Transcript for transcript_1xhkrALy5bBUcz8tYUb-8Ojv4qAmvSNt0_1921_3840_cbb0530277ef_9.txt

transcript_1xhkrALy5bBUcz8tYUb-8Ojv4qAmvSNt0_1921_3840_cbb0530277ef_9.txt

[00:02:28]  Among the candidates have joined
[00:02:33]  Hello, how are you?
[00:02:36]  Hi, sir.
[00:02:37]  How are you doing?
[00:02:39]  I'm doing good, sir.
[00:02:41]  Thank you.
[00:02:44]  So let's start with you introducing yourself and then following by, let's say, any project that you have worked with that you feel that is worth mentioning.
[00:02:58]  Yes, sir.
[00:02:58]  Thank you for the opportunity.
[00:03:00]  I'm Suhas Palappa, a computer science undergraduate at Institute of International Engineering College, Hyderabad, with a D.S.
[00:03:06]  And CGPA of 9.13.
[00:03:09]  And apart from the college academics, with my interest to learn new tech stack and modern technologies, I have taken a course inside the NextWeb Academy.
[00:03:18]  And starting from the basics of full stack, like from the HTML, CSS, to the advanced courses like React, Notions, and till date, I'm continuing with these courses.
[00:03:32]  And also, apart from these courses, I have been pursuing more courses like generative AI, large language models, and also a part of
[00:03:41]  let's say some data engineering type.
[00:03:45]  So based upon these courses, I have learned various technologies and I have done projects based upon the MUNSTAC, based upon the generative AI, and based upon a part of large language models.
[00:03:57]  So, and coming to the projects.
[00:04:00]  So coming to a generative domain, I have projects in automation workflow and part of large language models.
[00:04:09]  Like for the example, in the automation workflows, I use N8N, an open source tool to actually design the workflow.
[00:04:19]  To example is a DTS resume analyzer.
[00:04:22]  So where it is a telegram board that we build to analyze the resume, to screen the resume.
[00:04:28]  It provides the candidate score and the limitations and what are the corrections present inside the resume.
[00:04:35]  So it gives a clear idea to the candidate.
[00:04:38]  So that is one of the generated AI project which have built using N8N workflows.
[00:04:44]  And coming to the LLM, a project which determines the interviewer, AI interviewer, generally built using the short-term memory, like because of some API issues and limits.
[00:04:57]  So I've generally used short-term memory, Langchain,
[00:05:00]  and also Langraph to work with multiple agents and also integrated it with text to speech and speech to text conversions using assembly air and also the
[00:05:13]  the AIs for speech to text conversion.
[00:05:16]  Let me hand it.
[00:05:21]  I will forward the exact name.
[00:05:22]  But yeah, and for the memory, in-memory AI, it is generally a shorter memory used to store conversations.
[00:05:31]  So due to the limits, I only included the five conversations.
[00:05:36]  Like the AI will ask only the five questions to the candidate.
[00:05:39]  And using the blockchain and the RAC concept, retrieval documentary generation.
[00:05:44]  So this maintains this state accurately.
[00:05:48]  And coming to my experience, I have experience in two companies.
[00:05:53]  First one is at Port Robotics, where I worked as a trainee as a software developer.
[00:05:58]  And my role mainly was as a tester, where I tested the ongoing project, cooking automation, where I fixed the bugs which are causing the errors.
[00:06:09]  And my second experience was at NIT Varangal, where I worked as a research intern, where our task was to build some machine learning models that could detect early detection of some diseases.
[00:06:22]  And a disease given to me was renal cell carcinoma.
[00:06:25]  It's a kidney cancer.
[00:06:26]  So I worked there for almost three months.
[00:06:30]  and getting a decent accuracy.
[00:06:34]  Okay.
[00:06:36]  You mentioned Langchain and Langraph that you have worked on.
[00:06:39]  But what use cases you have worked on?
[00:06:43]  Right.
[00:06:45]  Coming to use cases, they were like dependent on my learnings and coming to the project in the project point of view, like Langchain, I have mentioned, I have taken the use case like the steady assistant as a basic project.
[00:07:00]  So we have built an LLM.
[00:07:04]  So where I've used some agents, I've created agent using external tools, memory, and also the tool decorator to
[00:07:14]  get some tools from external sources or to build in tools such that the core logic can be defined using a Python function.
[00:07:24]  And then I've used some memory AI tools.
[00:07:27]  And this helped me to get a good experience how to build the agents using the Langchain.
[00:07:34]  And a part of Langraph as well, it's a top of AI Langchain.
[00:07:38]  So to actually configure better on multiple agents, as I said, multiple tasks can be taken at the same time.
[00:07:49]  Okay.
[00:07:51]  So when you mentioned, let's say, LLM or an agent, so what is the difference between an agent versus an LLM model?
[00:08:01]  So coming to a difference, so LLM, it's a large language model.
[00:08:05]  It's a deep learning model generally trained on massive or complex data, higher amount of data, such that the responses could be taken from those data and that could be generated to the user.
[00:08:20]  And coming to the agent, it is an
[00:08:25]  Virtual assistant generally built using tools, memories, and external tools like using APIs, databases, memory.
[00:08:35]  So it says that it can perform action, it can take its own action, and it can generate what I mean, it can complete that given task to it.
[00:08:47]  But what is the processing for an agent?
[00:08:51]  How does it decide what to do?
[00:08:55]  Sorry, sir?
[00:08:56]  So like how an agent, let's say you mentioned that agent is a virtual assistant and he works to complete a task given to it.
[00:09:06]  But what is the core?
[00:09:12]  Like the intelligence part in an agent, where does that come from?
[00:09:17]  So the LLM performs intelligent, so it generates the responses.
[00:09:22]  So using Gemini API key or Gemini model or Croc or any other model.
[00:09:28]  So we generally create an LLM inside it, pass the LLM as a source, like a model to generate the responses to the user.
[00:09:38]  So that creates the content or generates the content.
[00:09:44]  So what kind of memory that you have used in the Langchain along with your LLM?
[00:09:52]  So, yes, sir.
[00:09:53]  So in the specific project, so due to the API restrictions, like we have limit free credit limits.
[00:10:01]  Using short-term memory was valuable because of some limits.
[00:10:05]  So in memory server, so it is generally a framework inside the language.
[00:10:10]  So we generally imported it inside the code.
[00:10:17]  We used it to save the memory.
[00:10:18]  So in the project, we generally included five
[00:10:22]  conversations between the AI agent and the user for that normal random interview.
[00:10:27]  Like interview was based on subject as well, like HTML, CSS, Python, self-intro.
[00:10:33]  So for each subject, there were only five questions.
[00:10:35]  Like the agent could first ask about a question and based upon the RAG, like retrieval augmented generation content.
[00:10:44]  So it would ask based upon the previous responses and it could generate the
[00:10:49]  new question.
[00:10:50]  So these are all included in the short term memory because it is more valuable because storing only five questions is not suitable for long term memory.
[00:10:59]  So it's efficient.
[00:11:02]  So but for let's say five question, why did you implement a RAG setup?
[00:11:09]  Because if you have into a rack, like what was the embedding that you used and how did you?
[00:11:18]  Yes, sir.
[00:11:19]  So like as from the project point of view,
[00:11:27]  First question, so the memory, like,
[00:11:30]  Sorry, sorry, the RAG point.
[00:11:32]  So RAG generally retrieves the information from the vector embeddings from the database.
[00:11:39]  So like if a candidate answers a question in the interview and the agent needs to ask an additional question or based upon the previous question,
[00:11:51]  So this question needs to be matched.
[00:11:53]  I mean, the similarity search must be happening between the question and between the answer.
[00:11:59]  So the agent needs to think or search for the question and ask such that it doesn't mess up or doesn't make any other, I mean, any other type of question.
[00:12:09]  It should be at least a similar kind of question, like with respect to the subject.
[00:12:13]  So the RAC concept generally created such good illusion because it could reduce the hallucinations.
[00:12:22]  Like if an agent asked a question related to AI and suddenly the agent misconfigured and asked a question related to some databases that doesn't add up inside the generated AI interview setup.
[00:12:37]  So using retrieval augmented generation, so it reduces those hallucinations from the AI.
[00:12:43]  It makes the decisions and performance accurate and it focuses on the content.
[00:12:49]  And coming to the memory, sorry, the question was?
[00:12:55]  What kind of embedding that you used?
[00:12:59]  Yeah, coming to the memory.
[00:13:00]  to the embedding.
[00:13:00]  So there was a framework generally imported inside the, like we use Google Colab to go down.
[00:13:10]  So using this framework, we generally use ChromaDB as our vector database to store the vector embeddings.
[00:13:19]  So these embeddings were converted, sorry, the text was converted into numerical
[00:13:26]  vectors having the embeddings having various numerical values.
[00:13:30]  So based on the similarity says those all embeddings were performed and this
[00:13:36]  similarity meanings or similar meanings were actually brought up into the project for dynamically asking the better questions to the candidate.
[00:13:49]  So when you say, let's say in your case, you mentioned that you're using RAG because you want to ask users similar type of question, right?
[00:13:57]  Yes, sir.
[00:13:59]  So, but wouldn't this create a loop for the agent?
[00:14:04]  Because let's say he has asked one question, let's say related to state management.
[00:14:11]  To the user and you have let's say similar question like very similar question but just related like in a different words or different
[00:14:24]  format wouldn't your agent gonna ask the same question again to the user if you're using for to finding similar question instead of follow-up questions
[00:14:35]  Yes, sir.
[00:14:36]  So it's not about like completely rag based, like it's a part like it's a it has a let's say a 5% role in that.
[00:14:45]  So asking the follow up questions is mandatory, like it is placed like the rag concept generally came here to actually make the agent does not go out of the context.
[00:15:00]  context window.
[00:15:01]  Like instead of wasting, like it only has the opportunity to ask five questions and those five questions must be meaningful.
[00:15:09]  So the RAC concept was just a bit and most was based upon the blockchain tools and everything.
[00:15:16]  But the main motive there was to actually
[00:15:20]  make the agent stand on its single path, like single domain or single subject.
[00:15:27]  But if you are saying that you are using RAT for maintaining context, right, to limit the context and make the agent perform in this context window, right, don't go out of it.
[00:15:40]  But does this, like, what RAT is used for, like, what is RAT?
[00:15:47]  Yes, sir.
[00:15:49]  So, RAG is useful when you want to have a lot of data, which is you want
[00:15:58]  to be loaded in the context memory conditionally when we want to search for it, right?
[00:16:03]  Yes, sir.
[00:16:08]  So let me like, yeah, let me clear it out.
[00:16:10]  Like seeing the RAC concept is true, but in this project, like the RAC concept was actually a bit like, it's not completely based on RAC.
[00:16:22]  Like I included it because to maintain some relation between the each questions as that the user can, or the candidate can go in as flow.
[00:16:32]  But the major part will be taken care by the LLM and the blockchain to generate the response, to generate the question, to give the feedback to the user based upon the
[00:16:45]  speed, like the responses from the user.
[00:16:49]  And yes, sir.
[00:16:54]  Okay.
[00:16:55]  So, uh,
[00:16:58]  I think on the embedding part, you didn't mention which model or thing that you use to create the embeddings.
[00:17:06]  Yes, sir.
[00:17:06]  So embeddings part, I don't remember the exact framework, but yeah, we imported, I remember I imported a framework.
[00:17:19]  Embedding framework like I don't know exact name but yeah I've included the embedding vector search
[00:17:30]  Vector embedding, it's a framework I imported inside it and it has a built-in method and using that method, so the text or the token used to be converted into its vector in a numerical form
[00:17:44]  such that this can be stored inside a database of vectors.
[00:17:50]  For performing similarity searches, chunking and other operations.
[00:17:54]  So how a similarity search actually works with a vector?
[00:17:57]  Let's say you have 200 vectors in your database and now you want to find a similar vector.
[00:18:07]  Like what is the, it works, how does it work actually?
[00:18:11]  Yes, sir.
[00:18:12]  So similarities are generally work based upon the embeddings that were created.
[00:18:17]  Like if there are two words, if let's for easy understanding, let me take two.
[00:18:22]  words A and B having similar meaning.
[00:18:24]  So if the LLM process after embedding the A word into the database, it has its own vector numerical representation of a vector having various decimal values.
[00:18:38]  Now the second word which has been like
[00:18:42]  the second word, the B word, which is generally embedded, will have the close embedding values, numerical values, with respect to the first one.
[00:18:51]  Maintaining the similarity, like that numerical values having close values, like if the values are like 0.81 and other one has 0.79.
[00:19:01]  And so on there are various embeddings in the numerical vector values so if the values are too close and too
[00:19:09]  close to each other like having almost the same value so this determines that there is a similar meaning between the two words and it upon
[00:19:20]  generating the same context, I mean same word.
[00:19:24]  So it could use the embeddings to generate new responses based upon the understanding the actual meaning of the words based upon the meanings, understanding the meaning based upon the other word and this word.
[00:19:39]  So it is just based on word or does it take the entire sentence in account?
[00:19:43]  Let's say you have 10 sentences, each has what in them.
[00:19:47]  So would the similarity score come same for all questions?
[00:19:51]  No, sir.
[00:19:52]  So that varies.
[00:19:53]  That will be like it will be based upon the sentence like meaning.
[00:19:58]  So.
[00:20:00]  Like, let's see.
[00:20:03]  Does it work like it calculates word by word?
[00:20:08]  Or because that would be very slow process.
[00:20:13]  No, sir, that wouldn't be a slow process because it has its various numerical represented vector has a new various numerical values represented inside it.
[00:20:23]  So each value, each number has its own like identity.
[00:20:28]  So which takes it near to its similar meaning word inside the database.
[00:20:33]  So it doesn't make much time to search for a word or place a word.
[00:20:42]  So that really depends on the two, three factors of vector embedding and the search might things work.
[00:20:53]  So the first thing which I was trying to ask you is how the vector search works.
[00:20:57]  So basically, the similarity score is find out based on cosine similarity.
[00:21:03]  So between the vector embeddings that you created.
[00:21:07]  So each embedding that you're creating, let's say it gets a representation in a 3D space, right?
[00:21:14]  So if you have heard, if you have learned 3D geometry, right?
[00:21:17]  Yes, sir.
[00:21:19]  So each of your embedding is a vector in 3D plane.
[00:21:22]  And what you do is you find out a cosine similarity, which is between the like how much divergence is between those vectors.
[00:21:31]  So basically how the cosine.
[00:21:35]  Angle between them, all that stuff.
[00:21:37]  So mathematically, that's the work, how it works.
[00:21:41]  It defines similarity.
[00:21:44]  And the speed and all also depends on first is what kind of embeddings you are creating, right?
[00:21:52]  You, what are these,
[00:21:55]  Let's say you have a 10,000 page document, right?
[00:21:59]  You cannot be creating, embedding for each of the sentences, right?
[00:22:05]  When you have to find something.
[00:22:07]  So you chunk it down.
[00:22:09]  create different summaries, let's say you might be converging the old responses into a summary of your might be converging different paragraphs and creating
[00:22:23]  embeddings for each paragraph and trying to then find maybe a similar paragraph to it.
[00:22:29]  Also, there are which model you might have used while creating an embedding that would differ in terms of how the embedding is actually created.
[00:22:46]  And you mentioned which DB that you used?
[00:22:50]  Chroma DB.
[00:22:52]  Chroma DB, okay.
[00:23:01]  Just to give me a second of evidence, getting a different.
[00:23:16]  So.
[00:23:20]  So ChromaDB is like a relational database or like a non-relational database?
[00:23:25]  At the core.
[00:23:28]  So ChromaDB generally have not gone much deeper into such type of database.
[00:23:37]  But I feel it is a non-relational database generally because it needs to store such a long
[00:23:44]  As you mentioned in a 3D plane, like there's such vector embeddings inside it based upon sin message, like having a plane, having the negative values or positive values.
[00:23:57]  So at all varies based upon the word, based upon the context.
[00:24:02]  So I majorly think it has a non-relational database.
[00:24:10]  But we are also mentioning...
[00:24:23]  So let's say you have mentioned adaptive questioning, right?
[00:24:25]  So what was this?
[00:24:27]  Just the prompt based adaptive questioning or you have like a defined an algorithm to how to find a similar adaptive question?
[00:24:37]  No, sir.
[00:24:37]  It's a prompt based questioning, like having a, let's say, best and good prompt.
[00:24:44]  Like I've designed a prompt giving to the LLM, like it's an agent again.
[00:24:48]  So that agent works on asking up the follow up question or adaptive question to the student.
[00:24:56]  So based upon previous context, so generally the follow up.
[00:25:00]  Question will generally vary based upon the terms like used inside by the student inside the interview system.
[00:25:09]  So based upon the same as vectors and embeddings.
[00:25:15]  So the similarity search will generally work on here.
[00:25:18]  So as the follow up question.
[00:25:21]  And the follow-up question generally is that depends upon the answer the user gave to the agent and the terms he has used inside his response.
[00:25:35]  Okay.
[00:25:39]  So, okay.
[00:25:46]  You also mentioned one project called Crypto Reconciliation.
[00:25:50]  Yes, sir.
[00:25:51]  What was that project about?
[00:25:54]  So Crypto Reconciliation, that was a project like an organization generally conducted a project.
[00:26:02]  Like we were given some transactions or data and we need to perform the and we need to design an engine like a back-end project so as that there's transactions we need to
[00:26:17]  That was only the backend project.
[00:26:19]  So we need to create some APIs.
[00:26:22]  So reconcile and reconcile.
[00:26:25]  What were you reconciling?
[00:26:26]  Like what were you reconciling?
[00:26:29]  Crypto transactions?
[00:26:31]  Yes, sir.
[00:26:31]  Transactions.
[00:26:33]  Like we got the data.
[00:26:36]  Like we got a data in the form of like Excel sheet in the CSV format.
[00:26:41]  So we generally use those data.
[00:26:44]  We need to find the matches.
[00:26:46]  Like we have two types of Excel sheets.
[00:26:48]  Like we have two types of data having the transactions by the same customer, by the same bank and by the same.
[00:26:56]  Like they were seen and they were like there some transactions were matched and some transactions were unmatched like we need to configure that like we need to create apis which will create some unique ids for each reconcile
[00:27:12]  process and
[00:27:14]  Upon this ID, upon running this through this ID inside the backend, so based upon that such ID, the backend should demonstrate how the number of matched transactions and matched transactions
[00:27:29]  and other in progress transactions like there were status, transaction status.
[00:27:34]  So it needs to mention all the like
[00:27:37]  the overall count or the overall, let's say, the summary of two transactions merged together.
[00:27:44]  So we need to create an engine and we need to provide them.
[00:27:48]  So for that, generally, I use Node.js and Express.js for generally on the server side and for creating the APIs for each one.
[00:27:58]  So coming to the reconcile first API, that was reconcile API, it would generate a unique ID based upon the transaction history.
[00:28:07]  So that unique ID will be then passed as an API key parameter.
[00:28:13]  So that using this parameter, we could generate the response, we could generate the summary.
[00:28:19]  So it's not like a generate, we could get the summary, we would get the response of all the transactions based upon that runner ID.
[00:28:30]  Okay, got it.
[00:28:34]  I think that's all from my side, sir.
[00:28:37]  If there is something, maybe a follow-up assignment or something, I'll let you know.
[00:28:41]  Okay.
[00:28:42]  Okay, sir.
[00:28:44]  Thank you so much, sir.
[00:28:46]  Yes, sir.
[00:28:47]  Can I ask you a question?
[00:28:48]  Yeah, please.
[00:28:50]  Yes, sir.
[00:28:52]  Searching your company in Google, I found some...
[00:28:57]  Detailed description and I got an excitement to know like there it was mentioned like this company major focuses on web technologies and generative AI and like agentic AI technologies to build solutions for the companies.
[00:29:11]  So what is the motive like how could you actually understand the
[00:29:16]  motive of various businesses to build the agents that are specific to their businesses or business logic.
[00:29:25]  Okay, so how does this actually work?
[00:29:27]  Let's say first part mostly consists of, because you can do let's say thousand things, right, in general, but
[00:29:36]  Most of them would not be used by that company or it's not relevant to that company.
[00:29:40]  So first usual step, which comes the most human part, right?
[00:29:44]  Let's say even these days, execution has been built up with AI, right?
[00:29:49]  You can write a lot of code with AI.
[00:29:51]  But the first part, which is the main important part, which comes with the consultation.
[00:29:55]  Where you understand the business requirements, you understand
[00:29:58]  like what would work
[00:30:00]  What is their current workflow?
[00:30:03]  Even the system would work in there.
[00:30:05]  Let's say you went, you're looking, let's say, just giving an example.
[00:30:15]  Went to let's say a factory right it's a manufacturing factory so you said i'll give them i'll give you a chatbot that let you know about
[00:30:24]  uh what is the operational uh efficiency of all your machines and blah blah blah which what are the shifts and all that stuff uh that doesn't create a very uh
[00:30:35]  good business use case for them, right?
[00:30:37]  Because that's a, you have to figure out what is a gimmick for them and what would actually move the needle for them.
[00:30:44]  But if you go and say, I'll automate different fragmented areas, which you're currently operating in silo into one integrated setup.
[00:30:56]  Then it makes them for a business.
[00:30:59]  So it's more about how you understand their business first and then apply your technical knowledge to solve a particular problem.
[00:31:07]  Okay.
[00:31:11]  Yes, sir.
[00:31:12]  Thanks, Ross.
[00:31:13]  Thank you, sir.
[00:31:23]  Suhas, you can drop off the meeting.
[00:31:28]  Okay, sir.
[00:31:28]  Thank you.
[00:31:35]  Aman, can we ask the next candidate to join us?
[00:31:37]  Sure.