chance and then you will actually see that you you get much more efficient with it. Thank you. I'll definitely take this advice. I will try it. Maybe I should start with paying the ChatGPT for version. I I think it's worth. How those many bucks? You know, I wouldn't I wouldn't pay the money for Copilot because I'm not a programmer and I think I I wouldn't need it, but. ChatGPT Plus is really worthwhile. Got it. Thanks, Robert. Before letting you go, I actually have one more thing to ask. I know the PMD has reached a new version seven. There are some breaking newses from there too, right? Totally, totally. Thanks for asking that because. You know that was that I
[27:57] think that was the how we learned to got to know each other and and I'm still very active in that respect not not a not writing code but you know last year I was doing a lot of conference talks and
[28:16] the the most interesting thing is that actually Salesforce. Is way more active, you know. Salesforce, you know, they have the code scanner, the code scanner, which should eventually fully replace checkmarks and all the other tools which are required for AppExchange partners, and, and the code code scanner is using PMD. So Salesforce already was active, like using it and promoting it, but now they actually have teams which contribute to PMD. You know, something very important has has changed in in PMD seven. We now have a open source Apex parser, and that's you know that that might sound very esoteric, but when I ported PMD. To Apex from Java, I was not able to write my own parser. So I, you know, I knew that that's too hard. So I took close source parser from the Eclipse IDE, and that was always a pain because you know people could not when when when the language got a new feature. Nobody could actually enhance it. We had to wait for Salesforce, and and what happened in in the recent years since we talked, like really smart people from Google and Financial Force, they built. Open source Apex parsers, and now in version seven, it it got replaced. So Pmd now even has a has a fully
[29:46] open source Apex parser. Do you use the Kevin Jones? Yeah, Kevin Jones from Financial Force. And another guy from from Google who's now with OpenAI, they they they build like yeah we have two Apex parsers now in PMD which are open source, and you know and for Salesforce to actually embrace that that that was a lot of like convincing work, but you know now Salesforce really said hey we will contribute to PMD when the new language feature of Apex comes out you know. A lot of stuff happens, like the null coalescing thing operator, or you know the thing that you can add to SOQL statements to make them safe. You know the the user mode, system mode. Whenever something is introduced into Apex, P and D has to understand it, and that is much better now. And I'm I'm really thankful for Salesforce to to have given the commitment that they. Not only use PMD, but also like give back. They will contribute. I think they they just started to sponsor PMD, even money wise, and that makes me very happy because that that's the goal. That they don't need me anymore. Yeah, it's really good to hear. But other than the parser side, do we have any like user facing features announced to four seven? I mean, or too many. Since since you and I talked has been a while. So what I did last
[31:16] year was basically telling the world that you can use PMD to scan everything. You know, not only Apex because that's what people don't know. You can you can scan your permission sets, your profiles, your custom fields, your flows with PMD, and you can write your own. Rules without actually programming Java, that was really something which I would say the Salesforce ecosystem didn't understand. So I really made a lot of sessions about how, you know, custom rules with XPath. You probably have seen that. So you really you basically all Salesforce metadata, all Apex can be can be observed by PMD, and this is a big thing. All right. Maybe one day you have an AI supported in the PMD. Yeah, and in the beginning, I really found that ChatGPT sometimes finds more issues in code than. PMD, but it can also be the other way around. You know, when I talked to Chris Peterson and Daniel Bailinger from the Apex team on Dreamforce, they said they want to have PMD to actually control that the quality of auto-generated AI-generated code and metadata is still good. You know. So, so not AI replacing PMD. PMD actually as a safeguard to prevent
[32:47] that AI is creating messy stuff. All right. One last question: If people want to learn more about using PMD to create the custom rules, those X path, I mean those things, need a bit the learning curve to build your own custom rules. So, where should they go to learn this stuff? I mean, they can, they can, they should follow me and and check check my my my LinkedIn and Twitter and, because they're you know most of the the things have been recorded. But I also created a custom GPT. You know, you can basically create a Chat GPT for yourself, and I created which helps people with P and D. You know, you you can basically ask it a question. You know, it's it's it's like a virtual Robert. I I will give you the link, and maybe you can put it in your show notes. I'll put it in the show notes. No, it's. It it it has my talks. It it has instructions on how to tell people. Wow! Excellent place to close the session. We're coming back to the AI again. Right, saying that you don't need me anymore. There's a virtue. Thanks, Robert. That's all. Hey, thanks for your time. Was great. Yeah.
