What we talked about
Harshit Kohli is a Senior Technical Account Manager at AWS, where he serves as a trusted advisor to enterprise customers on cloud strategy, architecture, and AI adoption. He started his IT career in 2010 at Infosys in India, built his way through Tech Mahindra, HPE, and Cloudera, moved to the US in 2016, and joined AWS in 2022. He is currently completing a PhD in AI, has published over a dozen research papers on mental health prediction, churn forecasting, and healthcare AI, and presented at the MCP Dev Summit on real-time streaming to agents.
Show notes
Harshit Kohli is completing a PhD in AI while simultaneously serving enterprise customers at AWS as a senior technical account manager, a combination that gives him an unusually clear view of the gap between what companies expect AI to do for them and what it actually requires. His answer to that gap is consistent across both roles: the data has to be ready before the model matters.
What we covered
- A Technical Account Manager at AWS functions as the extended team of an enterprise customer, not a salesperson, not a pure architect, but a blend of both. Kohli describes the role as someone who “thinks about the customer’s business as their own business,” engaging at every stage from initial cloud adoption through workload optimization, cost stabilization, and long-term architecture decisions.
- The biggest misconception Kohli encounters about AI is that it has a magic wand. He identifies three distinct misunderstandings: that AI can solve everything, that everything is AI-fit, and that AI can fix bad data. He illustrates the last point with an example of 100 flower images where 70 are blurred, no matter how sophisticated the model, unclear input produces unreliable analysis, not accurate results.
- He published research at an international conference on predicting mental health issues from wearable device data, using Kaggle datasets to compare model accuracy and precision across different algorithms to identify which model performed best on that specific data type. He has also published on predicting diabetes and on customer churn using sentiment analysis.
- At the MCP Dev Summit, Kohli presented work on streaming real-time context to AI agents, solving a specific limitation of the request-response model that MCP typically operates on. The result allows an agent to receive continuous live data and respond to interactive queries like “how many anomalies occurred in the last hour” in real time.
- When judging hackathons, he described seeing ideas emerge around AI-assisted crop prediction for farmers, models that analyze weather patterns to help farmers choose which crops to plant given a forecasted climate. He said that five or six years ago, these proposals would have been considered assumptions rather than buildable solutions.
- His advice to engineers in India considering a career in cloud or AI: go deep on fundamentals, not just surface-level familiarity. He argued that understanding the mechanics of services, even without reaching expert depth on all of them, is what allows someone to connect the dots for a customer and earn trust. Certifications and courses are the mechanism, but “if your fundamentals are not good, it is not going to help you in the longer run.”
About Harshit
Harshit Kohli is a Senior Technical Account Manager at AWS, where he advises enterprise customers on cloud strategy and AI adoption. He has worked across Infosys, Tech Mahindra, Hewlett Packard, and Cloudera over a 15-year career spanning two continents, and is currently completing a PhD in artificial intelligence.
- LinkedIn: https://www.linkedin.com/in/harshit-kohli-99801543
- Website: https://amazon.com
Episode 152 of the PreVetted Podcast.
Full transcript
Federico Ramallo (00:00) Welcome back to the Prevetted Podcast, where we spotlight extraordinary people and remarkable talent reshaping our world. Today, I am joined by Harshit Kali, his Senior Technical Account Manager at AWS, where he serves as a trusted advisor to enterprise customers on cloud strategy and AI adoption. He holds…
a PhD in AI, actually he’s working on a PhD in AI, has published research across machine learning and healthcare AI and built his career across two continents over more than 15 years. Harshid, welcome to the show.
Harshit Kohli (00:39) Thank you, thank you, Federico, and so nice to be part of this podcast and happy to meet everyone. Thank you.
Federico Ramallo (00:46) I’m honored to have you here today. So for people meeting you for the first time, can you tell us a little bit more about what you do?
Harshit Kohli (00:49) Same here.
Sure. My name is Harshit Goli. As Federico introduced me with each one of you, I am currently working as a senior technical account manager at Amazon Web Services. I joined AWS in the year 2022 in November, and I’ve been an integral part of the organization since then. So I started my…
IT journey back in 2010 from Infosys back in India. And since then I’ve worked for multiple employers, like I’ve worked with Mercer Consulting, I’ve worked for Indian consulting firms like Tech Mahindra. And then over here I’ve worked for Ditchins Consulting, where I worked for Ford Motor Company, and then I worked for Hewlett Packard, I’ve worked for Cloudera, and now I’m working for AWS. So.
It’s been a journey where I’ve worked with multiple employers. And then while working with them, I also made sure that I’m also upscaling myself by, like I already was having a bachelor’s degree. So I went ahead and did the master’s degree in information technology, which was focused on big data systems. And now with this AI booming, I decided to have a doctorate in artificial intelligence. And that is where I think one and a half years back, I started my PhD.
And right now I’m still working over it. It’s almost like a second job for me because I have to invest lot of time and effort. But yeah, I mean, it’s like getting the benefits out of it. I’m trying to use the skills from there, apply it in my job and then trying to learn as I’m progressing. Apart from all of this, I’m also fond of writing research papers, mentoring folks, being a speaker, being presenting.
being like doing a lot of tech presentations, visiting a lot of conferences. Yeah, that’s about me.
Federico Ramallo (02:50) Amazing. Amazing. How was your experience moving from India to the United States?
Harshit Kohli (02:57) Yeah, I I moved to US in November 2016 and it was a completely different experience for me. I mean, always I’ve heard from my friends and known ones that US will give you a lot of opportunities. There is a lot of room of working through. You will learn a lot of things. It’s a lot of opportunities as we say, right? So when I came over here, I saw a of good differences.
the place where I was working with where I am right now. So it was sort of learning new culture, meeting new people, understanding new things. And then the way the tech evolves here is massive compared to any other place, right? So we have all the good and growing startups here, all the big techs here, big IT here, right? So the amount of ease of accessibility in terms of all the techs here is massive.
And that was the number one thing which I learned here and I really appreciate here. Plus the work culture is superb. I really appreciate the kind of things I’ve learned here and I have given a lot of good opportunities from my companies here to grow and excel. yeah, so it was a massive difference and I learned a lot in these years.
Federico Ramallo (04:16) Amazing, amazing. Can you tell us a little bit more about what a technical account manager at AWS actually do?
Harshit Kohli (04:23) Sure, so I will explain it in a broader way so that everybody can understand. So the way I will start here is, we all know that AWS has a lot of customers worldwide, right? So you have big customers, you have small customers, you have enterprise customers, you have strategic customers, right? So when you are dealing with big customers or enterprise-based customers,
you need a voice from AWS who can work in close collaboration with those customers and who can work like an extended team for those customers. Okay. And those will be the folks who will be working backwards from the customer needs and will be helping those customers day in, day out. It could be related to cloud adoption, AI adoption, optimizing their workloads, or making sure that cost is optimized.
you do regular kind of benchmarking with them, you do a lot of reviews for them. So it could be anything, right? So they could demand you things like reviews to solution designing to adapting more services and whatnot, right? So as a technical account manager, a person is needed who can hear the customer’s voice, who can convert those needs into the requirements and then can help those customers because
A customer cannot be at every place and for everything, if they are supposed to go and create a ticket or engage somebody from AWS, they will not know whom to engage and whom to work with. So that is where a technical account manager is an important element in the puzzle, where they work closely with their customers and being their strategic advisors so that they can get the most out of the cloud and they can adopt the cloud and AI services as they are progressing along with AWS.
Federico Ramallo (06:13) Right, right. And what is the difference between a technical account manager and a non-technical account manager?
Harshit Kohli (06:21) Okay, so there are multiple roles, right? So we do work as part of one team here. So what happens is we have account managers, we have solution architects, and then we have technical account managers, right? So when it is in terms of sales perspective, understanding the sales health of your customer, understanding how the customer is progressing, what’s the sales trajectory of the customer, how the customer is growing, account managers handles all those things.
contract negotiations and all those kind of things. Now when it comes in terms of solution designing or helping the customer with a solution or giving them kind of a, helping them with a kind of architected review, all those things fall in the plate of solution architect. Now technical account manager does most of the things which you can consider is like an extended team for a customer. So what happens with technical account manager is since it’s like an extended team of
the customer itself, they start from day one and they are till, mean, whatever the customer needs are, they will be there. They’ll be the first point of contact for the customer and they will be like, they will understand the customer needs and then that’s how they’re going to work with the customer. So it’s like a blend of both technical and strategic kind of a need where you understand the customer needs and then try to see what fits the best for the customer.
So as we know, not all t-shirts sizes can fit everyone. So that is where you as a person decide that, okay, if my customer has X requirement, whether X will fit or whether Y will fit. So you are kind of a decision maker for your own customer.
Federico Ramallo (07:57) Right, right, makes sense. I used to work at Microsoft doing technical pre-sales and it was, even though it was pre-sale, was a different environment, but it was a similar situation where we would help clients understand based on their needs, you know, figure out which Microsoft product would fit best for the requirements, right?
Yeah, in your case, you’re more involved on the implementation, right?
Harshit Kohli (08:23) Implementation, service adoption, optimization, it could be standardization. So a technical account manager what I see is it is fitting into every stage for the customer. If the customer is spending 10k and they want to grow technical manager is handy. If the customer is stable and wants to stabilize their workload technical account manager is handy. If the customer is already sitting at a very higher price and now they want optimization kind of things okay how can I make my cost stable.
I don’t want to have 50K increment bill every month, right? I want to make it stable. Now that’s where also technical account managers are handy because they will give you certain kind of optimization frameworks on which you can work and stabilize your bill, right? So from what I see, from what lens I see, technical account manager are valuable piece of puzzle at every stage of the customer, whether it’s a starting stage or it’s a very mature stage.
Federico Ramallo (09:18) Right, right. And you have one leg on both areas, the technical and the commercial, right?
Harshit Kohli (09:21) Yes. Yeah.
And that’s how you kind of understand your customer needs. So if you consider an account manager or a technical account manager, which is less in technical skills, that person is supposed to understand the customer needs not that efficiently as a technical account manager who is highly technical will understand. So let’s say if I am having AWS certifications.
And I understand the way how AWS services interact and work. And if a customer throws me a question, hey, you know how this particular service is going to interact with this service. Can I do a review of this particular service? Now, if you don’t know anything about that service, if you’re not aware about what the service does, then you cannot even think of engaging somebody. You will not know which person to engage, when to engage, how to engage, right? So for connecting those dots, at least you need to know the basic technical stuff.
which will help your customer and like help you to earn trust with the customer.
Federico Ramallo (10:20) Right, so you have a T-shaped knowledge, right? So you have general knowledge and then specific knowledge on some parts, right? Yeah. And then, you know, when you need more depth knowledge on a specific area, either you research or you find somebody in your team that can help you figure out the specifics, right?
Harshit Kohli (10:28) Yep, exactly.
Exactly. So like if I talk about technical managers and solution architects here, we are having a kind of a level 200 knowledge on most of the services. But if you want to go in depth, there are specific certain set of services on which we go in depth. Like if you take my example, I am very much into analytics, streaming, Gen.EI. So if it’s that kind of a discussion with the customer,
Either I can lead it myself or if it’s a very corner case where I think that, okay, I might not have that much amount of knowledge, I will engage a specialist for my customer. But for all these services, can very briefly and very in detail fashion, I can have a detailed kind of discussion with my customer. But there could be services where I might not have a level 300 or level 400 knowledge.
Federico Ramallo (11:16) Right.
Harshit Kohli (11:32) and that is where I’ll let my customers know that yeah I mean I understand you have this kind of a requirement maybe I’ll go back and check in for some other specialist who can help you.
Federico Ramallo (11:42) Right. Right. And it’s hard to know all the services in AWS because you have so many and the level of complexity is so high. Right. And it doesn’t make sense to know them all. What it makes sense is to understand how they work. mean, the best way that I’ve been able to explain this is you don’t need to be a mechanical engineer to drive a car. Right. You just need to understand how it works. Right. Now…
Harshit Kohli (11:51) Yes.
Yep.
Federico Ramallo (12:09) Right. If it breaks, right, you actually, a mechanical engineer would not be the best fit for fixing a car, you know, unless you want to design a carburetor injection system or things like that. Right. You know, but having a mechanic that knows the basics will be able to fix it faster. Right. And, you know, but if you want to fine tune a car, then yeah, you need a mechanical engineer. Right. So,
It’s different world for different situations. what I’ve been able to do myself is to understand things in boxes. I don’t know exactly how it works inside, but functionally understand how it works and where it fits in the whole world of everything else. And then I can drive the car. I can use it basically. And then if I need more…
then I can reach out to the experts, right? And you can be that, you know, that channel to the customer, right? Either you solve it or you refer it or you find it internally.
Harshit Kohli (13:09) Exactly.
Yep, exactly. So the way we consider technical account managers is that you are, as I said, the extended team members of your customers, and you can consider your customer’s business as your own business. So the way you think strategically, the way you think about your business, you have to think about your customer. And that’s where you help your customer grow strategically, technically, growth wise, everything.
Federico Ramallo (13:39) Right, right. So you have published research on machine translation, right? Healthcare, AI and turn forecasting, right? Can you tell us a little bit more about what your research was about?
Harshit Kohli (13:46) Yeah. Yeah. Yeah.
Sure. So I did around, I think, 10 to 12 research papers in the past one and a half years. So I started my research paper journey early last year. That is when I started my PhD as well. So I have a bunch of research papers scattered across various kind of domains. There are healthcare, financial as well. So the most recent paper, which I did as part of an international conference.
is with respect to understanding like how the, I mean, how the mental health patients, right? And I mean, how you can predict by having the wearable devices, how you can predict some kind of a mental health issues for the patients, right? So for that, we use a kind of a datasets. I mean, there are a lot of datasets available online, Kaggle dataset is available and there are a lot of datasets you can use.
And then based on that, you can utilize various models and you can try to see which models is working with this kind of data set. And you can see in terms of accuracy, precision, which model is giving you the best results, right? And then kind of you infer the results based on those findings that, okay, this particular model is working with this data set and is giving me the highest accuracy. And that’s how you’re doing a research.
on a particular data set using a particular model by testing various models ahead. So the recent one which I did was with respect to the mental health, finding out the way on how to accurately predict the mental health issues in the patients. And then in the past, I’ve did similar kind of research papers with respect to diabetes as well, predicting diabetes for the patients. And then in finance sector, did for…
I did a research paper on how to predict the churn of the customers, right? So there could be sentiment, right? Like for example, sentiment analysis is a big field within AI, right? So understanding and hearing those signals on what the customers are expecting and utilizing those sentiments and putting values ahead in front of those signals or sentiments is a way to find out, okay, if I get a sentiment for a customer saying good or bad, what this turns out to.
And based on that, you have a kind of a data set and then you predict that, okay, based on sentiment analysis, customer A is supposed to get churned or customer B is a safe customer. They have high hopes with the organization. They’re not supposed to get churned. So those were a few research papers. And then I did some research papers around the same kind of things where, I mean, lot of my research papers was around finance and the healthcare.
And one of the recent conferences which I attended, was MCP Dev Summit, where I also presented one of the works which I was doing is how to stream the context to the real time kind of a scenario. So if you want to stream the context, if you want to stream the real time data to the agents, how you can do that using MCP. Because your MCP works on a request response kind of a model.
But if you want to send the data in a real time fashion that you’re sending the data and as soon as you’re sending the data, is analyzing the anomaly, it is reporting it and then you can interactively trying to converse with that saying that, okay, how many anomalies in the past one hour? How many such events occurred? Okay, what’s the number one anomaly in the past one hour? So you can do that as well in kind of a real time fashion.
Federico Ramallo (17:16) interesting. I didn’t know you could do that. Yeah, the best way that I understand machine learning is that you have a lot of data, and you have to kind of make sense of it, right? mean, before LLM, was, it was, you know, done, I would say manually, but probably this is the wrong term, right? But because it was through mathematical models, right? But you will try to find patterns.
And then from those patterns, you would make sense, right? So you could have bank records and be able to predict with some level of accuracy when a user is going to leave the bank, right? Based on those transactions, right?
Harshit Kohli (17:53) Yep.
Yep, exactly. So it’s more about predicting and analyzing your data. So data is the core backbone of all these findings and all these learnings we are talking about. So the more healthier your data will be, the more better your findings will be. So in earlier back before AI and AI came into existence, as you rightly said, there were mathematical kind of equations and models like
mathematical kind of formulas and equations which were used to predict all these things. So you are having data and then you are using all these kind of tools, mathematical equations and formulas to predict these things accurately. And it was kind of, I’ll say, semi-manual, not fully automated, but with machine learning and AI, lot of these things have been encapsulated. They have been abstracted. Now you have access to these models. You use these models.
You write a small code or a program to use those models effectively and then whatever your use case is based on a use case, you can utilize those models to get the results from it.
Federico Ramallo (18:56) Right, right. Yeah, I think that AI opened to automate a lot of those processes and also the speed of processing that data,
Harshit Kohli (19:06) Right, right. So I’ll give you a simple example. So for example, let’s say you’re watching a live soccer match, right? Now earlier what used to happen is you’re watching a live soccer match and after the soccer match is done, you are being given a task to generate insights or to generate a summary out of that soccer match, right? Before AI what used to happen is you used to watch the soccer match, you used to note the important kind of
time lapses in that soccer method. Okay. At 15 minute, a goal happened at 17 minute and offside happened at this much minute this happened. And then you will kind of write all those main, main points. And at the end, you will try to summarize it. Right. But now with the help of AI, you can just use a model, which accepts a video and generates the insights. It’s as simple as that. Right. You just pass a video through the AI model and you ask your model, write a prompt to saying that, okay, in just this soccer video,
and generate the insight, identify the players from different teams wearing different dresses, emphasize on the major kind of highlights in each time lapse and also finalize who won and who won the player of the match. So the more smartly you will give the prompt, AI is going to analyze the entire video and it is going to give you a nice summary out of
Federico Ramallo (20:28) So basically you can just skip the whole machine learning process by using AI.
Harshit Kohli (20:35) Not exactly machine learning,
say the manual process which used to do earlier, like by watching the video and everything. mean, see machine learning was there. I mean, even before this AI thing started, right? It was always there, right? All those models, fine tuning models, those things were there. But now with AI, whatever things were being done manually or you were
writing let us say, lengthier logics or complex logics for handling those things those have become very smart and very kind of easy to handle.
Federico Ramallo (21:03) Yeah, that makes a lot of sense. The level of decisions that a model can make based on the data, they can, or it can find patterns that it becomes harder for us to detect, right? So that’s very interesting. So what is the biggest gap between what companies think of…
AI will do for them and what it actually does.
Harshit Kohli (21:28) So yeah, so the biggest gap I feel right now is that most of the organizations feel that with this AI coming into the world and AI things coming into the world, AI has got a wand kind of a thing, which will solve everything for the customer. But that’s not going to happen. As I said, data is the key backbone. If you have, let’s say stale data, your data is not perfectly ready.
AI doesn’t have the magic wand to correct the data and then generate the insights for you in a better and a seasoned way. Right. So it all depends upon your data and not everything is AI fit. I mean, I have seen a lot of people I talk to, okay, you are doing X, Y, Z, Y is not possible in AI. I’ll say it’s kind of a, it’s kind of a good thought process, but not always.
There are certain limitations what AI can do versus not. So these two are the biggest, biggest, I’ll say the biggest kind of misconceptions the communities have right now. Number one is you need to understand that AI cannot do everything. AI doesn’t have a magic wand. can solve everything for you. And the second thing is that not everything is AI fit. So you need to categorize on what fits with AI versus not. And
A third, a very universal kind of a set rule is that your data should be ready for AI. So AI cannot fix your data, your data has to be ready for the AI.
Federico Ramallo (22:58) Right. It’s formatting, accessing, and then processing of the data, right, that has to be ready for AI.
Harshit Kohli (23:06) Correct,
Now, I will relate this with an example. So let’s see, let’s say I’m given a task where I have to ingest about 100 images. Those are kind of plant or flower images. And then I have to categorize which flowers are those or which plants are those using some AI models. imagine like out of those 100 images, 70 images are super blurred.
You cannot even identify what kind of a flower or what kind of a color it is, right? Only 30 images are crystal clear where you can find out what the flower is, what the image is, what the color is. Now, AI models, we have very smart and very good models. The models will try their level best to go and find out what kind of a color it is, what kind of a match it is. But if the image is not clear, it can give you the best kind of an analysis. It’s not going to give you an accurate analysis.
So for the analysis to be accurate, your images should be clear so that they can be analyzed and the best results can be drawn out of it.
Federico Ramallo (24:10) Right, right, that makes sense. And in the analogy you were talking about of the image, have you had issues or found issues where you see the models kind of like…
I think this is called token optimization, right? So providing a a better approach for the model to find the results, right? Rather than just letting it fill it out, right? And just in that exploration, you use a lot of token or resources or, know, are different ways to measure it, right?
Harshit Kohli (24:49) Yeah, I mean, I have faced this scenario where if there are blurred images or the images are not clear, a model will start hallucinating. They will start hallucinating giving you kind of saying that, okay, as per the analysis, this could be team A, whereas if you find out it could not be team A, mean, the jersey color might be different, the team name might be different, the player name might be different, but they’re just hallucinating because
They are just trying to predict based on the analysis they have done.
Federico Ramallo (25:22) Right, right. It’s the same example you were talking about, you know, the football match analysis. I mean, if the information is not there, it’s going to try to do the best estimation. And when that becomes too unpredictable or too unreliable, then we call it hallucination, right?
Harshit Kohli (25:33) Right.
Correct. Exactly.
Federico Ramallo (25:42) Yeah. And I think that the best way that I’ve been able to kind of understand this is that the models are very optimistic, right? Whatever you say, they’re always going to try to please you, right? And they’re always going to try to figure it out, you know, give you an answer.
Harshit Kohli (25:57) It’s all about
training those models. So the more the data you will provide, they’re going to get trained. Right. So there will be a situation when they will start pulling the information from the earlier trained data and they will try giving you that information. Right. So a lot of times you’ll have to instruct the model that, okay, if you’re giving a prompt, tell them that, okay, I need zero has a hallucination or zero assumption. predict or analyze, or just give me the inference on what you see, not what you,
I mean, not what you expect or not what you think is right. So I think prompt engineering is also a key element here because the more detailed prompt or the nicer prompt or the more clear prompt you will give to the model, the model is going to give you the results out of it in a better way. So your prompt has to be strict. It has to be crystal clear so that the model is not hallucinating or not producing the results which you won’t expect.
Federico Ramallo (26:28) Right.
Right, right. I’ve seen the development of skills, which is basically a text. The way that I understand it, and correct me if wrong, but it’s basically prompt injection, right? You’re adding, ⁓ right, right. So that way, when you type something as a user, you don’t have to explain everything because the skill kind of, you know, have the steps and the rules and whatnot, right? But I’m wondering…
Harshit Kohli (27:04) Yep, front engineering.
Federico Ramallo (27:19) What has been, what are your thoughts between, you know, the limits of where skills can go versus model optimization, right? Can we, you know, where have you seen that limit happens? Where it makes sense to just find other models or optimize, fine tune a model, right?
Harshit Kohli (27:37) Yeah, so I think the way you choose a model, now for example, there could be there could be multiple models to do a specific task, right? So for example, if you are processing video files, there are multiple models which could do the same amount of thing. But initial analysis is important where you need to understand, okay, how much big is my video?
whether that much amount of video can be analyzed by this specific model or not because every model has their own limitations. Some models accept an hour of video. Some models say that, okay, we cannot accept an hour video. We can just accept a maximum 30 minutes of video and we can produce the analysis, right? The initial analysis is important. And then always we should not rely on a single model. We should have multiple models which can.
give you results and then you will have something to compare with you can say okay model A is giving me this much accuracy versus model B which is giving me this much accuracy and then the closest result you find with whichever model you should try to fine-tune that model as much as you can so let’s say model A gave you 95 % accuracy model B gave you 97 % accuracy now it will be worth putting efforts into model B’s accuracy and try to reach that to 99 or close to 100 % right
rather than sticking on model A and trying to improve that. So, I think models, yeah, I think model selection and once the model selection is done, then model optimization or model tuning makes a lot of sense.
Federico Ramallo (28:59) I see.
Right, right. That’s very interesting. The other thing I’ve seen happening is models evaluating models, right? The output of one model is evaluated by another model, right? And what I’m kind of inferring, it’s this idea of models. I mean, for most cases, it doesn’t make sense to fine tune models because it’s cheaper or easier to just…
go and find other models. The catalog is so big, right?
Harshit Kohli (29:36) Yeah, I mean, it depends. I mean, it depends upon your use case because you might have a use case where you will have a limited set of models. So as I said, if you have very lengthy videos like two hours or three hours video and you want to generate an analysis from it, not all the models will allow that, right? I there is a very limited set of models which will accept two hours or three hours length of video and give you the exact summary or analysis out of it. So I think ⁓
Federico Ramallo (30:00) Right.
Harshit Kohli (30:01) As AI is progressing, we are getting more and more of models in the market. So soon we’ll have a lot of options. But I think deciding amongst if you have a use case, select the top two or top three models and then try to play around with it, test it. And the number one model which you think is giving you the closest kind of result, try to work with that and try to fine tune it and optimize it the better you can.
Federico Ramallo (30:26) I’ve been trying different models and I find overwhelming the variety of models that are available. And there’s a point where I say, okay, I should just pick one and try it out. But then even if when I do that, it’s hard for me to judge between one model and the other because I mean, I can, because it’s a non-deterministic process, so it’s hard for me to figure it out, right?
Harshit Kohli (30:51) Yeah, I mean, again, when you work on different models, and this is what has happened with me. Sometimes I’ve also seen that the structuring of the output or the structuring of the response is different with different models. So let’s say some tables are giving you a response in tabular format. Some are giving in plain text format, which it’s hard to read, right? So I think it all depends upon what kind of things you want to get it done from the model. And then you have to, again, as we discussed, you have to be smart and
designing your prompts so that the model can generate the best results for you.
Federico Ramallo (31:24) Right. Right. Interesting. What recommendations do you give to the audience on how to use AI tools on their workflows today?
Harshit Kohli (31:35) So if I talk about using AI in your day-to-day work, I’ll say whichever workplace you’re working on. So let’s say you’re working in tech. So most of the tech companies, are using, I mean, they have either started exploring AI or they have some kind of meeting where they have some kind of a roadmap to use AI in future, right? So I think not all of us can.
have a kind of AI center of excellence team where we think that, OK, since AI is booming, I want to work in AI itself. I don’t want to work on anything else. Technically, that’s not possible because there are other verticals as we discussed where you can still use AI in bits and pieces, but everything cannot be completely AI. So my recommendation to the audience will be that in whichever vertical or whichever sector you’re working with, try to explore the use cases on how AI can fit into your existing
sector or existing kind of a existing team. So for example, if somebody is working in human resources, if somebody is working in marketing, if somebody is working in sales, right. So right now there is a heavy influx of tools or heavy influx of applications which are getting launched in these sectors as well, which are AI ready, right. So all you have to do is you already have an application which is backed by AI, you all you only need to just provide them certain details and it will prepare a nice somebody for you.
So let’s say if I’m a sales engineer or if I’m a sales manager and I have to pick something to the end customer, I just provide some kind of initial kind of data points with the model or the application and that app will generate my talking summary for the customer. I just need to prep it up and be in front of the customer, right? So it’s all about how you have to utilize it smartly. from what I can recommend here is that try to use
the AI tools in your existing sector smartly and that’s the long way to go.
Federico Ramallo (33:27) Right. Yeah, that’s very interesting. I’ve seen so many SaaS companies trying to back fit AI into their products. They think that putting a chat prompt, that’s it. I think there’s more to it, right? And having a full integration, it’s a much deeper complex problem, right? And then in my personal experience, the issue that I’ve been having is,
integrations, right? I mean, you can interact with a lot of models, you can build some agents, but then, you know, the level of integration you can have, it depends on having an MCP or having, you know, a level of integration with your current applications, right?
So cloud adoption has been going for over a decade. What do you think are enterprises still getting it wrong?
Harshit Kohli (34:13) I will say sooner or later organizations have started realizing that cloud adoption is a requirement. I still feel that there are organizations who are not on cloud. They are still on on-premise. I do know a lot of organizations who are not still ready for cloud. Maybe they have a compliance issue or they have certain kind of issues.
A basic mechanism which I’ve seen in customers who are using cloud is lift and shift, where they just take out their whole system and just shift it to the cloud, right? And once it is on cloud, whether it is right or wrong, it doesn’t matter, then they start optimizing it, right? Plus the other thing what I’ve seen is most of the teams within the different verticals, have knowledge gaps or skill gaps within cloud.
So I think these are the core areas where the teams or the customers or the end community should focus on. That if you are going to adopt cloud, if you’re going to work on cloud, this knowledge gap or skills gap should be minimized. You can start prepping up your team, helping your team to understand the basics of cloud, have them prepare for certifications, do multiple kinds of trainings courses for them so that they can understand the basic mechanics about cloud. Otherwise the problem will be that once the lift and shift is done.
and you start operating your production on cloud, then you’ll start onboarding more and more workloads on cloud, right? Your team will eventually not get time to upskill themselves or remove the skill gap. So I think investing on filling up the skill gap is super critical, which I, in my opinion, should be emphasized.
Federico Ramallo (35:45) Right. That’s very interesting. mean, we used to think of the cloud as this big, unique thing. Amazon has been one of the first innovators in this, right? But then we realized, well, it’s another computer, right? So having that mind shift allows you to understand, you’re still using resources, optimizing those resources still becomes important, right?
regardless of the billing, I mean, I’m coming from technical side thinking resource optimization, it’s the goal, right? Regardless of whether it’s your hardware or the cloud hardware, right? And I think that what Amazon did with the virtualization,
you know, of the cloud allow you to have access to different resources has been brilliant, right? The I, I, I come from a school where a school of thought where you had to set up your own rack, your own server, right? And I would, you know, just being, being bolting, you know, the servers and setting it up and making the connections and everything, right? And now, and AWS was able to bring this idea of
Harshit Kohli (36:36) Right. Yep.
Federico Ramallo (37:00) one command line and boom, you have a server running, right? ⁓ That has been an industry changing. Now with that, then people start thinking of this cloud as this monolithic with infinite resources, which it is, but then you have the building site that says, well, you have to optimize it, right?
Harshit Kohli (37:03) right.
Yep, definitely.
Federico Ramallo (37:23) with great powers come great responsibility. Yeah. So you also mentor engineers and judge hackathons, right? And also you review academic papers, right? What do you see in the people who are going to make it?
Harshit Kohli (37:25) responsibility.
While I was judging the hackathons and I was reviewing the research, I’ve seen one exciting pattern. Nowadays, there are so many good ideas and so exciting ideas coming out of these hackathons and these that if you go back five years from now, six years from now, you can just sense it. mean, okay, while I was five years, six years back when AI was not there. I mean, these kinds of things were just
I mean, these kind of things were just assumptions. They were not practically true. But now when you think about these kind of ideas, are, I mean, people are proposing so nice ideas, so kind of corner cases ideas.
And when you’re seeing those in front of you, you’re reviewing them and you’re trying to judge those ideas as part of hackathons. You kind of enlightenment in your case that, okay, maybe these things are possible in the next gen, right? So this is the kind of thing I’ve been appreciating a lot when I was doing the hackathon judging and I was reviewing those papers. recently I think I was doing a hackathon judging too
months back around. there were many more ideas around eco-friendly atmosphere and bio-agriculture and how to maintain those kinds of things and how to utilize AI to predict the weather so that the farmers life can be preserved. I mean, they can be made sure that, they are going to grow such crops which are healthy for a particular climate or weather, right? So things have gone to that far. So
Earlier, mean, you would imagine if it’s a disturbed weather, if it’s a rainy weather or kind of a winter weather, and if you’re not preparing for that, all your crops and everything will be kind of destroyed, right? But now with these kind of models and these kind of solutions coming out of the box, we still have a ray of hope where farmers and other sectors and other industries can utilize them to predict things in advance.
Federico Ramallo (39:37) Wow, that’s amazing. I mean, it reminds me to the movie A Brilliant Mind, right? Where, you know, I don’t remember the name of this actor, but, you know, his character was, you know, was able to connect the dots that for everybody else was, you know, was brilliant, right? And I think that that’s, you know, having AI models that can make those connections, I think that’s amazing, right? And the people behind it that can…
Harshit Kohli (39:38) Yeah
Yeah.
Yep.
Okay. Yep.
Federico Ramallo (40:07) can set it up in such a way, right?
Harshit Kohli (40:09) Right.
Federico Ramallo (40:11) Yeah, I’ve seen that the bar to be able to implement an idea has lowered with the AI adoption. I’m wondering how much of that allows us to prove good ideas or bad ideas before we even implement them.
Harshit Kohli (40:27) Yeah,
I mean, there is a mix of both in the market right now. I have seen a lot of good ideas where people are looking for honest opinions, honest feedbacks, and they are looking to improve their ideas, enhance their ideas. And on the parallel side, we do have kind of fabricated ideas where you think that those ideas don’t have that much amount of depth, but still people are trying to portray that it’s a big idea. So both kind of things are existing.
But from what I feel is an idea is an idea, whether it’s a fabricated idea or it’s a good idea, unless you prove it out. So I think the next step, which we recommend all the people participating in hackathons or while I’m reviewing any kind of research papers that, okay, I would like to see these things in action. Okay. Let’s try to gather a sample data set and let’s see how you can try out this idea.
Federico Ramallo (41:21) Right, but validation becomes such an important part now, right? Yeah, and having the lower bar to build a prototype or to just test it out becomes much more important because now, you know, weeks of development. You just build a prototype, I’ve coded it or whatever, and then you see if it works, right? Yeah. Yeah, I mean, I think that becomes such an important…
Harshit Kohli (41:25) Exactly. Yep.
Yep. Yep.
Federico Ramallo (41:47) It’s delicate when you’re on a hackathon and you have somebody excited about implementing their idea. I think it’s important to allow them to explore it, right? Because that’s, I mean, maybe that idea could have no legs, right? But that person is going to have another idea.
Harshit Kohli (42:02) Exactly.
Exactly, because at times I’ve seen people when they’re working on a specific idea, while they are implementing it or they are prototyping it, maybe somewhere in the middle, they realize that, maybe what we thought is like just 10%, but what this will come out is massive. Right?
Federico Ramallo (42:21) Right.
And being able to detach from, know, the idea might not be successful, but that doesn’t say that, you know, you as a person, as a founder, as a hackathon participant, you’re a failure, right? It’s two different things, right? Being able to keep that motivation becomes so important, right?
Harshit Kohli (42:32) exactly.
Exactly.
Federico Ramallo (42:41) Yeah. So you have built a career that spans engineering, architecture, research and customer advisory, right? How do you think about what to learn next?
Harshit Kohli (42:54) So right now I’m focused on getting my research usually completed. And then I like to interact with customers a lot. And that has been my kind of a go-to thing. So I like to excel towards that itself. And I love to do sort of conference presentations, do a lot of presentations. So I’m heavily inclined towards kind of a customer facing roles, customer advisory roles.
So yeah, so I think while I’m working on my research, I’m learning a lot of AI related concepts, AI related domain knowledge. I will try to implement this in my customer facing roles itself. So that is my kind of a go-to kind of a framework for the next five to 10 years.
Federico Ramallo (43:37) Amazing. So for someone in India right now looking at a career in cloud or in AI or were thinking to move to the US, what advice would you give them?
Harshit Kohli (43:51) So I think the only advice I will have is they might be working in different vertical, different sectors, but they should try to go deep inside what they are working on, not just learn at an abstract level, at a level 200 level. They should dive deep. They should understand the core fundamental core concepts. And then with this AI and cloud and everything booming upon, they should try to
learn skill sets which can help them in the longer run. So for example, if somebody is working as I was mentioning, somebody is let’s say working in HR vertical or marketing vertical, right? So try to learn the skills which can help you build tools specific to your marketing or HR sector utilizing AI. every, mean, not every AI will require coding or anything, right? Some kind of AI will be
simple kind of a drag and drop kind of a tools where you just utilize those and then you can just build a sample app out of it, right? So try to learn those concepts, understand what AI is doing and how AI can help in your specific vertical, whether it is HR or marketing or sales, it doesn’t matter. diving deep and learning the fundamentals is super critical because if your fundamental is not good, it’s not going to help you in the longer run. And then once you…
or stick to these two core concepts, I think you will have enough opportunities, whether it is India, whether it is US or India continent, you will get immense opportunities out of it.
Federico Ramallo (45:22) Amazing, Harshid, we’re running out of time. I truly appreciate you being here today. Any final remarks before we wrap it up?
Harshit Kohli (45:32) No, I think that’s the, this was great. I really enjoyed the conversation with you and I really appreciate the audience. So yeah, hoping to have much more sessions like this. Thank you, Federico.
Federico Ramallo (45:43) Thank you, Harshit, for being here today.
Harshit Kohli (45:44) Thank you. Thank you, Mice. Thank you.