Welcome to Tectastic, where we navigate the intersection of technology and business, uncovering innovations that redefine our world. Swathi Young. It is so lovely to have you on. It's fantastic. Thank you for being here. Thank you, having me, Kristen. So you have an interesting role in the world with being that fractional CTO consultant, and it's at a really turbulent time in most companies' technology space as we all try to adjust to AI and how it impacts our business.
And the thing I would like to understand from you is what do you seeing the reaction being, I've had a lot of questions asked of me, but I I'm curious. You've got a broader lens. Yeah. Lots of interest. Right? So, would categorize broadly into 3 sectors because I being a fractional CTO, I have my hand in a couple of different sectors. So 1 is I'm a CTO for a sustainability startup sustained chain, which was a established by United Nations.
It's a platform that has an AI engine that's been running for the last 2 years. That connects people across sustainability areas, grassroots, startups, innovators working in sustainability space. So that's the nonprofit sector, but very, very growing, sustainability sector. And the other clientele I have is mostly in public sector since I'm based out of Washington, DC. Uh-huh. I hear a lot from public to our agencies as well as some, health care and fintech commercial clients as well.
So I think the most risk averse, but curious are the public sector as it's bound to happen. Right? So so they're very curious about the use cases, understanding the use cases. And this is a conversation I've been having for 8 years.
With the public sector leaders because this was way back, like, in 2018 or so when I started leading teams to implement traditional in learning, especially started off with the research in lung cancer space, and then started speaking to a lot of CIO CTOs in the public sector space, especially of health and human services. So they were very curious about the use cases. And, obviously, nowadays, the conversation is not from traditional ML, but moved on to generative AI.
So 1 thing is about identifying use cases that obviously matched with their business goals because that's most important. And I always lead with business first. What is the problem that you're trying to solve? And I'm not going to wrinkle the pixie dust of AI until I understand what are some really hair on fire issues. And then I say, okay. This is something that a generative AI can't solve.
At the same time, there's a lot of speculation around whether it's responsible AI or also a lot of security concerns because the the public, discourse is that OpenAI is using all these large language models based out of publicly available information and some not so much. We know that from the New York Times case against OpenAI, but what they really don't know is the end price version of the APIs of OpenAI are as secure as your AWS. Right?
So they will as any SAS platform, they will provide you audited security documents. So there's a lot of learning and, presentations that I do to demystify that aspect of security. It's not that they are going to get the public available. The chat, GPT app, it's more of tapping into the APIs and the powerful use of those APIs. The other thing is obviously around Responsible AI, and we are in a great space today that did OpenAI APIs not the only option. We have Claude. We have Mistral.
We have so many other Amazon bedrock provides. So many foundational models. So take your big hugging face if you want to open source. So the way I look at is in this strategic role, I do a lot of education and advocacy to demystify how, you know, when the rubbermaid the road. It's not the tragedy window that the enterprises would see. It would be the horsepower of the large language model that they have to leverage and solve their business.
Problem. So in a nutshell, I would say there is apprehension. There is, concerns about security and bias and so, those issues that depends on the correct foundational models that you want to leverage, add drag on top of it. So That's the conversation I'm having. There's a lot of education and advocacy that I'm doing. It's interesting because you you touched on a couple things in there that are really important. The 1 that I'm really curious about, though, was the UN Sustainability AI piece.
I was CTO at trade lens, which was backed by UNS security initiative, which is their outreach to developing countries to help them build up their infrastructure for supply chains and all that kind of fun stuff. And I was surprised at how forward looking that group was. That tend to think of the public sector as being extremely slow rehab. Right. This 1 is more of a nonprofit, but extremely fast and accelerating space that we are in.
So 2 years back, our platform, which is think of it as a LinkedIn for sustainability because it truly, truly connecting different players in the sustainability ecosystem using AI. Although the differences is not open to the public, it is by invite only. So we have a lot of startups there. So our first version was a reinforcement learning engine.
That's a recommendation engine that connects say if you're an impact investor in regenerative agriculture, it would connect innovators to the impact investors in working in that specific sustainability area. But this version that we're currently working on and not yet inaugurated, which is under the covers. So if you have any questions, I think we can have a another conversation, but just at very high level, it's how to sort of bring the old machine learning models.
And integrate it with large language models. That's what we are working on in this next phase. That is an interesting process to try to figure my own company is doing a very similar thing. The benefit of our older way of doing machine learning, we wanna bring that forward and pair it with the of modern generative LOMs and that model. And it's not obvious how you end up doing that because they are fundamentally different. Yeah. They're very structure.
Yeah. And I think the use cases where it's very much driven by is that the right use case to bring them together. Sometimes traditional machine learning works for a particular use case. More often than not, we keep finding a traditional heuristic works, and there's no reason to throw a large language model. Right. At them at all. You mentioned not wanting to sprinkle the pixie dust around.
Mhmm. And that's the most common problem I run into if people will say, oh, we we tried this, but it only had 50% reliability on its output Mhmm. As soon as a human in the loop, right, checking that? Like, right. You're using it for the wrong thing. It's not magic. It's still technology, and it's only as good as the data that's been trained on. It's only going to be as effective as that. You're asking if you do things outside of that core use case.
And that's the moment that I'm worried about because it happens repeatedly with technology. Something comes in, it looks like it solves all of our problems. Everybody gets hyped up about it. We all jump in, and then the realization that just new technology. It's not magic hits, and there's that letdown. Mhmm. And the opportunity that was presented by it gets lost you know, in that inflation of the hype level.
Yeah. Yeah. A lot of the companies that traditionally would have been very slow to adapt to something like this. Jumped right in immediately. We're saying, yes, let's go. And even so far as doing layoffs in anticipation of being able to eliminate human roles when they hadn't yet figured out how to implement the technology. And so I think we're gonna see a rebound very, very who's on that, that approach.
Yeah. Absolutely. And I think I would like to say that where we are today, there is yet not a large language model to clean the data for you. Right? Even if you use a foundational large language model, take your pick from hugging face to cloth to OpenAI. You can take your pick, but it's only going to be as good as your data and your lineage of data is so important and the provenance of data is important and what, you know, the data elements.
So there was a very interest used case that I was talking to somebody from 1 of the public sector agencies here in Washington, DC. So they were trying to see that the COVID vaccines and their reach within the United States, and there were so many deserts. There were pockets of geographical areas in the US where there were not many vaccines, right, vaccines taken or given and so on. So when they were trying to collect the data and do a regression model, these are outlier in the data.
And as a data scientist, you have to make a decision what to do without liars. And there's a whole technical process around outliers, but even business wise, you need to make a decision whether it's mission for an agency or a nonprofit or a business organization. What do you do without liars?
And sometimes people often not consider it or taken into account, but in this specific use case, imagine you're not getting the context and the desserts of the back scenes are because there is no access to health care in general. Right? So there are pockets of this country where there is no to a health care in their particular geographical area. So you are losing the context just by eliminating the outliers of data. Right?
Information resides, and too often, we mess the information architecture by looking into the data picture. So I always emphasize on context of the data and your business processes and processes in general when you're especially.
I think this is a very good example that it's very easy to delete your data and and actually remove the outliers, but it has a huge impact If you're trying to understand what is the impact of people not being given vaccines in this country, then you have to understand these outliers and what's causing it. And similarly, we I'm sure you have seen a lot of data that's either corrupted. It's not existing.
And sometimes you don't even know the data lineage, which OpenAI itself as a lineage question, and now they're entering into partnerships like Nos Corp being the latest. But previously, there have been, like, New York Times actually put a case against them. So so data lineage in your organization is so important. And and that's the conundrum for a startup, like sustained chain. We are a startup.
So initially, you don't have a large amount of data, but we still ran our machine learning engine and continue to run on smaller amount of data. And for that, we are using the reinforcement learning engine I think everybody has to make a conscious decision. If you don't have data that you can generate and you're augmenting the data, what do you do about it? That's most startups.
But in large organizations, you have lots of data, but you have either outliers, corrupted data, missing data, and it's not enough to look at the spreadsheet and just or a CSV file and say, okay. This data doesn't matter. There's a rich context behind that. He reminded me of a particular weird 1 we ran into with a former company, and that was that the operations team had become reliant on a core set of metrics, and we realized that it wasn't telling them what they thought it was.
And so we corrected that, and they asked us to change it back because even though the new data might have been more useful, it might be more correct. They were measuring their performance and how they, they were running the business. And people's bonuses were tied to that number. And when we corrected it, all of a sudden, all their numbers came, the number was on time delivery. Of the core metric.
And it was like, well, it used to be a close to 100% because we effectively were lying to it ourselves. We were adjusting that last minute, like, oh, it's not gonna be there today. It's gonna be there tomorrow at 4 pm. Even though a month earlier we were gonna say it's gonna be there today, we would update the number over at the old data. And then present them with they're almost always perfect. And all of a sudden it dropped, like, 25% on time. And so people were freaking out.
We're like, can't make it better if we have bad data. If we have bad information, we have bad decisions and we're incentivizing the wrong things. We actually had to fight that all the way to the CEO because it was a, like, we're not getting better. We're actually getting worse. At least we're updating this number more frequently. That's actually how we figured it out. We started looking at how, the rate of change on the fields and went, that shouldn't be updating. That should be persistent.
Oh, no. So the outliers are where we looked for a process improvement opportunities and places to make interesting decisions. That's usually where you find out things aren't perfect. Exactly. That has been my experience as well. And the other thing, obviously, is a lot of, conversations around but you give it different names at the Galaiyai or Biasinai. And I had the good fortune and of a project I undertook. This was in 2019, and we published it during the pandemic.
The fall of 2020 was the ethical AI framework for the US government. So it was a collaborative effort between government public sector commercial and as well as lot of folks from universities. So it was a very enriching experience. I was leading the bias, part fit and, was instrumental in bringing it all together and putting the framework together. So a lot of learnings there, although it has been 4, 5 years since I worked on it's been continuous discussions on that topic.
And even technology wise, I look at it 2 ways, right, for responsible AI, there are things to do outside of technology, whether it's regulation, bringing the right people, whether it's your legal teams, HR teams into the conversation, And the third thing is what do you have technology wise? Even that is under research. If you talk about explainable AI, there are still techniques like line, which are still being evaluated and continuously used.
And I know right now, we are also doing more, red, what they call the red light testing as well.
So so that's another hot topic of discussion and, definitely I am a strong proponent of, also, what are some of the things you can do outside the technology solutions, whether the diversity in your engineering and data science teams matters and unlike traditional software, don't leave out your legal, especially if you're in a regular industry like finance And Health Care, you need to have your legal in the room for any decision making, using an, an AI, even if with human in the loop.
Sometimes if you're using whether you're using large language models or, deep learning techniques, Sometimes the explainability would not be so clear. So those are some things you have to be aware of and also HR because to your point about people are getting laid off. I don't think in this interim that people more people are getting laid off. I think it would be more of a hiring freeze, but we do know need people, right, to understand.
I say that, always that the data scientists in the technology teams do not understand your business process. Even when I was working with the lung cancer surgery department, we had to sit with the lung cancer surgeon, even 1 hour or 2 hours of his time to understand the process, right, because we won't even know how to understand data. They don't explain it to us, and my team of data scientists, same with the radiologist. We were working on lung cancer CT scans. We had no idea.
They had to explain to us, walk through their process completely. So there is a, rich domain expertise, and these, some of my friends are not in the world of technology. They were like, oh, Swati, you're like, the, person in the room. Why don't you go through? What should I on? What certifications? And I was like, no. You're underestimating the domain expertise you bring to the table.
As a domain expert, if you know the right question, questions to ask the data scientists, that's the best learning you can do because the data scientists could do algorithm but they won't understand the context, the data, the business processes. So I think it's like a 2 piece in a part. They have to work with the business users, domain experts. To make the algorithms relevant. It's, it's funny because it often comes back to the same lessons we've learned so many other times.
Whether it's the constant evolution of AIs we're dealing with right now, ML before I did a science before that, just software development in general. Yeah. The, the first principles of those are the same as they are now. Context matters. Don't expect anybody to make the right decisions first. And it's gonna be on the learning evolution we go. So plan for obsolescence, plan for mistakes to occur and be adaptable.
And too often, at least in my experience too often the pushback that you'll get is that some cost policy kind of baked into the business decisions. Mhmm. We've already paid for it, therefore, and it has value as it is. Please don't break it. It's like, but it's not doing what you think it is. It's not providing the value you think it is. We need to adapt. Yeah. I keep telling people whenever I talked to him. I was like, well, what do you do today? You build software now. What are you doing now?
Same thing. New technology. Same rules apply. Yes. You're gonna have a lot more data scientists in your world now. You're gonna have people that are a lot more capable on the data side. Understanding it, you need them, but you should have had them anyway. Like with Wayfair, we had a very large PI team that was there largely to help us find where we had inefficiencies and we could resolve it.
And then that became our ML teams, and that's sure now is the AI type teams to the same group, what are they doing? Saying that they were always doing where we inefficient, how can we make it better? Yeah. And I'll although I think a little caveat to that, I would add is that maybe an addendum is that with ML or even LLMs, it's just not 1+1 equal to 2 with traditional software. It's very rules based.
And, you know, you are expecting an output based on some rules and that you write in your code, I think the biggest thing is that you have to be very patient for your LLMs to work. Sometimes the observations that the learning could take could take anywhere between 3 to 6 months. It's not like you're implementing a software in 3 months and you're getting benefits like your automating your procurement cycle.
That's a software, but there is your predicting which of your suppliers will do on time delivery that will take some time for it to kick the machine in place by learning the history, making predictions. Right?
So I think that is 1 difference, and that's where business leaders who are in a hurry either to realize the return on investment or to realize even the value of the problem that is trying to solve could take time, especially I think where data is smaller, the time would be longer, even with larger amounts of data. I think it takes some time for the prediction machine to churn and and and act, you know, optimal levels as expected by businesses. A very good point.
The learning aspect of it and the assumption that if but we fed it all this data, that there's a feedback loop that has occur for the training to work. And that is, did my prediction meet, for example, like, the outcome expected? No. Okay. What happened? Reinforced continue on. Yeah. If you ask, like, you had a person there that had to learn to experience, these magic technologies, we keep calling AI are no different than that, and they're just faster and could, like, handle more data.
Or they have had to be in existence for many years. Right? So today when Facebook has a billion users, the machine is working really well. We don't know how it was, right, initial machine. They didn't they might not have had even machine learning. So the power of whether it's LinkedIn or Facebook that have machine learning engines at scale is that they have been in business for long. They have been tweaking the algorithms for a long, long time.
So if beware, any other business or startup has to emulate those, it's going to take some time to 1 build the data too for the training to kick in. We only have a couple minutes. I just realized that, and I want and I feel like I I could sit here and talk to you forever. Yeah. And neither 1 of us had the time for that. So I wanted to give you a couple of minutes just to give, like, final thoughts to the audience.
Yeah. The final thoughts are, I think, professor, a Galloway for those of you who follow Scott Galloway, said it correctly. It's not AI, which will rule the world, but people who know how to use I will rule the world. Right?
So I would advocate that whether you're in technology field or not, just bring yourself speed with what's happening in AI or at least just use chat deputy for everything, even though even if you're a lawyer or a surgeon, right, I think it is there is a lot of, power in using the technology, exploring the technology because it's it's going to be here to stay. The Pandora's box is already open. Handover's box has already opened a green. Absolute pleasure having you on. Thank you.
Thank you, Kristin. It was a pleasure. And that's a wrap for this episode of Tectastic. Wanna thank you personally for joining us, and we'll see you next time. Until then, keep exploring, and stay curious. Thank you for listening. If you are new here and enjoy content, please subscribe. It really helps us out. And if you are a regular listener, thanks so much for your continued support. Overwhelmed by tech debt, Discover Vala AI, the solution to tech challenges with the simplicity of a click.
No engineering background? No problem. Valla AI enables anyone to effortlessly tackle tech issues, freeing up your time from tech headache Make tech debt vanish with Vala AI, where your tech solutions are just to click away.