He Built a RAG Project as a Student, But Couldn't Explain His Own Code

Indexed: 2026-05-22

[00:00:00]First thing I'll be honest, your performance is not good. Hey Vishnav, really nice to meet you.

[00:00:05]>> Hi sir.

[00:00:07]>> So Vishnav uh welcome to this mock interview. So we are going to take around 45 minutes I would say roughly to get this interview done. Especially we focus initial 40 minutes on the questions I want to ask as well as back and forth the technical questioning.

[00:00:21]Right after that we are going to go into the review of the interview. basically how you think your interview went, how I think your interview went as well. Okay.

[00:00:30]>> Okay.

[00:00:31]>> Uh so with that in mind, uh let's start with your introduction and we can take it up from there.

[00:00:35]>> Hi, I am Vishna currently pursuing uh engineering. Recently I have been focusing a lot on backend systems and a native application specially around and contextware workflows. One of the project that I have currently building is uh PR review agent uh using embed embeddings uh retrieval pipelines GitHub workflows and LM LLM based code reviews.

[00:00:57]Uh what really interests me is broader problem about the context is engineering make sure that uh AI gets the context I did right.

[00:01:06]>> Awesome. That's great. And I was actually going through your resume.

[00:01:10]Okay. So you did do freelancing works as well, right? Uh so that was uh opensource contribution so I added it under the freelance work.

[00:01:20]>> Got it. Makes sense. So it's always great that you know you are working for opensource projects because these are the projects that make us you know at the end of the day better engineers.

[00:01:28]>> Yeah.

[00:01:29]>> So just give me one minute. I'm just going through your projects and everything and we can continue after that. Yeah.

[00:01:36]>> So one interesting thing I did observe is you have some Java experience as well. So let's start from there.

[00:01:43]How do you think for example if I ask you to build entire AI application one of the college project you have done or whatever in Java and completely shift it from Python to Java. So how would you do it if that is the request?

[00:01:59]>> So it means I have to shift my total technology to Python.

[00:02:02]>> Java no total technology to Java not Python. So you have created your project in Python because generally that's what AI projects generally people do right they get examples which are already in Python. So you used Python and you created your entire project but suddenly I don't know maybe uh I am your professor and I started to tease you or whatever I wanted to ask you to completely shift this project back to Java for example. So how would you approach this problem?

[00:02:28]>> Uh what I think of is now is uh I would take down the all the details what I use in my Java Python project which I have to shift in Java like uh for back end I have used fast API in the Python. So I would ship to the spring boot where where I have to use what we can say else. So I'm not >> so for example what about packages maybe there is you know blang graph and all those things right and there is if there is an ML model you need to you might be using pytor so how are you going to shift these packages or code related to those packages >> okay >> so I don't have a idea about it current >> no worries I mean it's just curious question from my side so you also mentioned you know C++ and SQL right >> uh so yeah I use C++ plus pro plus 40 DSA >> practice for faster execution.

[00:03:20]>> Cool. I mean let's start with one very easy DSA problem then okay because I seen there is a uh you know you solve a lot of problems in late code as well.

[00:03:29]Yeah, >> the question is basically uh insertion sort. Okay, you need to implement insertion sort without actually writing the code. You need to explain it to me. Okay. Okay.

[00:03:41]>> So, how would you do it if I ask you to do it now?

[00:03:44]>> So, any example where >> so basically I have an array of 100 elements. The elements are already sorted 1 2 3 4 5 6 7 100. So, I need to run insertion sort on them. Okay. So how would you write the logic for insertion sort and what happens if the array is already sorted? Do you think insertion sort performs well or it does double work? What happens?

[00:04:08]>> So I think it would work. Insertion sort sort will work on the array.

[00:04:13]>> Obviously it works right because it's already a good sorting algorithm. It should work.

[00:04:18]>> But what if the array is already sorted?

[00:04:20]Uh does insertion sort immediately give it that it's already sorted or it does lot more work. So what is the situation there? it would uh give that array is sorted itself.

[00:04:29]>> Uh so it doesn't do much operations.

[00:04:31]You're saying it will immediately arrive at the suggestion that array is sorted.

[00:04:35]Okay.

[00:04:36]>> Yeah.

[00:04:36]>> Cool. Uh let's move to the next section.

[00:04:39]So can you just explain to me the entire rag application you have created right?

[00:04:45]>> Okay.

[00:04:46]>> Whatever the rack technologies you have used. So can you just share your screen and show me the code and exactly walk me through the application so that you know we can have uh back and forth discussions. I mean that would be very interesting.

[00:04:58]>> Okay. So can we start at the main file basically where the program starts running and can you walk me over APIs as well as what are the logic associated with each of these API?

[00:05:09]>> Uh should I go through the um what do we say agents or uh >> so let's start with the main.py Pi okay explain me what is happening there >> okay >> and then let's show me what exactly happens after you know for example we start at main then where the program goes for example are there any specific files which are you know discussing APIs from there we'll be going to services basically how the service logic is written so so that I can understand entire application right >> okay so I have done the uh uh this is the mainpay of the project itself uh I have implement the logger so whenever any error occurs. So if someone is solving the any errors we he should know that the like what we see a PR review agent is shutting down. So it will showing the terminal what actually is happening. So uh the person who is doing the work uh get the better idea of it. I use the synchronous dev function here where the PR isn't are start get started then P.

[00:06:10]>> So I have few questions here Vishnov.

[00:06:13]Can you explain me what is this async as a as s y and c what does it do?

[00:06:18]>> So the the main function of async is when the two what we can say when the uh I'm not able to get it right now.

[00:06:26]>> No worries no worries. Uh okay let's go to the next one. Uh let's please continue your explanation. Uh >> so I have the uh verify github signature. What this uh does is like when the it it checks whether the uh PR which is coming through the GitHub uh is genuine or not is it fake or not it's uh avoids it. So we have uh got the open up the GitHub where we get the uh signature of the GitHub and it's verified whether it's a GitHub PR or not.

[00:06:58]>> Okay. I again have few questions here.

[00:07:01]>> Okay. So you are getting this GitHub web hook secret right from settings. Why not get it directly from ENV file because you I have already seen that there are few ENV files there right? So why are we using the settings and then getting it from there?

[00:07:15]>> So what I think is that I have put the all the files in but uh it would make the more modular to use what I think of >> and is there any other use uh instead of modularity that settings.py Pi file that you are using right so is it giving you >> if you can you open the settings.py Pi file or if wherever the settings are there.

[00:07:37]>> Uh I have put in config.

[00:07:39]>> Okay. If it's in Yeah. Let's go there.

[00:07:41]>> Yeah.

[00:07:42]>> So what actually you are doing here?

[00:07:44]>> So I have the class of setting where I have uh given all the um settings or tokens and all where I can use that directly in my uh order files like GitHub token is in the string form secret.

[00:07:59]>> So why not get it from ENV again? you are getting here also into from ENV only right now instead of directly getting from ENV now you put here and from here you are getting those sequences >> so I can't get it directly from env what I think is uh >> you can get it there is a package called env in python >> uh you can use it and get it directly wherever you want >> but um word exactly I cannot think of >> no worries no worries I'm just trying to get more from you but the reason you did this is if you see you're telling Gemini API key API keys string str right >> so what happens if it's not a string by mistake someone has put number or someone has completely left it >> you are going to immediately get an error >> when the settings this file is getting com you know it's getting used right instead of while the runtime is happening you are going to get from env and then you found out it's not string sometimes your functions might take take them directly and you get unexpected errors right so that's why this >> whatever the data types you want those environment variables to be you are going to manage from this particular config.file file so that your error is much earlier than later in the system life cycle, right? Yeah.

[00:09:12]>> So, >> yeah, I mean that's awesome. Let's go ahead please uh let's say go back to that previous file we are discussing and we'll start from there.

[00:09:20]>> So, uh here the app dot uh post I have used here for the injection purpose. It shows that injection has started and it gives the full uh repo name and all that uh stuff.

[00:09:32]>> Okay, that's awesome.

[00:09:33]>> Yeah. So this add the rate app.post what is this at the rate there? Uh sir I think so it is the tool calling.

[00:09:40]>> Tool calling.

[00:09:42]>> Yeah.

[00:09:42]>> Uh if I say that is wrong what would be your answer?

[00:09:47]>> Sorry.

[00:09:48]>> No worries.

[00:09:49]>> Controller. Sorry. We can use it in controller where what we can say.

[00:09:53]>> Have you uh for example have you ever heard about decorators and these kind of things in Python?

[00:10:00]>> Uh no sir.

[00:10:01]>> Okay. Okay. No worries. Let's continue.

[00:10:03]Uh let's go to the logic part.

[00:10:05]>> All right. Should I open through agents or >> Yeah. So let us maybe go to this fetcher.py >> uh and see what is happening there.

[00:10:15]>> So in fetcher.py uh it's get the client token from the GitHub. Then uh it it started the fetching the PR and PR repo name and it checks if the repo name is correct or not. So if it cannot find the repo name it's uh use the uh exception like the could not find the repo name and it's particularly raise it. So when the error comes the one who is using the terminal should get to know where the error has happened and it's same as for the PR number also uh does the exemption if he could not find the PR number it raised the error second the uh >> so what if I remove that raise statement there what happens >> so uh nothing uh serious would happen but uh whenever some error happens and which is because of the uh what we can If we not able to face the error or face the PR so error it would give exception but we will not particularly able to know the where the error has happened.

[00:11:18]It would make easier >> okay >> to check whether the error >> you are anyways logging it right for example you have logger error which is going to log an error in your wherever you're putting these logs >> and maybe you can put a filter on top of it right maybe whenever there's an error just send an email to few of the users so that will also give me information why do I need to use raise there >> so it I think so it will particularly raise it what >> so you are saying raise to what raise to terminal Basically I get an information that the error has happened as a developer. Is that what it is doing?

[00:11:53]>> So no I am not able to >> No worries. No worries. Let's go. Let's go to the next step.

[00:11:58]>> So uh it iterates through all the files in the uh GitHub repo and check for the change files. Check the status of the change file. What files are added? Uh what files are deleted and here the I uploaded the patch file. So for example if there is any >> this changed file uh sorry to interrupt this changed file right what is it >> uh it the so for example any PR is there uh it it checks in that file what has changed that you get it >> yes I got it so you have created this change file right so I think you might have mentioned file name should be this status should be this and you have given certain rules to it right >> yeah and I have implement the patch also >> okay got it So now the changed files itself is an array and you are appending this change file object to that right.

[00:12:48]So now the append is happening at the end of the array for example.

[00:12:52]But if I want to append at the start of the array every time I get a new changed file what code change can I do here?

[00:13:00]>> What we have to do?

[00:13:01]>> So basically append what happens there is an array. Whenever there is a changed file object that is coming it's adding to that array at the end of the array.

[00:13:08]Right? Because append that's what it does. But what if I want to add at the start of the array every time I get a new object?

[00:13:15]>> So I'm not able to do.

[00:13:17]>> Okay. Okay. No worries. I mean one easier way is just you know you can reverse the array or I mean there are multiple ways to do it. There is a function also which can actually add at the start of the array. Uh let's continue from here.

[00:13:30]>> Okay. So what uh what was your question exactly? We have to get the last change file to the first. So basically now what happening this for loop every time whenever some new file is coming it is going to get added at the end of the array. So array has started increasing towards the right side. I mean if I can say that >> but what whenever a new file comes for example if you have four numbers already >> the fifth one instead of getting added at the end it should go to 08 index that is the first index.

[00:13:58]>> So how would you do it?

[00:13:59]>> So I don't know but >> no no worries no worries it's so let's go to the next part.

[00:14:04]Should I go to next function or explain this also?

[00:14:08]>> Uh so I think we have covered some portion of the code right. So I will ask you a few questions. Basically we I see there is a docker compos yml file.

[00:14:17]>> Yeah.

[00:14:18]>> So why do I need to use docker at all?

[00:14:21]Um if I remove docker what happens?

[00:14:24]>> So I have used it to deploy it but my laptop don't support docker.

[00:14:30]>> Uh acha. Okay. So I mean why do you need docker? Maybe let's talk start from the basics. Why would someone actually use Docker in their applications? So what happens when um I have installed some dependencies uh some requirements all the uh or the what we can say Python versions all the files versions and you have I you have to uh what do we say clone that my uh project in your system and what the versions that I have used in my uh this project might now have might not be in your uh laptop or PC whatever And do docker use that uh all the stuff at the once. So whenever someone use the docker uh what you can say docker gets all the information at one place. So whenever you uh install docker on your PC clone my project uh it would not give error that this python version is mismatch or something.

[00:15:30]>> That's true. I mean actually even though answer was longer it's what it does basically if I can put it simply right.

[00:15:37]Can I can I say docker provides a sandbox environment >> with certain predefined packages and everything you know already defined so that wherever you deploy it you get the your application gets the same environment right so >> okay yeah yeah right >> exactly I mean whatever you said also you know it's related to the same thing uh but let's go to the next one so I think now at least I got some understanding on the application logic perspective right this is you know a engineer or fullstack engineer anyone does the same thing but let's go to the a part of this application. So can you explain your rack pipeline to me how it currently works?

[00:16:13]>> Uh should I explain or show code?

[00:16:15]>> Uh first explain then show me the code like how it exactly >> I would explain from the uh top to bottom what happens. So what happens the when it would check for the in the chrom whether the uh the fetch items are there in my uh files or not then only retrieval happens. Yeah, there are no file no files then injection injection happens. It would inject the all the files and then the retrieval part is going to start. So it uses the a for the chunking where the abstract syntax tree which is the specifically for the Python code whether which chunks the files into functions and classes.

[00:16:57]>> Okay.

[00:16:58]>> Yeah.

[00:16:58]>> So what is your vector database currently? Uh I was chroma maybe.

[00:17:03]>> Okay. So first obvious question is why chroma? Why not any other vector database?

[00:17:09]>> So first reason was the uh it is uh it can be used locally and good fit for the projects or the prototype. Okay.

[00:17:17]>> And yeah it it fast on the local also.

[00:17:21]>> Okay cool. Uh so how would the retrieval happens currently? Is it just semantic similarity or do you have something else as well? So I have the hybrid uh search.

[00:17:32]So you you use sem semantic similarity also and a keyword search.

[00:17:37]>> So for keyword search what are you using currently which metric 25 >> and uh for example you are storing the embeddings right in vector database. Uh but for the keyword search where are you getting the chunks for or from?

[00:17:51]>> What side? So for example you're chunking right and you're embedding those chunks and storing it in vector database and using semantic similarity you're going to get either top case chunks from it right >> that is the semantic part but for the keyword search part where are you storing uh the chunks or the data are you getting again from vector database or you are getting from somewhere else >> no so from the vector database it works the hybrid approach worksh together it would get semantic ically and the keyword search which make it more precise to use.

[00:18:26]>> Okay. Then I have one question.

[00:18:28]>> There is a file called BM25 Wishnob Bosler right in codebase index.

[00:18:34]If you can you open that file. So what is this file?

[00:18:37]>> So I'm not getting uh >> so basically what is this file? Why is there why is this file present in this?

[00:18:42]Uh so uh today I have check uh tried to apply my PR agent on my one of my friends project which I have for it that time it has created I don't know what it does for now >> no worries no worries just a curious question because it has some relation I feel like with BM25 okay uh uh but anyways you further research on top of it >> uh let's go to the DB or database part of it what database you're using for application currently >> should I go to database files?

[00:19:14]>> Yeah, >> sir. I have SQL almighty.

[00:19:19]>> So can you just explain this code? Not to induct but what is exactly happening here?

[00:19:23]>> Uh so what happens when the uh first my pent started it's first initialize uh database checks uh checks for the database is it working or not?

[00:19:35]>> Yeah.

[00:19:36]>> Okay. So if I I mean as I'm looking at it right >> I see there is try you are yielding a DB object finally and you are doing finally. So what is finally doing there basically?

[00:19:49]>> So I didn't get it. Sorry.

[00:19:50]>> So in get DB yeah that particular function there is a finally statement right. What does finally do generally?

[00:19:57]Where do you use that particular statement in Python? I'm >> sorry I'm not able.

[00:20:01]>> No worries. So basically what finally does is right finally is nothing but at end of everything for example if you're running a specific function right after everything it's going to run. So what you're doing you are doing db.clo closed that means after every functional things are done finally we'll run this particular statement and your DB connection will be closed so it's not leaking no leaking the memory and all those issues that come up with right open DB connections >> so that's what is happening here uh so that is a simpler language or top level answer but there is much more nuance in top of it so let's go to the next section of this interview okay >> so this is a good very good project which is using rag And there is a you know hybrid system for retrieval and all those things which are good standards.

[00:20:49]>> What I think of is I have to go more depth in Python itself.

[00:20:54]>> That's fine. That's fine. There is no issues. Okay. At the end of the day even if you answer like one question correctly if you're sure about it that's great. If it's not it's fine because then you will be learning on this top of this right. Basically wherever you went wrong you are going to answer those questions. So there is no issue in telling wrong answers. it's more important to correct these mistakes and learn from them. Okay, that is the only thing. So, let's go to the next uh part of this interview. So, the next part is basically I want you to use AI uh if you have an ID already, right? Maybe I don't know you might be using Visual Studio only or you might have some other maybe codeex or something and create a small chatbot application but I want you to ask a model to create this application and while you are creating it let us you know have a back and forth conversation.

[00:21:43]Okay. So you are going to use copilot or something.

[00:21:46]>> Uh no sir uh I have open code.

[00:21:48]>> A okay >> can I use it?

[00:21:50]>> Yeah you can use any anyone no issues.

[00:21:53]Okay so the goal is to create a chatbot.

[00:21:55]It could be any kind of chatbot you want. But before you actually build it, tell me what kind of chatbot you want to build and also we are going to discuss more on for example how would you design rag architecture for this chatbot right so just you have back and forth conversation with myself as well as think how you're going to ask AI to build this application so that at the end of the day because interview is very short interview right >> previously people used to ask you to write a file that would be have been one file you would have taken entire interview now the because of AI we got this chance where we are building applications in the interview itself because it's that fast so I want you I want to see how you actually work with AI and how you come up with this kind of application so just talk to me as well while you're typing >> so what is the problem statement I have to use uh make a chatbot >> yes >> that uses rag or >> any kind of chatbot yes >> okay we can say or we can say the goal our goal is is to create chat What requirements such as? So I would give it uh requirements mile like uh it should retrieve the uh documents. Okay.

[00:23:17]>> So first question is do you think it's a good prompt that you have given currently? Uh no sir what I should have mentioned uh so what I was thinking like uh I should give chatbot the more context about what I'm going to be like I should give the file structure what exactly my each file has to do and then get the uh code from it.

[00:23:39]>> Yeah. So by for example if you can give the field right basically is it a financial quand or it could be medical chatbot or whatever based on the field itself chatbot architecture kind of changes here and there right so yeah but anyways we got some answer so can you explain me what it is saying >> okay so it has given the architecture uh so it says the uh from the data we can extract the raw docu documents such as PDF txt files or any other docs and it start with the injection pipeline where the our documents would get uh chunks and embedded using any embedding models and storing the vector database. Then comes the uh retrieval pipeline where the uh it will retrieve uh from the query of the uh user and uh gives the relevant chunks to lm to generate the answer.

[00:24:37]Then it comes to generator where the llm's call the prompt plus context uh goes to the uh generator where we get the final response from it. So it has used the orchestr which is the main part which should call the functions one by one in sequential manner.

[00:24:56]>> Okay.

[00:24:56]>> And it has also give the config uh folder which we have created earlier uh in my project. So currently will it give me a you know hybrid drug or will it only give me?

[00:25:10]>> So sir before that I have to mention him that which technique I have to use to deter.

[00:25:16]>> So currently is it already deciding what kind of rag it is going to do based on the answer it has given.

[00:25:21]>> Okay let me check it has given tech. No sir I'm not getting what you said. So question is did it already decide the methodology or how it's going to?

[00:25:34]>> We have to decide what's currently you say it hasn't decided it right.

[00:25:38]>> Yeah.

[00:25:39]>> Okay. Uh let's go to the next step. Next what would be our prompt or we have to use the junking strategy here. What we can use uh for now we can use the fixed size uh chunking and later we can we improve it and for embedding.

[00:26:18]So while it creates right so what is the context length for the model that is currently answering to you >> the context length >> so we haven't decide the context l for the model >> no what model are you using currently in this chat >> so I haven't uh I haven't using uh Germany Germany >> so what would be the context length of the Gemini model you are going to use >> I don't know exactly >> no worries because based on the context length right your questions might change. You can ask much more broader questions because if it has more context length, you can get more modules at the same time. Right? Once you get the model again, you can run some kind of testing or ask it to review the code that it has generated. If the model itself has small context length, you might need to go one by one. Maybe talk about rag and ask it to create rag and then go to the next section. Right? So that is good. Uh let's see what it has created. So can you explain me what is done in this round? Huh? Uh let me get so I have to explain file or something.

[00:27:18]>> No, no, just what was uh the you know work that has done >> okay though it has started with the injection pipeline where we have to load the documents. Uh it's uh it's comrade txt files and all this stuff and we have done the fixed side chunking is had use the,000 words of chunking splits. So we can use here overlap. Huh? It has done the thousand character chunks with 200 words overlap. If the our data is on the half of the page and next is on the half of the page, we don't get the context in the,000 words itself. We can use the overlap for the 200 to get the more context and okay. And for the chunking documents, it has uh attached the metadata for the source file. For example, um if the we have chunks the documents in the smaller parts, it would gives the id to each smaller part. So whenever uh so user queries about the uh something about the machine learning part and we have we get the smaller chunk relevant to it and then while giving it to uh LLM we will give the full part of it uh using the metadata ID.

[00:28:36]>> Okay, got it. Can you go to ENV file once?

[00:28:40]>> Yeah.

[00:28:41]>> So obviously we'll supply Gemini API here. So before that uh what are the file types that our current application supports?

[00:28:51]>> File types.

[00:28:54]>> File types means uh the data we are going to give.

[00:28:58]>> Yeah. What kind of file types? For example, py txt txt files. Uh we can also use the uh doc documents files.

[00:29:08]>> How would you know if the code is on the left side? Which file would you? search.

[00:29:13]So I'll give you option basically only you can open one file. Okay. So which file would have information where what kind of documents this particular application can chunk actually >> uh txt file.

[00:29:26]>> No no the code is there on the left side right so you have logic you can I can ingest any file I can put my image also.

[00:29:34]So you that's good you have clicked the inest.py there there is an option right what files it allows. So can you tell tell me what is the logic it has written basically you know what files are supported.

[00:29:46]>> Okay so yeah for now it's supposedly txt.py.js.ts CSC CSV files PDF is not yet supported we have >> so for example if I upload a file >> which is not in this list or tupal or whatever it is. So what happens when in the application what happens in the back end if I upload something else? Okay, for now uh it gives the value for the PDF. But what I can see the it it gives the exception where it skips that file if I uploaded the file that's not mentioned in here.

[00:30:20]>> So for example, I have uploaded maybe in a new file type it's called XF. Okay. So I'll come to try statement. If it sees it's not present there, so it will go to LF. It sees ext is equal to PDF. It's also not PDF. So it will again go to else now continuous right. So should I need that continue statement at all is it required?

[00:30:44]>> Um what we can uh I think so this statement is not needed we can do is the uh LCF here and if the mention file is not present we can directly skip or raise the error that uh this file is not supported. So let's see for example if the file is XF can you just run me through what particular lines of code will run >> dot XF >> okay so what will you do uh it will check the for here >> that if a file is actually there it will the path of the file dot and it will that also check if it is a PDF or not for now and so it will raise the value error that it's not a PDF file then it will else continue So code will the continue and it will give the exception that the file is not uh as mentioned above.

[00:31:35]>> Okay let's stop there what this continue statement does.

[00:31:41]>> Uh so as we see uh if these files are not present KSMD MD it will go to else if PDF not uh if that file is PDF it would say PF PDF not supported then it would continue the code and check uh check. So continue means it will again continue from line number 27. Is that what you're saying?

[00:32:01]>> Uh yes.

[00:32:03]>> Okay. Uh let's leave it there. Okay. Not an issue. Let's go to the the can you come down slightly to the uh line 38. So yeah. So here we have fix it size chunk whatever the logic you have given and chunk documents also. So can you explain me the line number 52? What exactly is happening there?

[00:32:20]>> Okay. uh questions for dogs and sorry I'm not getting >> no worries just read through this particular line what is happening can you read it out loud >> yes um so l for foreign text it will goes through that uh chunk chunks pi chunks checks if the raw chunks and if it >> so raw chunks is what basically >> um the chunks it has uh taken from the documents >> okay So what is enumerate doing for those chunks?

[00:32:54]>> So I'm I'm not aware of >> no worries. So okay. So let's continue with the chat. So I think some of the application is already ready. Let's see.

[00:33:17]Okay. So one thing I love is even in the real professional work life right where you're writing a lot of complicated code it's always good that you are asking it to explain what it's going to do basically what's the plan for it is before you know actually the asking it directly to give the code right it's also good if you can just give some kind of technical documentation or some kind of system design you have already decided if you can give it it will come up with much better code so that's what we do because even now 90% % or even 80% of code is being written by a in even good product company. So writing with a is not an issue. It's more about the understanding whether what it is exactly writing whether it's following what we are saying it or not. Okay. Uh cool. I mean let's stop it here. Okay. I think our application is at least at the initial level. Uh so now the question is more about the next point. So you have I have seen you working and creating an AI application using AI and also we have discussed about the code you have written. Now let's come to the point where we are going to discuss about your performance. Okay.

[00:34:27]>> How did you perform? So let's start with your opinion. How do you think you performed? What are the areas you think you need to improve? Uh you can stop the sharing now. Um and what are the areas that you know you think you performed well? Then I'll be giving my review. So what I think is that I don't have the actual knowledge of the Python uh syntax or what it actually is Python libraries where where I have lack exactly and the main problem is that I'm not able to explain uh things properly in my what I was supposed to explain.

[00:34:59]>> Okay. Uh anything else do you think any place where you have actually performed well?

[00:35:03]>> I don't think so. what I have prepared uh for back two back in two days uh I haven't per performed that well yesterday only I have had my one mock interview there I have perform well I don't know what happened >> no worries so the reason I asked you right first thing I'll be honest your performance is not good okay because I don't have any problems with you no not knowing packages these things that is fine I can get explanation for packages but actual the logic of the code what is happening For example, if there is any finally statement, if there is continue statement. These are all base Python >> keywords, right? And secondly, the logic itself. For example, how the chunking is actually running in the code. That is very very important to understand because AI will give you code very beautifully. Sometimes it will check if there are any errors. It's it checks that npixes syntax errors. But logical errors are something which AI can also do quite a lot. And even if it does only once that once is enough to break your production application right. So every line that is being generated by a you should be having the ability to to at least read and understand it. That is first thing and secondly I mean good point is you are able to explain the concepts at least well for example what is rag how these kind of you know techniques work but the bigger issue as you mentioned is the understanding of the code that a is you know writing.

[00:36:28]That is first thing. Communication wise it's good. No issues. Communication doesn't mean you need to speak entirely or fluently. Some interviewers are also give you option to talk in in a much slower way or you can if the interviewer knows Hindi right you can also talk in Hindi. That's fine but you are able to at least explain to me what you are thinking in your mind. That is good.

[00:36:48]Okay. But obviously the more uh fluent you are the better you know you can talk. So in short amount of time you can explain more. That's the only positive you get coming to AI topics because your profile and everything is related to this chatbots or rag and all those things interview has diverted into that direction. Generally in real companies interviews one round will be like this another round would be completely you know Python problems maybe they give you two problems or three problems and one more round would be just managerial round right where they want to understand what is your aspiration why you want to come into this company what is the projects you have worked on and all those things also happen right so never think that one interview if it goes wrong it means every interview goes wrong okay >> okay >> because this is one interview where it was more focused on you understanding the code part >> code >> even if you to see the channel itself there are different kinds of interview right >> so every now every time interview changes so no issues that is first and really really you know all the best in the coming upcoming interviews for you now is the time we can you know discuss some questions that you have >> so I want to know how can I improve in this uh two three days better so that I would able to >> okay so one thing you could improve is whatever projects you just understand the code. So just maybe ask a itself to see you know explain few of the keywords that are present there and when you are going through the code if you don't know what enumerate is right just ask it what enumerate is actually doing these are just five or 10 keywords but they are very important it's not like you don't even know keywords then it's going to cause issues see continue what it does it go to the next line immediately or it skips entire thing and go to the loop again so different things happen for different keywords right so understand that first the keywords and logic whatever the logic you have written whether it's actually making sense third thing whatever uh external packages or external technologies you are using one question I have asked about is docker right so if you're sharing the screen and interviewer sees docker they want to know why did you use docker do you know if the docker is actually useful do you know what are pros and cons with docker so that means whatever you put in the code you need to know I'm not asking anything more Just whatever you put in your code, I want to know if you know it or not. Okay, that is first thing you need to learn. So cover this in the next one or two days. Thirdly, the technologies, right? BM25 or uh whatever like the methodologies have adopted just have some kind of refresher on top of it. Uh so how how much percentage would you go to uh know would you give to semantic retrieval and how much percentage would you give to keyword retrieval? Like how would you decide those percentages? these kind of questions that can come up on your resume. So that means give a resume whatever you have written every single line of word in less 3 days or whenever your next interview is just give some time and maybe ask GPT right I mean that's the easiest way you can do it to ask some questions on your regiment >> so so that you can prepare that as well and thirdly from now on try to at least write few codes by yourself >> don't write everything with AI at least till you get comfortable with syntax maybe write one or two functions okay then you can give the remaining then obviously once you are very comfortable in understanding base Python structures again you can use AI no issues >> okay >> so entire theme is you need to understand the code that that is the theme of this interview okay any other question >> ra concept but what I think is I have to improve on the python side >> exactly >> got it >> um any other question uh >> no you only tell any other side where I have to improve myself So I would say this is the one first if you do this exercise for next few days that would be great but if you are targeting for good product companies right it's also good to have some kind of base knowledge on DSA it's not like you don't need to buy hard it but try to implement few sorting algorithms and just try to solve easy to medium lead code questions not for you know all the companies but if you are actually going for an interview in a good product company which is actually asking DSAs in few of their rounds it will help you at the end of the day right it's a problem solving skill that is also something I would refer. Other than that, yeah, I think we discussed everything uh that you know makes sense for you in next 3 days and also in broader future pipeline perspective as well. Uh so thanks Vishnov really have a great day. Thanks for attending this interview and wish you all the best.

[00:41:25]>> Thank you sir.

[00:41:26]>> Bye-bye.