BraindumpDiscussion on Tornadoes and Video AI Project
Summary:
- Meeting involves discussion between Alex and Monash.
- Monash recently graduated and lives in Bloomington.
- They talk about recent tornado warnings and related damage in Bloomington.
- Monash developed a video AI project to summarize long lectures.
- The AI uses both transcripts and visual frames for summarization accuracy.
- Challenges encountered included issues with AWS Lambda and integration with Langsmith.
- Monash prefers coding by hand over using AI tools for programming.
Content:
So, kind of have like a middleman help keep the stuff going, help keep the meetings concise and that sort of thing. I see Alex is joining in here. Hey. All right. Can you hear me? Yeah, Alex. Hi. Yeah. Nice to meet you. All right. Yep. Yeah, it is. It's nice meeting you too. Yeah, it was nice meeting you. I'll see you later, maybe. And I will hand this meeting off. Hello, just give me one sec. I'm just finishing with a client here, sorry. Just all day. Okay, so how do you, first of all, how do you say your name?
Is it just Monash? Yeah, just Monash. Oh, okay. All right, easy. Okay. I was always wondering when I saw your name when I was reviewing that exercise. So, you graduated last month, right? Yeah. Okay, cool. Are you still in Bloomington or did you go, are you somewhere else? I'm still in Bloomington. I'm still living in Bloomington. Man, did you, I'm guessing, I mean, you're probably near downtown or somewhere around here, but those tornado warnings are getting crazy. It's like every, it's like every week. Yeah, so I think most of the tornadoes are from the east side and a little bit north, north side of the Bloomington, the Hussie Code.
Yeah, I think around, yeah, I don't, you know, I mean, I don't see that much destruction anywhere. Oh, I, I drove, I went to Indy, was that yesterday? I went to Indy and then I saw, it was like trees were just ripped in half. It was, it was such a large area and it was, I've never seen anything like it before. It wasn't like they were pulled out of the ground. They were just, it's almost like they were cut halfway up the tree and it was just gone. It was crazy. That's crazy, yeah. I've never seen the last tornado, it came, you know, Huntington Bank?
Yeah, oh yeah, yeah, yeah. Yeah, so there's one branch on the east side and it got destroyed completely. Yeah, it's, it seems like it's becoming more and more frequent here, but hopefully it doesn't happen again anytime soon. Yeah. I'm glad you're, I'm glad you're safe. Glad you stay safe. Let's see. I think I had some notes. Oh yeah, your resume. The, the video AI thing. What, how did you get, what got you like started with that or interested in that or how did you, what happened? How did you end up working on that? The video AI, the one you talking with the RAG project or the hackathon one.
The REG or the RAG? Is that what you said? Yeah. The REG one, RAG one. Yeah. You're just asking me, how did I come up with that idea? Or like, yeah, what, I mean, I don't know, what made you decide to do that? Yeah, I guess. So during my last semester, I was where I had to listen to a long lectures. It could be around two hours, two and a half hours. The other guest speakers from other universities and I'd write a whole summary report on those whole videos, whole lecture. So what we were provided is that we just provide the Zoom link after the meeting and I had to go through the whole video, pause play, pause play for every five minutes, three minutes and listen to what they've been talking on and then try to put everything in a paper and then again the same thing to pass it on.
I try to put everything together. So instead, I was like, okay, maybe I should build something where I just pass the whole video and I just ask it in a normal question, natural language. And then it tries to give the whole summary or then or tries to figure out a specific part that part has been spoken about in the whole two hour lecture instead of me going through manually part by part. So that's the whole idea it came from. Was it mainly just the, I mean, was it like the actual content of the video or was it just like the transcripts?
Like the, what all did you need to do? Yeah, we used both transcripts and both the visual frames. Oh yeah. Okay. Jeez. We used whisper to get the audio from the video and then we use whisper basically transfer the audio to a transcript and then we pass the basically initially I planned for 10 seconds frame capturing and then pass it to cloud IQ to produce a caption for the frame. But then later I figured that there could be a lot of things happening in the 10 seconds. So I implemented a few other techniques where screen cutting technique where if a system detects there's a huge change in the pixels, it tries to capture that frame.
That would be, I don't, we use or I use a fathom with one of the clients. I don't know if fathom does that or if it's just based on the transcript. I'm not sure about that. Well, I guess they use it and they give us the transcript, but anyway, there was a lot of like evaluation or tests of the system. Yeah. How did you, how did you like figure out if, did you have to go through and watch every video to make sure if it was accurate or how did you make sure it was giving you good results?
So initially, uh, what I did was I went to three whole videos. I wrote 10, 10 questions from each video with the exact timestamp where the answer can be found. I found that very irritating because watching all the videos again, it made me a bit frustrating. So what I did was there's a transcript option in YouTube, right? Yeah. So I just copy pasted the whole transcript to a cloud GPT 5.5. And with the transcript and the, all the text has been, as I gave my 30 questions as in a future, future prompting where I give provide examples to the model and model tries to give something output similar to what I give.
So based on the transcript and the question, it tries to frame questions and the exact timestamp and the random check like one, one note out of 10, one of 20. And all of those things matched so that the, there's a total of 135 UI questions are there with exact timestamps on the exact answer that could be found. Cool. Um, let me see what else. Um, sorry. Let's see. Did you get a chance to uh uh check something out from there? Oh, I didn't test it, no. I've been too busy. I didn't, I didn't test it myself.
Okay. Um, let's see. Do, do, do. Do you prefer writing code by hand or using AI tools? Sorry, what was that? Which one do you prefer, writing code by hand or using AI tools? I mean, that's like a big question because the whole job has completely changed since AI. Like, I kind of miss being able to do it by hand because it was more, I mean, it was like fun. It was the, I mean, that's what coding was. And now it's... It's like a completely different job. It's... Yeah. I think I read an article. It was like how the magic of coding has been lost and everyone's kind of just depressed now because, yeah, we don't get to, like, solve puzzles on our own.
Yeah. But, yeah, I don't know. I think I... I preferred it before, honestly. It was more, it was more engaging. It was more fun. Mm-hmm. Okay, I agree. Let's see. I guess I could just say, like, what's a technical, what's a decision that you had to make with that project that's, it could have gone a different way? Like if, I don't know, did you, how did you approach it and, like, did you, was there a point where you could have went a completely different direction with the plan on how you implemented it or, I don't know, I guess this would be early on.
Yeah. So the initial idea was that I used to pass only the transcript or the actual video frames, but then later on I figured out maybe there could be information on the video that was not spoken out and that won't be transcribed. Maybe I think that was a decision where I took, and that actually improved a lot of accuracy of retrieval. Sorry, say that one more time. I think using the video frames plus the transcript has increased the accuracy overall instead of just asking from the transcript. Oh, yes. Yeah, sorry. I'm trying to, like, adjust my sound because there's just a lot of bass from my speakers and that's all, even the bass is all the way down.
I wonder if I could... I think another one is that I think my accent is also a bit off. Oh, no, no, I don't think, I don't think it's that bad. I mean, I... Where exactly, by the way? I'm from India. Yeah, okay. I had a physics professor and he had a much thicker accent than you. It was... But I, I mean, I got used to it. It was, that was like my favorite professor, honestly. Okay. Let me see. I'm just trying to, like, give me a second here. I wanna see if I can lower the bass because it is all I hear.
Here, can you say something now? Yeah. Oh, there we go. Okay, perfect. I just, I have like an EQ that I can adjust. Okay. All right. What else here? I guess... I don't know. What issues, did you ever run into any, like, actual issues or major blockers? Yeah. So, for logging, I used a application called Langsmith. Have you heard of, yeah. Yeah. So, and for the AWS server, I used Lambda, which is like low cost and it is, it returns on only when there is work and then turns off when there is no work. It does idle actually.
So, when I initially used integrated architecture, it is in such a way that user get the request what it's been asking for and Lambda dies immediately. It doesn't give another second for the request to come back through Langsmith for a complete tracing. So Langsmith always keep on loading, saying that it's still waiting for the request, but the request has been completely dead from the Lambda side. So later on, I figured out the issue. I gave this some piece of line code, some total flash where the Lambda passes everything first before shutting it down. So yeah, so that was one of the, what do you say, that was one of the decision I had to make and one of the problem I faced.
Okay. Let me think of like a hypothetical. What if it, what if everything worked perfectly, but like, but the AWS, like the charges were like going out of control? What would, what would be the first thing that you like investigate if that was the case? I would say, basically, I can, there's a, I have actually used a service called CloudWatch in AWS where I can see like what's the, what's the billing cost for each service that has been running on. Maybe I'll try to go there and look what the, what's actually costing us more. Is it from the LLM model or is it from the other services such as Lambda, Fargate?
Maybe, yeah, that's the first thing I'll do. Okay, cool. Let's see. I guess that's enough for that. I mean, I go to your, the other exercise, the charioteer exercise that you did. When you worked through that, what do you think? What, okay, AI, where did it help the most and then where did it make more problems or create more work for you, would you say? The great help is that instead of normally if I had to use no AI, I had to clone the repository and I had to go through each file what's currently being there and I had to analyze from the scratch, right?
With the help of AI, with just a simple prompt, I get to know what's the current status of the project and what am I supposed to do, what are the next steps I have to take. And then the planning phase is also very helpful. Do I add all the steps in different phases where phase one goes first and then with the help of phase one, we try to build phase two. In that case, I would say AI is pretty useful. And where AI gave me trouble was that. Yeah, I think there was a conflict between part one and part four of the assignment with part four is where user selects, users can select the multiple API service, right?
They like four out of five Star Wars API servers where the users can select to retrieve the data. So what AI did was that it gave a simple solution. Instead of implementing that in the phase four, why don't we just implement in the phase one? That actually negated the prompt I gave to it. I asked you to be in a specific order, but then AI tried to simplify the whole thing and put everything at a single time. Okay. Let's see. I don't know if you answered that. I was thinking, like, and, I guess that would be the same thing.
I was going to say, like, what issue did you catch that the AI missed? Let's see. Um, I guess, how about when we asked for the failure handling revision, how did you decide which users should see when the request failed? So this was the follow-up question, right? Yeah. Yeah, can you repeat the question one more time? How did you decide what users should see when one detail request failed? Yeah, okay, yeah. Even I have asked the same question to LLM. First I passed it to Sonnet. It gave me a plan, like how to implement that.
But then when I gave it to GPT 5.5, it first, I think, it gave some output where it is not the exact approach that it would be taking. There are some edge cases that the first model missed. And I think using, by using the second model, I came to know that there are some edge cases that I need to focus on when using both cross- cross mode, I mean, investigating both the models' response. I got to know that. Okay. I'm trying to think of what else I could ask. Um, I mean, honestly, you got through it faster than other applicants.
Some people took several revisions, so you did a pretty good job with it. Um, let's see if I have anything else. Sorry, I just... One moment. Um, I guess I didn't have 30 minutes worth of questions. Next time I'll have to have more, I guess. Um, let me see what is going on in this Slack. I forget who's after me. Oh, Simon. It's just Simon to wrap up after me. Um, cool. Um, I'm trying to think of anything else I get... I mean, we could just talk, really. Let me see. Uh, what else could I possibly have?
Okay, then can I ask something? Yeah, sure. You could, yeah, I guess you could ask questions. Yeah, sure. I should have just said that. Go ahead and ask questions, I guess, if you have any. Yeah. So, uh, what are the biggest immediate challenges a person in this role will face in the first two to three weeks? The biggest what challenges? Uh, after, after getting this position as a HR, what could be the biggest challenges one could face in the starting phase? Hold on, it is, the base is so... Let me switch to my, my, just the speaker.
Okay, try it. Oh, it didn't even change. Hold on. I really can't understand. Hold on. All right, wait. Why isn't it? Hold on. It's refusing. Maybe I can't change it once we're in a call. Okay, talk now and see. Yeah, so the question I asked was, is it okay now? It's just not switching to my laptop. Maybe I think you need to restart the application by quitting. Maybe only then the changes might apply, I guess. I'm not sure. Oh, wait, no, here. Okay, it's in, I have to do it in Zoom, not, not this, my OS.
Okay, speak, talk. Okay, there we go. That's, that's through my laptop speaker, so it's not so much bassy. All right, what was the question? Sorry. Yeah, so what could be the biggest immediate challenges a person could face in this job role in during the first few weeks? Oh, okay. Um, well, it's, honestly, it's not too bad. I would just say like the learning curve of all of our, we have everything documented in wikis, but it can be kind of like overwhelming at first. But I guess, I mean, I don't think anybody really has time. I'm not going to go over all of them.
You just kind of focus on what you're working on, and then we'll, I mean, we'll try to help correct you to the resources that we already have without really having to look at the entire. It's just a huge wiki library, basically, that we've created over time. I mean, other than that, as long as, I mean, you've done well with the exercise, so I think that kind of is just what we do every day, you know? Okay. Just MRs. Well, MRs, PRs, I mean, I guess it's the same thing. Merge requests, pull requests, but yeah, it's just like that format.
I mean, I, I don't think it's, it's that difficult or there's not much of a learning curve other than like just carefully, don't, don't look at everything all at once because that can be overwhelming, but yeah. I'd focus on one thing first and then move on from there. Yeah, it's, I mean, I'm still seeing new things that I haven't seen before and I've been here for almost two years, so. Yeah, it's, yep. I don't know. It's not too bad, honestly. Mm-hmm. Yeah. That was the only question I had. Oh, okay. What did you think? Are you also, sorry, what was that?
I was going to ask, who else have you talked to? Chase and... I talked with Chase, I talked with Kyle and Keith. Oh, Keith. Yeah. I haven't really talked to Keith much personally. Yeah, I think he's really a good guy. He's very friendly. That's good. I haven't worked on many projects with him yet. Cool, cool. What were you going to ask about something else? I think I cut you off. I think I was going to say that was the question I had in my mind. Okay, yeah, I was going to ask you, like, are you, are you also based in Bloomington or are you in some other place?
Yep, I'm on the west side of Bloomington. Yep. Okay, okay. Yeah. Yeah, Elliot is telling me that we turned us people from all over the U.S. working remote, so I just want to ask whether you're also based in Bloomington. Yeah, I think, was it Kyle's from, where is Kyle from? North, is it North Carolina or something? I think that's what, or South Carolina? He's, yeah, Kyle's not in Indiana, I know that. I think he... Yeah, in his LinkedIn profile, he shows somewhere else, yeah. Sorry, say that one more time. Your pet is like 20 pounds, right?
Or 40, 40. 40 pounds, yeah. Yeah, I think it's not a small. My pet is just like 4 kgs. I mean, 8 pounds. Say that one more time. Sorry. You said your pet is like 20 pounds, right? Or 40, 40. 40 pounds, yeah. Yeah. But I think everyone else that you've met, well, I think, I think, I'm pretty sure. Okay. Because I've met Keith once at a Christmas party, which we have that every year in Bloomington. Okay. So I've never actually met Kyle in person. Mm-hmm. Yeah. But it's fun. Yeah, I don't really have anything else.
I don't know. I guess I should ping Simon or something. Let me see here. Mm-hmm. Okay. I'm gonna ping him. You just told me that future people from all over the U.S. working remote. So I just want to ask whether you're also based in Bloomington. Yeah, I think. Kyle's from, where is Kyle from? North, is it North Carolina or something? I think that's what, or South Carolina? He's, yes. Kyle's not in Indiana. I know that. I think he... Yeah, in his LinkedIn profile, I think it's not Carolina. Oh, okay, yeah. Like I said, LinkedIn. Yeah.
But I think everyone else that you met, well, I think Keith. I think he's from Indiana, I'm pretty sure. Okay. Because I met Keith once at a Christmas party. We have that every year in Bloomington. Okay. So I've never actually met Kyle in person. Mm-hmm. Yeah. But it's fun. Yeah, I don't really have anything else. I don't know. I guess I should ping Simon or something. Let me see here. Mm-hmm. Okay, ping him. Say in the future website, I see that you have a pet. What is it? A pet. Oh, yeah, I have a dog, yeah.
Yep. Yeah. What kind of breed is it? He's a... He's on the ground there. It's like an Australian cattle dog, I guess. Okay. It's like a blue heeler, but he's spotted instead of just one shade. Okay. Yeah. But he's kind of small. He's only 40 pounds. 40 pounds. 20 pounds, I think it's not small. My pet is just like 4 kgs. I mean, 8 pounds. Say that one more time, sorry. You said your pet is like 20 pounds, right? Or 40, 40. Okay, 40 pounds, yeah. Yeah. I think it's not a small. My pet is like 8 pounds now.
What kind of pet do you have? I have a toy poodle. A what? Oh, poodle. Toy poodle, yeah. What's the word that... Toy poodle. What's the word before poodle? Toy. Oh, toy poodle. Oh, yeah, those are little. Yeah, really. How heavy? You said 8 pounds? Yeah, 8 pounds. That is like, that's tiny. That's like a cat. Yeah. Gosh. I mean, that's kind of, it's, you don't have to buy as much food. That's for sure. I don't know how people have huge, like, 100-pound dogs. So much food. Yeah. But I'd like to have one in the future.
Maybe after I get settled, I will get a big dog. Yeah. I can just coach on you. Yeah. Well, I haven't heard from him yet. He's not doing anything. It says four. Yeah, four. We've got three minutes, I guess. Although he did, maybe he's busy with the other thing right now. Hmm. Oh, he looked. He did the eyes emoji. I'm assuming he's going to be here soon. Well, I mean, it was nice meeting you. I hope you get the position. Yeah. It was nice meeting you too, Alex. Yeah. I don't know where he is. Maybe I think he was trying to join at 4 sharp.
Yeah, maybe. We got two minutes, I guess. So when was the last time you went to the mill office in person? Oh, the mill? It's been, I don't even know. It's probably been over a year. Oh. OK. I mean, Simon's there every day, though, but I don't... I was hoping maybe I could go to the mill like once or twice in a week. Yeah, I don't know if they really... I guess you could. I mean, I think you're supposed to have, like, pay for a membership, but I don't know. I mean, we could just be guests of Simon, like, if he's there every day.
I would assume. I used to go there for my previous job. OK. Yeah, I used to go there more often. And that's how I met Simon, actually. Well, that plus, I went to Kenzie Academy, which his wife also went there. So it kind of, there's connections all over. I mean, yeah. You just know people in the industry and it's, yeah. One minute. We'll see. He might be late. It's a campus-based online school, Kenzie Academy. Sorry, say that one more time. I just looked up the Kenzak and like, what is it actually? Oh, what is it?
Well, it doesn't exist anymore. Um, it was like a, I guess it was like a boot camp, a software boot camp or a coding boot camp that was like a 12-month program, but... Okay. Yep. Simon is here. Oh, maybe. Can't hear you. Just try this mic. There you go. Yeah, it's working, exactly. Thanks. All right. All right. Well, nice meeting you again. Nice meeting you, Telex. Bye-bye. Take care. You too.
- Meeting involves discussion between Alex and Monash.
- Monash recently graduated and lives in Bloomington.
- They talk about recent tornado warnings and related damage in Bloomington.
- Monash developed a video AI project to summarize long lectures.
- The AI uses both transcripts and visual frames for summarization accuracy.
- Challenges encountered included issues with AWS Lambda and integration with Langsmith.
- Monash prefers coding by hand over using AI tools for programming.
Content:
So, kind of have like a middleman help keep the stuff going, help keep the meetings concise and that sort of thing. I see Alex is joining in here. Hey. All right. Can you hear me? Yeah, Alex. Hi. Yeah. Nice to meet you. All right. Yep. Yeah, it is. It's nice meeting you too. Yeah, it was nice meeting you. I'll see you later, maybe. And I will hand this meeting off. Hello, just give me one sec. I'm just finishing with a client here, sorry. Just all day. Okay, so how do you, first of all, how do you say your name?
Is it just Monash? Yeah, just Monash. Oh, okay. All right, easy. Okay. I was always wondering when I saw your name when I was reviewing that exercise. So, you graduated last month, right? Yeah. Okay, cool. Are you still in Bloomington or did you go, are you somewhere else? I'm still in Bloomington. I'm still living in Bloomington. Man, did you, I'm guessing, I mean, you're probably near downtown or somewhere around here, but those tornado warnings are getting crazy. It's like every, it's like every week. Yeah, so I think most of the tornadoes are from the east side and a little bit north, north side of the Bloomington, the Hussie Code.
Yeah, I think around, yeah, I don't, you know, I mean, I don't see that much destruction anywhere. Oh, I, I drove, I went to Indy, was that yesterday? I went to Indy and then I saw, it was like trees were just ripped in half. It was, it was such a large area and it was, I've never seen anything like it before. It wasn't like they were pulled out of the ground. They were just, it's almost like they were cut halfway up the tree and it was just gone. It was crazy. That's crazy, yeah. I've never seen the last tornado, it came, you know, Huntington Bank?
Yeah, oh yeah, yeah, yeah. Yeah, so there's one branch on the east side and it got destroyed completely. Yeah, it's, it seems like it's becoming more and more frequent here, but hopefully it doesn't happen again anytime soon. Yeah. I'm glad you're, I'm glad you're safe. Glad you stay safe. Let's see. I think I had some notes. Oh yeah, your resume. The, the video AI thing. What, how did you get, what got you like started with that or interested in that or how did you, what happened? How did you end up working on that? The video AI, the one you talking with the RAG project or the hackathon one.
The REG or the RAG? Is that what you said? Yeah. The REG one, RAG one. Yeah. You're just asking me, how did I come up with that idea? Or like, yeah, what, I mean, I don't know, what made you decide to do that? Yeah, I guess. So during my last semester, I was where I had to listen to a long lectures. It could be around two hours, two and a half hours. The other guest speakers from other universities and I'd write a whole summary report on those whole videos, whole lecture. So what we were provided is that we just provide the Zoom link after the meeting and I had to go through the whole video, pause play, pause play for every five minutes, three minutes and listen to what they've been talking on and then try to put everything in a paper and then again the same thing to pass it on.
I try to put everything together. So instead, I was like, okay, maybe I should build something where I just pass the whole video and I just ask it in a normal question, natural language. And then it tries to give the whole summary or then or tries to figure out a specific part that part has been spoken about in the whole two hour lecture instead of me going through manually part by part. So that's the whole idea it came from. Was it mainly just the, I mean, was it like the actual content of the video or was it just like the transcripts?
Like the, what all did you need to do? Yeah, we used both transcripts and both the visual frames. Oh yeah. Okay. Jeez. We used whisper to get the audio from the video and then we use whisper basically transfer the audio to a transcript and then we pass the basically initially I planned for 10 seconds frame capturing and then pass it to cloud IQ to produce a caption for the frame. But then later I figured that there could be a lot of things happening in the 10 seconds. So I implemented a few other techniques where screen cutting technique where if a system detects there's a huge change in the pixels, it tries to capture that frame.
That would be, I don't, we use or I use a fathom with one of the clients. I don't know if fathom does that or if it's just based on the transcript. I'm not sure about that. Well, I guess they use it and they give us the transcript, but anyway, there was a lot of like evaluation or tests of the system. Yeah. How did you, how did you like figure out if, did you have to go through and watch every video to make sure if it was accurate or how did you make sure it was giving you good results?
So initially, uh, what I did was I went to three whole videos. I wrote 10, 10 questions from each video with the exact timestamp where the answer can be found. I found that very irritating because watching all the videos again, it made me a bit frustrating. So what I did was there's a transcript option in YouTube, right? Yeah. So I just copy pasted the whole transcript to a cloud GPT 5.5. And with the transcript and the, all the text has been, as I gave my 30 questions as in a future, future prompting where I give provide examples to the model and model tries to give something output similar to what I give.
So based on the transcript and the question, it tries to frame questions and the exact timestamp and the random check like one, one note out of 10, one of 20. And all of those things matched so that the, there's a total of 135 UI questions are there with exact timestamps on the exact answer that could be found. Cool. Um, let me see what else. Um, sorry. Let's see. Did you get a chance to uh uh check something out from there? Oh, I didn't test it, no. I've been too busy. I didn't, I didn't test it myself.
Okay. Um, let's see. Do, do, do. Do you prefer writing code by hand or using AI tools? Sorry, what was that? Which one do you prefer, writing code by hand or using AI tools? I mean, that's like a big question because the whole job has completely changed since AI. Like, I kind of miss being able to do it by hand because it was more, I mean, it was like fun. It was the, I mean, that's what coding was. And now it's... It's like a completely different job. It's... Yeah. I think I read an article. It was like how the magic of coding has been lost and everyone's kind of just depressed now because, yeah, we don't get to, like, solve puzzles on our own.
Yeah. But, yeah, I don't know. I think I... I preferred it before, honestly. It was more, it was more engaging. It was more fun. Mm-hmm. Okay, I agree. Let's see. I guess I could just say, like, what's a technical, what's a decision that you had to make with that project that's, it could have gone a different way? Like if, I don't know, did you, how did you approach it and, like, did you, was there a point where you could have went a completely different direction with the plan on how you implemented it or, I don't know, I guess this would be early on.
Yeah. So the initial idea was that I used to pass only the transcript or the actual video frames, but then later on I figured out maybe there could be information on the video that was not spoken out and that won't be transcribed. Maybe I think that was a decision where I took, and that actually improved a lot of accuracy of retrieval. Sorry, say that one more time. I think using the video frames plus the transcript has increased the accuracy overall instead of just asking from the transcript. Oh, yes. Yeah, sorry. I'm trying to, like, adjust my sound because there's just a lot of bass from my speakers and that's all, even the bass is all the way down.
I wonder if I could... I think another one is that I think my accent is also a bit off. Oh, no, no, I don't think, I don't think it's that bad. I mean, I... Where exactly, by the way? I'm from India. Yeah, okay. I had a physics professor and he had a much thicker accent than you. It was... But I, I mean, I got used to it. It was, that was like my favorite professor, honestly. Okay. Let me see. I'm just trying to, like, give me a second here. I wanna see if I can lower the bass because it is all I hear.
Here, can you say something now? Yeah. Oh, there we go. Okay, perfect. I just, I have like an EQ that I can adjust. Okay. All right. What else here? I guess... I don't know. What issues, did you ever run into any, like, actual issues or major blockers? Yeah. So, for logging, I used a application called Langsmith. Have you heard of, yeah. Yeah. So, and for the AWS server, I used Lambda, which is like low cost and it is, it returns on only when there is work and then turns off when there is no work. It does idle actually.
So, when I initially used integrated architecture, it is in such a way that user get the request what it's been asking for and Lambda dies immediately. It doesn't give another second for the request to come back through Langsmith for a complete tracing. So Langsmith always keep on loading, saying that it's still waiting for the request, but the request has been completely dead from the Lambda side. So later on, I figured out the issue. I gave this some piece of line code, some total flash where the Lambda passes everything first before shutting it down. So yeah, so that was one of the, what do you say, that was one of the decision I had to make and one of the problem I faced.
Okay. Let me think of like a hypothetical. What if it, what if everything worked perfectly, but like, but the AWS, like the charges were like going out of control? What would, what would be the first thing that you like investigate if that was the case? I would say, basically, I can, there's a, I have actually used a service called CloudWatch in AWS where I can see like what's the, what's the billing cost for each service that has been running on. Maybe I'll try to go there and look what the, what's actually costing us more. Is it from the LLM model or is it from the other services such as Lambda, Fargate?
Maybe, yeah, that's the first thing I'll do. Okay, cool. Let's see. I guess that's enough for that. I mean, I go to your, the other exercise, the charioteer exercise that you did. When you worked through that, what do you think? What, okay, AI, where did it help the most and then where did it make more problems or create more work for you, would you say? The great help is that instead of normally if I had to use no AI, I had to clone the repository and I had to go through each file what's currently being there and I had to analyze from the scratch, right?
With the help of AI, with just a simple prompt, I get to know what's the current status of the project and what am I supposed to do, what are the next steps I have to take. And then the planning phase is also very helpful. Do I add all the steps in different phases where phase one goes first and then with the help of phase one, we try to build phase two. In that case, I would say AI is pretty useful. And where AI gave me trouble was that. Yeah, I think there was a conflict between part one and part four of the assignment with part four is where user selects, users can select the multiple API service, right?
They like four out of five Star Wars API servers where the users can select to retrieve the data. So what AI did was that it gave a simple solution. Instead of implementing that in the phase four, why don't we just implement in the phase one? That actually negated the prompt I gave to it. I asked you to be in a specific order, but then AI tried to simplify the whole thing and put everything at a single time. Okay. Let's see. I don't know if you answered that. I was thinking, like, and, I guess that would be the same thing.
I was going to say, like, what issue did you catch that the AI missed? Let's see. Um, I guess, how about when we asked for the failure handling revision, how did you decide which users should see when the request failed? So this was the follow-up question, right? Yeah. Yeah, can you repeat the question one more time? How did you decide what users should see when one detail request failed? Yeah, okay, yeah. Even I have asked the same question to LLM. First I passed it to Sonnet. It gave me a plan, like how to implement that.
But then when I gave it to GPT 5.5, it first, I think, it gave some output where it is not the exact approach that it would be taking. There are some edge cases that the first model missed. And I think using, by using the second model, I came to know that there are some edge cases that I need to focus on when using both cross- cross mode, I mean, investigating both the models' response. I got to know that. Okay. I'm trying to think of what else I could ask. Um, I mean, honestly, you got through it faster than other applicants.
Some people took several revisions, so you did a pretty good job with it. Um, let's see if I have anything else. Sorry, I just... One moment. Um, I guess I didn't have 30 minutes worth of questions. Next time I'll have to have more, I guess. Um, let me see what is going on in this Slack. I forget who's after me. Oh, Simon. It's just Simon to wrap up after me. Um, cool. Um, I'm trying to think of anything else I get... I mean, we could just talk, really. Let me see. Uh, what else could I possibly have?
Okay, then can I ask something? Yeah, sure. You could, yeah, I guess you could ask questions. Yeah, sure. I should have just said that. Go ahead and ask questions, I guess, if you have any. Yeah. So, uh, what are the biggest immediate challenges a person in this role will face in the first two to three weeks? The biggest what challenges? Uh, after, after getting this position as a HR, what could be the biggest challenges one could face in the starting phase? Hold on, it is, the base is so... Let me switch to my, my, just the speaker.
Okay, try it. Oh, it didn't even change. Hold on. I really can't understand. Hold on. All right, wait. Why isn't it? Hold on. It's refusing. Maybe I can't change it once we're in a call. Okay, talk now and see. Yeah, so the question I asked was, is it okay now? It's just not switching to my laptop. Maybe I think you need to restart the application by quitting. Maybe only then the changes might apply, I guess. I'm not sure. Oh, wait, no, here. Okay, it's in, I have to do it in Zoom, not, not this, my OS.
Okay, speak, talk. Okay, there we go. That's, that's through my laptop speaker, so it's not so much bassy. All right, what was the question? Sorry. Yeah, so what could be the biggest immediate challenges a person could face in this job role in during the first few weeks? Oh, okay. Um, well, it's, honestly, it's not too bad. I would just say like the learning curve of all of our, we have everything documented in wikis, but it can be kind of like overwhelming at first. But I guess, I mean, I don't think anybody really has time. I'm not going to go over all of them.
You just kind of focus on what you're working on, and then we'll, I mean, we'll try to help correct you to the resources that we already have without really having to look at the entire. It's just a huge wiki library, basically, that we've created over time. I mean, other than that, as long as, I mean, you've done well with the exercise, so I think that kind of is just what we do every day, you know? Okay. Just MRs. Well, MRs, PRs, I mean, I guess it's the same thing. Merge requests, pull requests, but yeah, it's just like that format.
I mean, I, I don't think it's, it's that difficult or there's not much of a learning curve other than like just carefully, don't, don't look at everything all at once because that can be overwhelming, but yeah. I'd focus on one thing first and then move on from there. Yeah, it's, I mean, I'm still seeing new things that I haven't seen before and I've been here for almost two years, so. Yeah, it's, yep. I don't know. It's not too bad, honestly. Mm-hmm. Yeah. That was the only question I had. Oh, okay. What did you think? Are you also, sorry, what was that?
I was going to ask, who else have you talked to? Chase and... I talked with Chase, I talked with Kyle and Keith. Oh, Keith. Yeah. I haven't really talked to Keith much personally. Yeah, I think he's really a good guy. He's very friendly. That's good. I haven't worked on many projects with him yet. Cool, cool. What were you going to ask about something else? I think I cut you off. I think I was going to say that was the question I had in my mind. Okay, yeah, I was going to ask you, like, are you, are you also based in Bloomington or are you in some other place?
Yep, I'm on the west side of Bloomington. Yep. Okay, okay. Yeah. Yeah, Elliot is telling me that we turned us people from all over the U.S. working remote, so I just want to ask whether you're also based in Bloomington. Yeah, I think, was it Kyle's from, where is Kyle from? North, is it North Carolina or something? I think that's what, or South Carolina? He's, yeah, Kyle's not in Indiana, I know that. I think he... Yeah, in his LinkedIn profile, he shows somewhere else, yeah. Sorry, say that one more time. Your pet is like 20 pounds, right?
Or 40, 40. 40 pounds, yeah. Yeah, I think it's not a small. My pet is just like 4 kgs. I mean, 8 pounds. Say that one more time. Sorry. You said your pet is like 20 pounds, right? Or 40, 40. 40 pounds, yeah. Yeah. But I think everyone else that you've met, well, I think, I think, I'm pretty sure. Okay. Because I've met Keith once at a Christmas party, which we have that every year in Bloomington. Okay. So I've never actually met Kyle in person. Mm-hmm. Yeah. But it's fun. Yeah, I don't really have anything else.
I don't know. I guess I should ping Simon or something. Let me see here. Mm-hmm. Okay. I'm gonna ping him. You just told me that future people from all over the U.S. working remote. So I just want to ask whether you're also based in Bloomington. Yeah, I think. Kyle's from, where is Kyle from? North, is it North Carolina or something? I think that's what, or South Carolina? He's, yes. Kyle's not in Indiana. I know that. I think he... Yeah, in his LinkedIn profile, I think it's not Carolina. Oh, okay, yeah. Like I said, LinkedIn. Yeah.
But I think everyone else that you met, well, I think Keith. I think he's from Indiana, I'm pretty sure. Okay. Because I met Keith once at a Christmas party. We have that every year in Bloomington. Okay. So I've never actually met Kyle in person. Mm-hmm. Yeah. But it's fun. Yeah, I don't really have anything else. I don't know. I guess I should ping Simon or something. Let me see here. Mm-hmm. Okay, ping him. Say in the future website, I see that you have a pet. What is it? A pet. Oh, yeah, I have a dog, yeah.
Yep. Yeah. What kind of breed is it? He's a... He's on the ground there. It's like an Australian cattle dog, I guess. Okay. It's like a blue heeler, but he's spotted instead of just one shade. Okay. Yeah. But he's kind of small. He's only 40 pounds. 40 pounds. 20 pounds, I think it's not small. My pet is just like 4 kgs. I mean, 8 pounds. Say that one more time, sorry. You said your pet is like 20 pounds, right? Or 40, 40. Okay, 40 pounds, yeah. Yeah. I think it's not a small. My pet is like 8 pounds now.
What kind of pet do you have? I have a toy poodle. A what? Oh, poodle. Toy poodle, yeah. What's the word that... Toy poodle. What's the word before poodle? Toy. Oh, toy poodle. Oh, yeah, those are little. Yeah, really. How heavy? You said 8 pounds? Yeah, 8 pounds. That is like, that's tiny. That's like a cat. Yeah. Gosh. I mean, that's kind of, it's, you don't have to buy as much food. That's for sure. I don't know how people have huge, like, 100-pound dogs. So much food. Yeah. But I'd like to have one in the future.
Maybe after I get settled, I will get a big dog. Yeah. I can just coach on you. Yeah. Well, I haven't heard from him yet. He's not doing anything. It says four. Yeah, four. We've got three minutes, I guess. Although he did, maybe he's busy with the other thing right now. Hmm. Oh, he looked. He did the eyes emoji. I'm assuming he's going to be here soon. Well, I mean, it was nice meeting you. I hope you get the position. Yeah. It was nice meeting you too, Alex. Yeah. I don't know where he is. Maybe I think he was trying to join at 4 sharp.
Yeah, maybe. We got two minutes, I guess. So when was the last time you went to the mill office in person? Oh, the mill? It's been, I don't even know. It's probably been over a year. Oh. OK. I mean, Simon's there every day, though, but I don't... I was hoping maybe I could go to the mill like once or twice in a week. Yeah, I don't know if they really... I guess you could. I mean, I think you're supposed to have, like, pay for a membership, but I don't know. I mean, we could just be guests of Simon, like, if he's there every day.
I would assume. I used to go there for my previous job. OK. Yeah, I used to go there more often. And that's how I met Simon, actually. Well, that plus, I went to Kenzie Academy, which his wife also went there. So it kind of, there's connections all over. I mean, yeah. You just know people in the industry and it's, yeah. One minute. We'll see. He might be late. It's a campus-based online school, Kenzie Academy. Sorry, say that one more time. I just looked up the Kenzak and like, what is it actually? Oh, what is it?
Well, it doesn't exist anymore. Um, it was like a, I guess it was like a boot camp, a software boot camp or a coding boot camp that was like a 12-month program, but... Okay. Yep. Simon is here. Oh, maybe. Can't hear you. Just try this mic. There you go. Yeah, it's working, exactly. Thanks. All right. All right. Well, nice meeting you again. Nice meeting you, Telex. Bye-bye. Take care. You too.