🎙️ Ctrl + Shift + AI || Episode 1: "From Application Developer to AI Engineer"

[00:00:06]Hi everybody, welcome to the podcast series control shift AI. This is a conversational series where we talk to engineers from different backgrounds, stacks and are at different stages of their careers about their shift to AI.

[00:00:18]Some of our guests are just a few years into the industry while others have spent decades building software. So today we have our first guest Vive who has started his career in 2023 right when LLMs were open to public and chat GP was kind of skyrocketing. So let's talk to him and understand his AI journey. Hi V how are you doing?

[00:00:38]>> Hi Gri, I'm fine. How are you doing?

[00:00:40]>> Please introduce yourself to the audience.

[00:00:43]>> Yeah, sure. Uh hi everybody. I'm Vive uh 2023 CS graduate. I started my journey as a uh intern in wavemaker then got a full-time offer uh where I started after that I started learning web development and app development then later on recently I transitioned into AI engineering roles uh where I started building uh systems around AI uh using AI and for AI.

[00:01:08]>> Oh I like that you know building systems for AI and you know using AI. So you've mentioned that you have transitioned from web developer to an AI engineer right? Was this shift linear or you know what did interest you to in taking this leap?

[00:01:24]>> So when I started my career career like as you mentioned uh 2023 was very initial phase of chat GPT and the whole LLM wave. So I wasn't the early adopters of that. uh I did my development using traditional ways like going through stack overflow and community forums but later on there was this AI hackathon internally conducted where we wanted to build a chrome extension but none of my teammates had experience building a browser extension right >> so that was the first time when I personally uh use charge dip to build something uh using a coding then we were able to build that and uh it was kind of amazing what we could do with that chat GPT. Then later on uh there was this feature that we wanted to build. Uh we wanted to give a a user to uh ability to create their own personal theme just based on their prompt.

[00:02:21]>> So it was nothing but a basic LLM wrapper uh where they would define their theme and uh the theme variables would change and the user interface would change on the runtime.

[00:02:34]So these two things where I thought key uh there are a lot of possibilities and I was amazed key what we could build using LLMs and around LLMs.

[00:02:44]>> Nice. So starting from the nent stages of you know just prompting and getting the answer to get the work done >> to actually wanting to build something using AI and you know leveraging to build different solutions of it. I think it could have been quite a journey. So uh so these are the two stages that you have you know come across what do you like the most?

[00:03:09]So uh firstly the coming to the coding part like using AI we know like uh in early days we used to copy paste the snippets code snippets to our uh codes already existing codes right so we knew we already had a code base and we understood the code bases but later on if you see the it transitioned into inline autocomp completions the tools like super maven and tab 9 came out and you could they could directly suggest codes uh in And later on we saw a curs cursor coming out where the whole agentic system was inside the whole your ID. They had access to your file systems. They couldn't uh create new files and edit your files. So development became very easy in terms of like you have to just understand your system then you have to provide just a prompt or you have to better understand what the out better output looks like. Right? So that is not very interesting in terms of using AI to develop something but uh building systems around like which use AI to solve some meaningful use case that is a more fascinating and awesome if you look at that perspective. So Vive you mentioned about building systems right I'm sure there's a lot of learning that must be involved in that how does your learning curve look like is it a part of work that you do or you know is it selfdriven >> so as I mentioned uh we got into this AI hackathon >> and we built a use case using it was a basically LM wrapper right >> so these two things kind of in sparked the interest what else I could build with AI then later run I built couple of LLM rappers uh one or two chat boards then also I learned what rack like how we can improve the LLM context so I've built one or two rack systems small applications uh so this was my own curiosity after the hackathon but later on as a product based company we faced lot of customer use cases >> so one of the use cases was uh for a supply chain management company where they wanted to enhance the user experience by removing thousand plus CRUD screens and giving them a better experience by uh giving them one dynamic UI where they would just see whatever they wanted uh on in the runtime. So we built a MCP layer for this and a dynamic UI where LLMs would decide what UI to show to the user. And another use case was where we had to build a production rag system. So it was bit of both where it was my own curiosity and a work requirement also. Uh in the work requirement band we had to learn more uh technologies at the evolved but it was our curiosity and uh that we wanted to go deeper in the tech and build a production grade systems.

[00:06:07]>> Okay. By this you know um we understand that the evolution of LMS >> uh right when anthropic has standardized MCP because you mentioned about MCP right when it standardized the MCP uh we saw how powerful you know it can get when right tools are given to it and um you know going from just prompt engineering to actually uh integrating with systems. MCP has actually taken this AI world by storm I would say.

[00:06:39]So talking about that and you know listening to all your learnings I'm sure you must have your uh development practice like what does it look like?

[00:06:49]>> Yes. Uh so actually as you know I evolved from non from non using chat GPT to completely >> uh using coding agents and >> so uh recently my for currently my coding practices look like first I start build uh start with uh deciding the architecture >> and how the system would look like >> and how it should function and how it should not >> and then also you have to see it should it has to be scalable right if you want to make it to go to production.

[00:07:22]>> So you decide all this uh make the boundaries uh for the like uh also the coding conventions you also decide that >> and then you start the uh actual coding part.

[00:07:34]>> So I use uh coding models like GPT 5.5 and Opus 4 recently Opus 4.8 to decide all this architecture and also my own research. uh I still use stack overflow sometimes >> and after deciding all this I mean there are conventions uh for this agents like we can create cloudmd files for agent.mmd files >> uh then after that uh start with actual coding so this high-end uh models are used to decide the architecture then once that is fixed you can use any uh low reasoning models like on it or let's say if you're using cursor composer 2.5 >> to code the small parts but they don't have to think think much >> and I use very minimally Gemini models I mostly use them for media that is I had to create any image or any SDGs they are actually good with that and creating animations also so gymnasts are there and mostly I've been using as a coding we know uh every week we get new models and true >> and new coding ids and tools but uh recently I've fixed to few like coding ids or tools that is like codeex and cloud code cloud code mainly for and cloud code mainly for front end things like it performs better with front end >> yeah so this has been my uh practice is where you decide the system architecture and uh you have to know what is the right code. Right? If a model is uh giving you output something like giving you some code, it can give uh like infinite it can keep on giving code rate. You have to decide if the code is right or not.

[00:09:35]>> So you have to have that knowledge uh have that judgment if the code is right or not.

[00:09:41]>> So this has been my coding practice uh recently.

[00:09:45]So Vic, what do you like most about building with AI and what do you see in the near future?

[00:09:51]Uh >> so with the latest capabilities of models, uh so what I like is the language or the text stack is not a bottleneck anymore. So uh let's say there was this for example uh there was this hackathon uh where we wanted to build a mobile application. So it has to be a native mobile app. So the text stack required was cotling >> but none of us knew how to code with cotling. So there we kind of fully wipe coded it.

[00:10:22]>> So we didn't have to know the text stack. We just had to know what the output should look like. Right?

[00:10:28]>> So uh the knowledge about text stack is not a bottleneck anymore. I can build in any field or using any text stack. So this is interesting. Another thing the shift I'm seeing with coding agents or tools is that um they are eliminating the use of file edit or editors I mean so if you see codeex or if you see cursor and anti-gravity >> they have a separate agent window where you you just interact with uh models you don't have to open your file so they are going with this experience >> and also uh what I'm more interested is We are rightly using our phones to code.

[00:11:09]I mean laptop is is not required anymore because if you see in the cursor or codeex uh your code or repository is directly stored in the cloud.

[00:11:20]>> So you can make changes directly over there and it is com committed to your repo.

[00:11:24]>> So you don't have to open your laptop. I mean physical machine is not required anymore. And also this feature in CEX where you can directly work with your codeex uh environment directly from your phone. So you don't need your uh system in order to work on a project.

[00:11:42]>> So I feel like in future I won't need my laptop to code. I could directly uh spin up my project, commit and push and deploy probably from my phone itself.

[00:11:53]>> Yeah. So that is interesting.

[00:11:55]>> I second that uh thought Vive. I think most of us will relate to it and u your journey has been very interesting. Um thanks for coming over here and sharing your thoughts.

[00:12:07]>> Thank you Gaitri. Thank you for having me.

[00:12:09]>> Yeah. Thank you everybody. This was our first session with Vive and it was quite good to know his journey. Stay tuned to the space for more updates.