Sumit Kapoor

So I'm leading GenAI products at GSPANN, and other initiatives as well. I came over as part of an acquisition. I was founder and CEO at a couple of companies that I founded. Zorang was one of them, which got acquired by GSPANN about two years back. We had been building AI solutions since 2016, so we were a little bit ahead of the curve.

The initial problem that we were trying to solve was organizing unstructured data within organizations. 90 to 95 percent of the data inside an organization is unstructured. One of the problems that we saw AI could solve was, if you could train the models on the specific business vertical for that industry, for any customer, we would be able to take all their unstructured data and extract actionable intelligence out of it. For example, one of the first verticals that we worked on was real estate. That was primarily training models on contracts, because real estate deals with a lot of contracts.

Exactly.

Exactly. And then we could extract things like when the lease is expiring, which customers are more likely to turn over. For real estate investment groups, that was a great solution for trying to discover and take action before things come to a head. That was one of the first products we built. We actually got awarded for that at the AI DevWorld conference, as one of the top 10 ideas there. Still early days. As you mentioned, we were training our own models. Then these public LLMs came about, and we saw immediately that that problem could be solved very easily. You can take one of those publicly available models, fine-tune it for that particular vertical, and then extract information from there. So we pivoted at that point, and we built another product specifically for retail and product companies, primarily focusing on generating product content. Because that was one of the challenges when these companies were putting their products on the digital shelf: a lot of times they wouldn't have enough information. And one of the challenges with not having enough information is that these LLMs train on that data, the product information that you have on your pages. They've already consumed all the human-created data by now, so they need to consume more data. They have a voracious appetite for data. So we were creating product content, which also drives, by the way, engagement through answer engines like ChatGPT and Perplexity.

That is exactly where we started out: creating written content for the product. Product descriptions, product detail pages, feature bullets, etc.

You can say that. But one of the key things that we did in the platform is that it is based on the product company's brand voice and compliance and regulatory requirements. AI, as you rightly said, can generate a lot of slop if it's not on guardrails. So our core focus from day one was to provide those guardrails, so that the output adheres to their brand voice and any compliance or regulatory requirements, like FDA requirements, or NHTSA, because we have automotive customers. We have food and beverage companies as customers.

That is also there. But what we also found was that one of the weakest links was the prompt, because the output of these engines is only as good as the prompt that you feed in. So the ContentHubGPT product that I'm talking about has a prompt engine, a prompt layer built in, that allows us to provide the consistency and also the guardrails. It combines all the data, like your compliance requirements as well as brand voice. Plus there is a RAG database, which makes sure that the output is actually taking valid information.

Exactly. All of us have a different way of prompting, and we want to get to a certain result. Most of the time we know what result we want, but...

Right. And then of course it also helps, because right now another problem that is happening is token maxing. Everyone is burning through the tokens. And part of the reason is that the way we write prompts is very inefficient. And these models actually have their internal prompting language also. So we leverage all of that capability in our platform to make it really efficient and make it scalable. One of the other challenges is, when you manually prompt these engines, you have to prompt one product at a time. It's not very scalable. So we have made it a very scalable platform.

It's a pretty small team, actually, because it's a product development side. Most of my team consists of developers. There are some product managers as well as product marketing managers. And then I work with, of course, the external GSPANN sales team. One of the things that has been revolutionized with these AI models is, of course, coding, so we can operate with a very lean team. Plus, on the product management side, I act both as a product manager and as a VP of engineering. My background is, of course, programming, so I come from that experience. But then I've been leading products at IBM, and even at Zorang we built multiple products, and some of the products were led by me. So I have a lot of rich product experience. And now I could go back to my roots with the help of AI, because I had been taking a long break from hands-on programming. I'm now actually hands-on programming and delivering code.

Exactly. And now actually the value is understanding the business, the value that you're providing. Coding has become easy, really easy. So what that has led to is you can create a lot of code. The problem is, creating is really easy; when to stop is the difficult part. And that only comes with experience, and...

That only comes with experience. So I think for folks like us who have rich experience, it was difficult. We were kind of dependent on the back-end development team to actually realize those solutions that we had in our minds. Now that problem is gone. You still need teams, of course; you will still need teams to develop. But now even creating product definitions or requirements definitions is so much easier, because you can just translate it using these models. And so I'm using Claude Cowork to create the Jira tickets as well, while at the same time churning out code on my own.

Right.

I think it's going to open up tremendous opportunities also. If you look at it, when things become easier, you do more of it. If you look at video editing as a profession, yeah, it looks scary. But now as a video producer you can produce way more videos. You still need those editing skills. Going back to my point, it's easy to create now; when to stop is the problem. And that only comes with experience, and that experience is with video editors, that experience is with the business leaders who have spent thirty years in the grind. They understand the customers' problems, they understand the solutions that can solve that problem. Translating it has become really easy.

So I think we need to bridge that gap. That is the challenge. That is why we are seeing stress among the fresh grads coming out of colleges. They're not getting jobs. And I think that is wrong, because they need to go through the same phase to get that experience. We need to actually help train the next generation. So I'm actually hiring. In fact, what it has also opened up is an opportunity: experience doesn't count anymore in certain ways. Taste is what matters, right? And taste comes with experience, of course. It comes with the years of toil that you have done. So you can actually hire young, fresh graduates, and I'm doing that in my team. In fact, the majority of my team is fresh graduates. You give them these tools to develop, and then train them through, share your experience with them to develop that taste. And that will take a little bit of time, so you have to stick with them. Show them the power of these tools to cross-check their... For example, just the other day I was talking to my product marketing manager. He'd come up with some ideas and was sharing them with me. And I said, well, it sounds good, but I would take it through a taste test through one of these models. So I just took that same conversation I had and threw it into Claude and ChatGPT, and it basically came back with some ideas on how you can actually refine that idea further. So inexperience doesn't count anymore, if you can actually cross-check it. So I'm training all my team to take those ideas. They still have original, great ideas, but take them through the taste test through these models, and then my experience can help guide them in that process.

So next six months is a long horizon, actually, compared to the past quarter. Yeah, exactly. Next week, actually. It's pretty crazy, the way things are moving. It's so fast in terms of possibilities and ability. I think this is going to be massive. Some of the technologies I just saw this tech week are mind-blowing. Some of the things that people can do. We are looking at maybe technologies like the holodeck from Star Trek becoming real now, where you can actually interact with the characters in a live video that's playing out in front of you.

There definitely will be, for the short term. I think we saw the same thing when the Internet bubble happened. I think the same kind of consequences we might see over here. But then that unlocked a huge number of possibilities that were not feasible before. The same thing is going to happen over here. I think there will be short-term consequences till we figure out how to adjust to the evolving technologies, and things that were difficult to do yesterday are really easy now, which typically leads to more investment rather than lack of investment. But I would say the froth that is there right now needs to settle down a little bit, and people need to see through it.

Not AI right now. I think it's still early days for that. Things are moving fast. I know it is going to impact a lot of jobs.

Yeah, definitely. I think one of the challenges with data centers is it sucks in a lot of energy from the local grids, and that will definitely spike energy prices for the local economy, and that can have a detrimental effect, definitely. But on the other hand, from a job side, yeah, there will be an impact on jobs and on the way we were doing things. I think people are already seeing that. I'm seeing comments online. Folks are talking about how most of their day goes into prompting these models rather than actually being creative. But again, I think we need to evolve. The skills need to evolve. We don't have to be mundane programmers anymore. I think we saw that happen when we moved from assembly language programming to C programming to the scripting languages and all of that. It became easier and easier to develop applications. This technology is again revolutionizing, and it's a leap forward in having the ability to build applications. But you still have to consider how to scale, how to make it scalable, how to make it reliable, how to actually solve the business problem. What business problem are you solving? You still have to do that. And that is where the skills need to evolve, to actually start focusing on that. I think this will give us time to do that. Ultimately, change is hard.

Exactly.

It's not too worrying for me, because our nuclear missiles still run on floppy drives. They're not even hooked up to the Internet, so I think from that point of view we are kind of safe.

Exactly.

Yeah.

One is in college, studying pre-med. That is another industry that is going to be impacted and get revolutionized. In fact, it will be very beneficial, because we need more and more doctors and health care providers in this country. If you are able to diagnose for less cost, more efficiently, and scale that across a lot of patients, then maybe health coverage or health insurance coverage becomes a moot point. You can go to an AI-based facility. You will still need doctors to review the work they're producing. We may go back to the days when a hundred-dollar doctor visit would be able to diagnose the problem, rather than going to a doctor and spending tens of thousands of dollars today.

Exactly. But there's so much data that is being generated. Who has the time to review it? These AI engines can help review that. So that's what I'm saying: a lot of these industries are going to be revolutionized through AI, I think for the benefit of humanity. I have a very positive outlook generally, and I think a lot of new possibilities will open up. So I think people just need to be patient as well, and be resilient, as you rightly said.

Yeah, absolutely.

Right now it is. That is a challenge. I think we need to weigh the cost of that, and I think we need to compensate the right folks for that, like the content creators that these models have been trained on. They should get compensation. And if certain policies are not favorable to content creators, there could be another company that comes out and creates a platform which can compensate them better, and they will create content for that platform. Because at the end of the day, these models have been trained on human-created content, and they can be treated as glorified pattern-matching engines, right? And that pattern has to be created. Human ingenuity is what created those patterns. So if we stop that, then they will token-max into creating garbage, if these models are just building off each other, off the content that is created by these models themselves.

But what if Siri on your phone, or Alexa, can tell you what is AI slop, or that you're spending too much time watching slop? Or even curate the content for your taste and give it to you, rather than you doom-scrolling the whole day.

Absolutely. Yeah, absolutely. It is already happening. We are already seeing companies out there that distinguish what is AI slop and what is not.

So I'm overall positive. I think this is going to open up a lot of opportunities for all kinds of industries and all kinds of trades.

Thank you.