Building AI-Ready Data Platforms: Enterprise Strategies for Scaling AI | Agentic AI Podcast by lowtouch.ai

So today we have Vinod, a distinguished IT leader with almost 25 plus years experience, leading global teams and multi-million-dollar tech transformations at big giants like Fidelity, Mass Mutual, State Street, all the big ones in the Boston area it looks like. And most recently, we know it served as a fractional CTO for a stone peak. He holds master's degree from Harvard, Water and CTO certification, and the data science certification from MIT. Vinod brings deep expertise in AI, cloud data platforms, transformative projects in finance and environmental services. So let's start, I think it's a great resume. I'd love to hear your thoughts on some of the things that is happening in the AI world. So what is your, I'll start with the basic introduction, let you chat about what do you think the AI is, or what is the role of AI in enterprises in general you think. And what CTOs and CIOs should be thinking about in terms of how to adopt them. Absolutely. First of all, Rajit, thanks for having me. It's my pleasure to speak here. And thanks for that kind introduction. See, everyone is talking about AI, right? Every, every small forum, the large forum you go, everybody is talking about AI, co-pairers, chatbards, genAI, MMC, PCR, multi-modality, and so on, right? But the uncomfortable truth, today's topic is mostly around data platforms and stable business engines. AI doesn't just run on these heights, right? It runs on data, right? And not just the data, but it's reliable, trusted, and secure data, right? So if you are a CIO CTO or a business leader, you probably already been thinking like, what's my AI strategy or what's our AI strategy? But the question you should be putting back to yourself is my data platform, ready? That's where you draw a line between innovation at scale and pilot projects dying quietly, right? You see, a lot of companies spending way too much time on pilots, but they're not taking them to enterprise scale. To answer your question, when we say a ready platform, we're talking more than technology. It's about the foundation that you need to build. Like I said, data should be reliable, it should be accessible, you should have good guardrails, it should be secure. That's where AI can actually deliver value. The main thing is clean and trusted data, right? Because AI is not as only as good as data. You must have heard about it several times, right? If your data is messy, siloed, fragmented, incomplete, AI models will only amplify those problems. AI is more like a multiplier, right? If you have bad data, it's going to multiply, the multiply and give you the bad results. AI, if you have good quality data, it's going to multiply and give you positive benefits, better retro-investment and all that. So that's something that the leaders really have to keep it in mind. The second thing that I want to talk about is modern architecture, right? Scalable, intra-operable systems. Think like data fabrics, microsite fabrics, data fabric, lake house concepts, APIs for integrations, compute and storage separation. So you move fast without breaking if you are opt into this modern technology. You can still implement AI on the older ones, but I would say modern technology is going to be a little bit more seamless than working on the traditional platforms. The third important thing that I want to talk about is the governance and security. If you do not have governance, if you do not have the capabilities for a lineage and access controls and proper guard rails, it's not only innovative, but with those things, AI can be more innovative and you will end up building responsible AI, ethically AI with the right compliance. One thing I want to mention, I have seen in my previous organizations and even while speaking with a lot of leaders in the industry is that many people implement, they have quality data to a lot of extent. They have decent governance with security data address, encryption data address, encryption data, and transit access controls all this kind of stuff. But one thing I notice missing a lot is the cataloging aspect of it. A lot of the organizations do not have proper cataloging with their metadata and all of that. They don't have labeling, they don't have tagging, they don't have organization specific glossary and enrichment, the right mapping with industry, and the knowledge and things like that. To me, that is essential for AI. That's where you need to adopt into things like per views and callipras and unity cataloging from data. That's one thing that with that you will have the explainability, you will have a right semantic search, scalable adoption across the enterprise. Again, data is not good, AI is not going to, it can make it worse sometimes. Now this question is, can AI help in that? So let's say I don't have the right data, can AI help to fix that? It can see what I'm noticing is, it's not like implementing AI on top of a good quality data and well-governed data. You have a lot of inherent AI capabilities in all these modern technologies like informatikas or five trends when you build the pipelines, even sales force. These platforms are, they have enabled a lot of AI capabilities. Previously, data was traversing from source to target with many hops, delaying the SLAs in because of the lower compute capabilities and things like that. But now a lot of modern platforms have overcome those difficulties. Now people are talking about near real-time data or even predicting the future data, in many organizations. It's like people are talking the data as it is happening right now. So those capabilities are unlocked with some advancements in these technologies and tools that enable the pipelines and things like that. Awesome. And I think when it comes to the, I mean, from an implementation perspective, the good news is if you're using a Azure or AWS or Google, they do have private managed model hosting. So you can leverage that. Right. So the data will not leave your private account or a infrastructure or VPC. So you have the option to do that. Right. So now let's talk about the tools. I know there is, when it comes to data, fabric, data bricks, snowflake, come into mind. Right. So what's your take on that? How can they help? And how are they doing in terms of enabling AI? Yeah. See that's one of the hardest challenges the enterprises are facing. Right. Speaking with many people in webinars and seminars and various events, that's one thing that organizations are struggling with. Keeping the structured data and unstructured data side by side. So when we say structured data, like it's all well-organized, and columns and rows in a warehouse like an environment, right. Unstructured data is like a PDFs, transcripts, call transcripts, videos, or even IoT logs, right. IoT signals. This is all unstructured, transactional logs, network logs. All these are like little messy. They don't fit in a table. Right. But organizations tend to try to fit all of this data in a table like format. So that's one common mistake that I'm seeing. First, don't force everything into warehouse. You leverage the modern technology. Right. If you try to fit everything in a table like data warehouse structure, table structures, it kills the flexibility. Right. A lot of unstructured data remains untouched. Right. But AI requires structured data and unstructured data to give you the right outcomes. Dumping all of the data into data lake without governance is also not going to work well. This is where you go with these modern technologies where you can keep structured data and unstructured data side by side. Data lake concepts, Delta Lake concepts. Right. Data lake gives you warehouse like features. Delta lake gives you acid transactions, schema and forcements. So we have seen in all these modern technologies. So take that that's where that's one reason why I mentioned the modern architecture as the second thing after clean data as to what our introvisors should be wrapped into. Who by answering your question that way you can keep the scale of a lake but with the cleanliness and stability of a warehouse. Right. So that's where that's my two cents on your question basically. If you're talking about a fraud model, right. For example, if you're trying to direct or train a fraud model, combine structured data coming from your payment systems and transaction systems and things like that and unstructured data coming from customer complaints and emails and things like that. When it comes to healthcare, take the data structured data coming from claims tables and also unstructured data coming from your mara scans, doctors, nodes and things like that. So it is exciting. It's an exciting time. So now let's talk about some use cases, right. Can you touch on some of the use cases that you came across in finance and healthcare? Absolutely, right. So there are a number of them. I will touch on a couple of them because one of them is dear to to my heart because that's something that I have stood ground up at State Street, Trade Surveillance. So, if you take capital markets example, with real-time surveillance needs. If you take a NASDAQ, they have implemented something called SMART. SMART is like not a new thing for NASDAQ. They had it. It was all rule-based like several years ago. Now they implemented from their learning systems. It is the shift from surveillance purely rule-based to a learning system that surface about 5 to 10 percent of the alerts. That truly matters. Right? Previously they were producing a lot of false positives. People were sitting together to filter all the false positives. Now, with all the learning system implementation, they have cut down significantly and they are able to see the alerts that truly matter. So, faster intervention reduced false alarm, also on fatigue, fatigue all of that. The competitors are still, they definitely need on competitors. While competitors are processing two days old data for next two days or next five days to find out the real alerts, these systems are identifying the problems right when it happens. And customers, it doesn't get unnoticed by customers. They see it real-time. If not immediately, but eventually they see that. The medical, same thing with the medical system. I don't have a lot of expertise on medical field, but speaking with people, medical imaging is another thing. They have a lot of backlog. If you go to a doctor, you get the results one week later, two weeks later, because of the backlog of scans. Now, after implementing some medical companies are able to provide the results in like in hours after the after you go to the scanning. And then the speed at cancer treatments are happening. The delays have been cut down by 50%. So that's what I'm hearing from various people. Definitely is adding a lot of value. And we can talk about these things like in supply chains and in insurance claims processing all of that, right. But those two things come to my mind real quick. When it comes to privacy, right? When you talk about enterprise AI solutions. Privacy is a major concern and compliance, right? When you touch on that or what do you take on that is. Absolutely. Great question. Basically, privacy is absolutely essential. Thank God, organization, right? If you want to build a stable AI, build a customer, trust, earn trust from customers and regulators and even from your own employees, you need to implement privacy and you need to focus on privacy. Look, all models are hungry for data. If sensitive customer or employee employee data is leaked into wrong place, you're not just risking the compliance and fines. You are basically eroding the trust from everyone, right? Stability and consistency are very essential as well on top of this privacy. So I'll touch up on three things when it comes to privacy that quickly comes to mind, build privacy into the platform. Not it's not an afterthought. Use techniques like role-based access data, masking, tokenization and so on, right? Put some secure enclaves, so sensitive data never leaves the control zones. That's one thing. Build privacy into platform. Use governance. We talked about governance cataloging, pervuse, collibras, unity cataloging. That's another thing that can be used for the purpose of privacy implementation. Adopt privacy to preserve AI, right? So when I say that think about federal federated learning and think about synthetic data, differential privacy, you might have heard about it for some of the implementations are going through differential privacy data-masing. So these are implement these things while you train and innovate without exposing your rad data. So basically that's my two cents. At the end of the day, privacy is not a break on AI. A lot of people think privacy is a break, putting a break on AI. It's like putting a seatbelt, right? It allows you to go faster safely and gives customers and regulators good confidence that your AI engine can scale well. Awesome. And I think when it comes to the, I mean, from an implementation perspective, the good news is, if you're using a Azure or AWS or Google, they do have private managed model hosting. So you can leverage that, right? So the data will not leave your private account on the infrastructure or VPC. So I like the idea of, yeah, it's a seatbelt, right? You can now go faster, which is awesome. I know there has been, it comes to data, fabric, data bricks, snowflake come into mind. How can they help? Yeah, it's very near and dear to my heart, this particular question, great question, right? I get all of you excited to talk about these platforms. Fabric, I'll start with Microsoft Fabric. It gives you an integrated ecosystem, right? If you are a Microsoft shop, Microsoft technology implementations, whether it's data engineering, AI, real-time analytics, governance, everything is in one sass layer, right? So for AI, that means data scientists and business teams, they work off of the same fabric, endless handoffs, right? That's what enterprise is. So that's one good thing that I see. And then when it comes to data bricks, it has pioneered the lake house concept, right? It designed to have the structure and unstructured data side by side and data like use acid reliability and time travel and all that kind of stuff. So it means like you can build and train the models directly where the data lives, no massive copying, right? Snowflake is something that I have implemented in my previous organizations on both on AWS and Azure. It's stated as a warehouse. It's more when I implemented several years ago, it was for analytics and AI platforms and things like that. Now they have invested a lot on AI and they have a lot of AI capabilities with cortex AI and all of that. They have iceberg tables. They have native ML integrations. They have a great integration with marketplace with a lot of third party vendors and things like that. It has the data sharing capabilities with their partners. Snowflake is another wonderful platform. So it depends, right? But there is no, choose this versus the other. It depends on your organization's use cases, your ecosystem, your technologist stack. But when you talk to most leaders, the question they ask is, how can I implement AI or scale AI without breaking my business? Everybody has their businesses running on certain platforms. Right? How I can bring AI without disrupting my ongoing businesses. These platforms matter to answer that question because they quietly solve the problems that kills AI pilot from that kills AI projects with silo data and lack of governance and run away cars and things like that. To summarize in one sentence, I would say fabric is a winner when it comes to convergence with your B.I. data science governance and all of that. Data bricks is more it wins with flexibility, right? It's a winning point is it's flexibility with structured or structured data, streaming data and all of that. Snowflake is like it wins with the reach. It can easily seamlessly share the data with their partners and the clients. So that's my take on. I think the bottom line is really good news that you're saying, doesn't matter what your favorite platform is fabric, data bricks or snowflake, they are all AI enabled. Absolutely. Right? You should be able to leverage that, right? That's awesome. So now when it comes to, you know, we talked about different steps that enterprises can take from take it from take them from the current majority level to, so how do you prioritize? What do I do to go from what is the starting point? Ask a simple question. Again, this is our topic for today's conversation. Do we know where our critical data lives? Do we know the right use cases that we want to implement? Once you know the use cases, do you have the reliable data, trusted data to solve for those use cases? You don't need to build the entire platform in order to implement AI, right? You, what you call high value, low risk items and prioritize them, find out where your data is. And so that's how I would start. When a bit data maturity baseline, do audits or assessments across your organization and see where your value is going to be coming from and what kind of data that you have, that value that you want to deliver in the future. The other thing I want to talk about is prioritize the cases by business pain. Identify the pain points, right? Where do you have the right value coming for your business pain points, right? So that's one thing that I want to mention. Basically, paying to cost, revenue and risks. At the end of the day, that's what matters, like cost, revenue and risk. The other thing is identify your AI squared, right? Like, it doesn't need to be huge, it's got unless you have a one. of funding, identify what your squad together, put your center of excellence teams and have them align the technology to the business objectives and the pain points that have been identified and prioritized. And pick RUTRI pilots, pilot use cases, ROI, like not in months, just in weeks, right, with all the co-pilots, all the capabilities that are out there, put 4-6 weeks time frame, they try to deliver them with hard matrix, KPAS is another thing that the board is interested in, C-suite is interested in, and business, senior business partners are interested in, put the hard matrix, show them or reduce or rely on cycle times, lower false positives, like I said, value increase revenues, cost optimizations, risk communications, faster cash flows and things like that, right, show the hard evidence with hard matrix, the goal is to boil the ocean, right, it's a journey, right, you start slow, you just identify the high risk and low, sorry, low risk and high value use cases, experimented, and once you have gained the confidence, you try to product, like it, right, take it to the first environment, test it, reproduction, and then take it into production and keep monitoring it. So that's my two cents, is to create a flywheel, assess the maturity, win quickly, where it matters most, then reinvest in strengthening the foundation, that's what I would say, that's how it most from pilot to project. Thanks, we know, thanks for joining us, so I'll provide the we know, link, profile link, and the description, and I think the takeaways, your journey can start today, if you have any questions on any of the topics we discussed, please reach out to we know, or we in LinkedIn, we'll be happy to jump on a Zoom call and chat about it. It's an exciting time to lead a team, no matter what your role is, AI is real, and it's going to stay. If there are any questions, please reach out. Thanks again, we know, thank you very very much.