How AI Agents Are Changing Who Can Build Software

0:00 On this episode, the show I have with me Shosu Liu, he is Co founder and CTO at Loop. And we're going to be talking about how his team is leveraging STEM individuals and not software engineers to be involved in software engineering life cycle. And we're asking him how that is where that is and how that has shifted his team's productivity. 0:19 And we'll ask him a little bit about skills that he is seeing be a little bit more value in the age of AI. But Shosu, thanks for joining the time. Sure. Thank you, Amir. Excited to be here. Absolutely. All right. Before we start, Luke, what do you guys do over there? Yeah. So I I mean what Luke does is I'll type in the details, but I was given a very high level of what Luke aims to be, right? 0:41 What Luke does today we are a company that focused on industrializing service industry. And if I really step back, look at last 300 years of human effectively societal evolution. 0:56 We went from a Agricultural Society into a, you know, industrialized urbanized society. We with accumulation of income and we urbanized, we demand a lot more services. And the service economy has been growing since. And by the end of the 20th century, you know, we, we went through the information age transformation like today, like in the US, 80% of GDP actually come from services. 1:24 And, and, and you and me, we know we, we see this day in day out like AI has, AI is changing the paradigm how the service are delivered as an engineer, like software engineer, software development is fundamentally changed forever. Same things happening with legal through Harvey or diagnostics through Hippocratic. 1:47 And the word loop is focusing on our time on we want to accelerate this transformation of service industry, but we're going after the hardest service of those kind of categories. Think about supply chain logistic transformation like transportation, manufacturing, construction, right? 2:04 And we're effectively going after digitizing industrial, a lot of workflow that's the back office workflow doesn't really provide core value in those industry through AI and making people a lot more efficient so they can actually have more free time to really be applying themselves in probably more value generating or even more creative category of work. 2:26 So that's what we're Lube is focused on today and where we really start. And we're focused on supply chain. And with supply chain, there's a very specific category. No one even hears about it even in supply chain, let alone like even where I am, you know, SFB area, it's called a freight ODM pay loop. 2:47 Today we have about 400 customers with service about 20 Fortune 100 companies. And we help them understand how much can they spend on supply chain and help them to make sure that the the invoices are correct, they're accurate and we pay them in time and also make sure they are recording those charges accurate in their financial systems. 3:13 And the some of the largest company, I cannot even name names. I'm talking about trillion dollar valuation companies use loop. So we the entry can close their book and publish their earnings about how much they spend on transportation. And, and we're doing all of that right. 3:30 This is this industry is used to be all labour based, like you know, all over the world, BPO business process outsourcing and the world effectively taking over the market by increasing the amount we have done by AI and we can deliver this product, this product service faster, better and cheaper because the value we deliver is superior. 3:56 It's way more accurate, way more extensive. Now we're largely I compared to human bodies and I'm excited diving into some of the detail entry how we do that with you today. Absolutely. And I guess as we're trying to talking about AI within the engineering process, you mentioned that you guys leverage AI extensively. 4:19 I, I guess before we dive in, it's good for the audience to know what that looks like. So, so right now, obviously in the last year, a lot of things have changed. But right now, where, where do you guys use agentic? Where where's AI sit within the actual engineering development process? 4:37 Yeah, happy to kind of cover that really quick. So there's I mean there's obvious level when we talk about I think it's probably commonly shared across many companies. You know the the engineer software development process is 48. You know from from writing the specs, writing the code, deploying the code and then tested even though you have to roll out over the code are very much air driven. 5:00 Who then seen there like I think there's no significant difference between loop and other companies were pretty much on the cutting edge. But what I'm really seeing is almost the full software driven life cycle, not only on the software development side, but also the roll out, the validation of feature and the management feature in production. 5:22 Having more sensors than actuator integral with the right tooling might be observability, might be feature flag service that really amplify each individual engineer's impact. Like I used to work in big tech. We used to have a full team literally built on like feature flagging or feature rollout, like a test on certain traffic. 5:43 And you have like very dedicated tools for doing this now, like agents can do those at scale and and loop and we don't have to buy those tools. And it's mostly agent tech to me that actually create a biggest impact. That's that especially post the code merging, like that's where a lot of time actually being wasted in my opinion. 6:04 That's a lot of context switching and enable feature flag. We hope for half a day to make sure the metrics looks right and roll it out further. That's a ton of context switching that I see that having that part of the agent take that is extremely valuable. And the also the power user loop are really good at doing that that's risk will set them apart come out to power the average AI engineers at group. 6:28 Absolutely. I, I guess to talk about that a little, you, you mentioned, you know, leveraging agentic within the process itself and I guess you're freeing your software engineers to be more productive in other areas. 6:44 When it comes to specifically the areas you, you mentioned writing test harnesses, that's becoming something that you can get non software engineers involved in. What, what are some other areas that that you're seeing either agentic or non software engineers, you know, adding impact? 7:05 Yeah, I I was maybe I missed that back on qualify that statement a little bit, right. So there's a software development we do ourselves that's, you know, mostly like engineers that's very much AI native. If I really step back, right, what loop does we are a service provider. 7:22 We're not a SAS company. It's not like people people paying us to use our to the log into our web app to get a value. We have to deliver the value to our clients and that service delivery before loop was mostly done by human bodies. You know India, Philippine, a lot of this is the word BPO happens, right And what the loop is building there differently than the loops R&D team, engineering product and design, we call EPD team are mostly building harnesses on how this workflow actually done. 7:55 And we allow non engineers and in this case we call them AI OPS. They're mostly STEM educated, but they have a good sense of a data, extremely high data literacy. They understand very they have very, very good specs. Then our programming agents to deliver the service and loop provides our clients through AI. 8:17 So that's to me is the biggest different, biggest innovation that we were doing on loop because now we can customize like if I really think about traditional software, like if I think about the biggest software category, ER, like enterprise categories, Erps, CRMS, anyone who has used the hot spot or Salesforce understands, Oh, these are so many. 8:41 Like the two does so much it, it automates 85 ninety 95% of my work, but the last 5% that cannot be automated other than the last 5% span across different systems. And then my internal SOP doesn't actually work in this tools natively. 8:58 That's where like I spent 90% of my human time solving and having this sort of test like test harness that we created in our software product natively allow people to automate 100% and then last 5% of automation end up saving 100% of time. 9:18 And that's where I really see the value. And then by the way, these test harnesses are not being like they're being built by engineers, but they're being used by non engineers. And this non engineer end up writing way more code than Luke's engineers because we can now customize our product to the 100% requirement for each of our clients, which was not possible before and it's not possible today even in those large ERPCRN softwares. 9:48 You know, that's interesting because as we're kind of looking at AI and we're looking at what it can do. You mentioned that last mile, you know, up to now traditional software was written pretty darn close. 10:06 Like whatever the requirements were, they were, they were pretty close. Like I wouldn't say they're ever a fit. And then that last mile, companies have traditionally changed to adjust to that gap. Now with AI, that last mile, it's, it seems that it's fair to say that AI, with AI, you can get that last mile to fit the customer to make sure that what they're actually trying to do is what the application's trying, you know, getting out. 10:35 I guess when you look at the complexity of what it takes to produce a platform that delivers that last smile. I, I guess talk to us a little bit about is, is that a case of needing the four deployed engineers to go out in the field? 10:51 Is that case of is, is, is your AI platform, you know, maturing and learning what's happening, I guess at that from that standpoint? Yeah, I think that there there are a few different things. I, I definitely, you know, we, you know, loop loop is also hiring for deployment engineer. 11:09 So far different engineer very, very important actually getting in person, right. It's like just I was like in general, there are a few points, right. For deployment engineer is actually very, very critical in this process #2 is really how you, how do you organise and understand customers, business requirements, eventually manage that as contacts for your agents. 11:32 And lastly, in my mind is how do you actually accumulate like building a mode in this industry where in the day, right, the model are getting commoditized. But what really sets loop apart? What would set loop apart for me cloud Co work instance just by dropping a bunch of documents, right. 11:53 So I'll I'll start with the first one. So AI, you know, 40% is extremely critical and in would worry seeing that talking with the higher up you get like what's valuable? Why is it valuable? Very clear, like how do we actually create value for our clients? How do we get a commercial business deal like set up, but talking with the, the people on the ground doing their work in our customers organization may realize, oh, there's so many different use case edge cases that that's actually preventing this from being in the 1st place. 12:26 So doing that and the best way to actually qualify those into AIAS through 4 deploy engineers. So that's very, very, very, very real. And I do think that forward deploy engine, I think there's a few ways and people are leveraging 4 deploy engineers. 12:42 Some companies in fact are using them as very fancy alcohol information managers or solution engineering. They are more sales focused, customer focus, less technical. And Loop were really leveraging them as alcohol core product engineer. 12:59 They can, they can actually build a core product, but they can also talk to customer and also really understand navigate the customer organization dynamics. So it's very, I think Loop has a very different way of leveraging for deploy engineer. I would say very different, but it's probably 1 or few different ways than pure leverage for deploy engineer. 13:20 So that's very real on the loop. The second category on how do we actually build on software is very interesting is a loop where we're in very deep domain such as supply chain. And inherently we, we, we have, we have chosen to be highly our opinion on how we're actually building this domain. 13:42 We do a few things very uniquely #1 we do domain driven design and we believe there's a uniform domain and spanning across the supply chain network because everyone in supply chain, they all work with each other and they can clearly talk to each other and understand each other. 13:58 So there must be a common shared language they're using and how do we codify that language and teach them to AI? And with that trade off mean we are like at the same time, right? We are also doing billion payments. We're dealing with 10s of billions dollar payments per year as a company. 14:18 So we cannot afford certain sure, like heuristic or like, you know, hallucination in the model. So we don't do any rack ad loop. So you can think of our system organizing without has a highly strong skeletal structure which the domain model we believe is how industry operates and is curated by both industry experts as well as engineers. 14:42 But there's different flexible parts, maybe joints that's very much AI native that AI can actually do the work to customize it as much as we want. So in that world, you can imagine that the loops value were really, really cumulative value is this skeleton structure, which you know someone else can come up with if they spend that time. 15:06 But really the the the context that's being accumulated on those flexible parts in which we can be flexed. How can actually be flexible? What a sort of real world physical world entity or mapping that too, right? They're, they're very clear examples that you know, almost our customer has 20 different FedEx and they're in their system that's through merger and acquisitions and different, different sort of historical reasons. 15:33 But we know there's only two FedEx in this world. One is FedEx, the parcel shipper, the parcel carrier as well the FedEx Ground, the freight carrier. And our job is to reconcile the physical reality with our customer system so that anything we have learned about the physical reality, how FedEx parcel works and FedEx free works, we can apply them to our customer that can value through those, right. 16:00 I had a question about the forward deployed engineers. I may appreciate the context. That's actually really interesting to know. Obviously forward to play engineers are very hot right now. Everyone wants to hire them. You mentioned, you know, more technical, they could write core software engineering as well when it comes to hiring a forward to play engineer. 16:21 And obviously if they're engineers, they may not necessarily want to be involved in that particular role. I mean, how do you, how do you position the role so that it's attractive? Because not every engineer wants to be forward to play engineer, but obviously you're getting engineers that are interested in becoming FT ES. 16:37 That's an interesting point, right. If I really if I'm going to step back or if I observe for deploy engineer like we have this common scene within loop is that you don't find the forward different engineer who has 10 year experience mostly because #1 the 10 years ago only penalty has for deploy engineer and most of the people are there are very small number of people #2 for deploy engineer is not a terminal career. 17:02 It is passed to something else. And and you say for deployer is pretty demanding you have to go travel customer sites. So a lot of customer you can do both be evaluate on your interpersonal or national almost like the street smart aspect of your skill set, but also on your technical skill set, which already is pretty like is a hard skill set to develop and and curate, right. 17:28 So what we are seeing today is it's easier, I think relatively easy to for someone to already come in to the Ford deploy engineer role with some sort of personal trait or develop a soft skill that can demonstrate that can be trained and be successful in that regard versus someone who come out have have complete different personalities on the technical side. 18:03 I, I think that's actually in my mind that are way easier skill to get as long as I think there's some like very basic limiting factor on people's call it intelligence distribution, right? I might be physically intelligent. I'm very good at sports. I might be STEM intelligent, right? 18:20 Very good with numbers, logical concepts with the right people and in training the, the technical skills actually straightforward, but there's a minimum bar people have to cross and that's what we're seeing today. And with AI, actually the technical skills, it's probably even easier to train with the right personality because it's, it's a matter of pen tolerance. 18:41 The more pen you can tolerate, the faster you can learn the technical skill because the AI is already smarter than you on the technical skill. But no matter how do I build that from context, like going back and forth the AI into my intuition, which is building to my weights in my individual head, right? 18:57 That that's a matter of pen tolerance. The throughput is limited by how much pen you can tolerate. That's interesting. Yeah. I mean, I mean, that's that I think it's a fair assessment of, you know, you do need a specific type of personality, but also the technical skills. I guess when you're looking to interview and hire for. 19:15 I mean, we could just stick to the forward deploy engineer because I think a lot of people are doing it. You can obviously technically assess when it comes to the intangibles you're looking for. Those are a little bit trickier, you know those those communication skills, how do you weight that? 19:34 How do you go about assessing that balance of yeah, obviously you mentioned need to be technical, need to understand that, but also specific type of personality to want to be engaged with the customers in front of the customer, work hand in hand with the customer. How do you guys evaluate that? That's a good question. 19:51 I think we usually, so I can share our interview process, we do a, we do a effectively a commercial screening like just like literally. Based on past history, chatting about different situations, like it's pretty basic behavior interview that's very much focused on custom interaction. 20:09 We do a technical screen which is focused on very basic technical skills that we bring on site and doing a deeper dive on both the technical front and so on the commercial front. On the commercial front, right. So this is where a lot of measurement are. 20:25 It's it's not done by an engineer today. It's done by someone who's a lot more commercial focus, who's our head of head of deployment. He's ex founder. He has sold products. He's highly technical when the MIT right very he he, you know, he cleared the intelligence bar very, very well, but he also has been in the trenches. 20:43 He can build a lot of intuition around, Oh, can I put this person in front of customer, like in front of like Fortune 100 customers, head of a supply chain or COO that can demonstrate and show not only loops expertise, but also, you know, extra additional information that help people be successful in this demo, in this pilot, in this RFP, right? 21:04 That's a very, very delicate skill, right? It require a lot of confidence, a composure, but also like an art of speaking that gains, you know, presence or gains of sway on the other side. And again, them to share more information and now contribute the eventual success of the project. 21:23 So that that's I'll say it's definitely an art, right? It's the same thing that also it's like something that sales, right? You don't go to like MBA to become a counter executive to sell you, you, you actually learn from the industry. 21:39 There's no school teaching you to be a great salesperson. There's no such school that exists in the world. You have to actually be in a transgender. And the only people who have spotlight or talent are people who have gone through themselves can effectively evaluate other people. I love it. 21:55 So I was going to say that's I appreciate you sharing all your insights because I think a lot of people are looking to hire those forward to play engineers and obviously you guys are leveraging them as a part of the product as you've done outlined how they're useful within the process. 22:11 So I appreciate that if somebody does have a question, somebody might have a follow up question. What what's a good way of connecting with you and asking you about a follow up? Yeah, I mean, I'm, I'm, I'm pretty active on linkedinimihaveanemailthatseffectivelyshu@loop.com, which is my e-mail and I try to answer all my e-mail that's, you know, new grads or interns. 22:38 Yeah, so I think it's pretty easy to reach me. We actually are. We just finished our new gram hiring, but we do have a very different review on effectively emerging talents. We're actually very bullish on you as an intern or a very requested hiring those you know, for the for the 4006 summer, but also going to be coming from the next year as well. 22:58 Very cool. Obviously it's a super interesting opportunity as well. And, and, and appreciate you coming on and sharing. I know you're busy, so thank you for taking the time. Great. Thank you. Thank you. I mean great, great chatting. Absolutely. All right, that's end of this episode. 23:14 Be back again. Different guest, different topic. Until then, two things. Well, if you could share, this episode shows you did a great job of talking this about how AI is being leveraged within the product. How AI is helping the industry is obviously a gap within the industry of shifting from old ways of doing everything manually, heavy people driven to automating it with AI. 23:39 We also talked about how the forward deploy engineer is a model that is very useful and how they are going about hiring it. I know because that's a very difficult one out there. So I thought people would like that we talked about that. So share this episode with somebody who could benefit from it also like subscribe, comment, living how the show's going for you. 23:56 Until next time, thank you and goodbye.