Think AI Podcast - AI Sprawl Is Real | Ep. 4 with Priya Udeshi

Episode Date: April 8, 2026

What happens when someone with 17 years of IT strategy experience gets a front-row seat to the AI transformation inside a global tech enterprise? Priya Udeshi is Chief of Staff to the CIO and Head of IT PMO, sitting at the intersection of AI strategy, governance, and execution. She shares how she balances enterprise-scale agent deployments with democratized AI for every employee, why old governance frameworks are breaking, and what CIOs need to do right now to stop falling behind.In this episode:00:00 Meet Priya Udeshi: From PMO leader to Chief of Staff05:25 AI curious vs. AI enthusiast vs. AI skeptic: Where do you fall?08:24 Enterprise search as the real solution to meeting overload11:59 Why PPM tools fail and what actually works18:55 AI is in the passenger seat, I’m driving22:31 AI doubles capability every 3.3 months. What that means by 203024:00 Agent sprawl is real and governance has to evolve31:16 Hallucinations: What’s working to contain them41:05 What CIOs should be doing RIGHT NOW44:33 Measuring AI-assisted engineering with real KPIs47:06 Priya’s 7-year-old built a children’s book with AI52:04 Trust your gut: Advice to her younger selfIf this conversation sparked something for you, hit Subscribe and drop a comment with your biggest AI challenge right now. New episodes every week.

How can I use AI to make my life easier, both personally and professionally, and how can I just improve my own productivity? From idea, from conceptual idea to product, AI was a partner. Welcome to the Think AI podcast. Each week, we talk about the most exciting AI research, tools, case studies, and more. I'm your host, Dave Goir, and I've been working behind the scene in data and AI for over 30 years, whether you are an AI expert, skeptic, or something. between this podcast is for you.

Today I have a guest I'm really excited about, Priya Odishi. She's the chief of staff to the CIO, head of the IT PMO at MongoDB, one of the most important data companies in the world, and she sits at the intersection of AI strategy, IT transformation, and enterprise execution. She is not just watching AI happen. She's in the room where it gets deployed. governed scale across the global organization. Priya, welcome to the show.

Thank you very much, Dave. I am very happy to be here and excited to talk about a topic that is very top of mind for pretty much everyone in the tech industry. Great. So let's start with your story. How did you get involved with AI? And how has the evolution help you what you're doing today at MogoDB? Yeah, no, for sure.

I'll give a little bit about just sort of my background. So I've been in the tech industry for about 17 years. I've kind of grew up through the channel of project program and portfolio management. So I think being at that, like you said, the intersection of strategy, execution, operational discipline, it's really cornerstone to, you know, driving PMO leadership, driving modern IT portfolio leadership. So that's been my primary space. I've been at Mongo now for three years. So I joined leading a technical program management function under our CIO's office.

again, driving technology-enabled AI-powered strategic delivery across the enterprise. About seven months ago, officially assumed the chief of staff role. So again, kind of getting that inner corner office of the CIA, the CIO vantage point for all things AI. You know, I would say that my AI journey, you know, this word AI has actually been in practice in our, you know, tech evolution for many, many years pre-LLM era. And so I would say, you know, prior to this sort of an official launch, you know, kind of pre-GBT era, AI was really a wrapper for, I would say, all things, predictive analytics, machine learning, robotics process automation, right?

So like RPA capabilities, any type of automation that you could put some predictive algorithmic, you know, a boundary around was considered AI at the time. And so, you know, a lot of the project initiatives that, you know, I had led or I had, you know, folks on my team leading were really in that space. ChatGBTBT, you know, was launched just before I joined Mongo, you know, just as I was entering Mongo, you know, just as I was entering Mongo, We did, I would say, what most tech companies do. We started on our chat bot era. It was really focused on, you know, how do we, you know, we didn't want to just roll out chat GBT at scale, you know, across the enterprise. You know, you have your security.

All of those things are very important. So we don't want to, you know, start pumping, you know, Mongo specific information or internal company data, of course, into those public LLM. So we went on the journey of building our RAG infrastructure, connecting a GBT-style LOM to our data sources. We ended up building something called MongoSys. GBT, which is actually still in existence today and it is used across the enterprise. So that was really, I would say, the learning phase of kind of the AI evolution. Fast forward to the agentic

phase, which is what I think we're in now. We kind of have two paths, at least at Mongo. Enterprise agents and enterprise scale level agent, agentic workflows definitely is cornerstone to our AI strategy, looking at how we can improve the lives of sellers, right, across our go-to-market teams, like how can we build GTP, what kind of go-to-market agents to health sellers, get everything they need to know about their accounts, their customers, write at their fingertips. On the people team side, how can we build agents to address employee cases, submitting simple task workflows? So there's definitely a good portion of our AI strategy that is, you know, focused on infusing AI into the daily workflows across all of our customers across the business, which again, working in IT, our customers

are every employee of the company. But then we also are looking at how can we actually democratize the use of AI for everyone in the company. I mean, you know, we talked about this when we first met, you know, Dave, like you and I should be able to create our own agents. You and I can create our own agents today, right? It doesn't, I'm not, I mean, I'm an engineer by trade, but I don't sit behind a computer and code, you know, in my role today. It's a very different, you know, kind of vantage point. But how can I use AI? We're all thinking that. We're all asking that question every single day. How can I use AI to make my life easier, both personally and professionally? And how could I, you know, just improve my own productivity? So, you know, we, we do have, and

And that's a real thing that I think a lot of companies are facing is, you know, what we don't want happening is similar to every. And I would say being in tech, we've, you know, we talked about this the other day too, like the notion of shadow IT, right? There's, you know, when an enterprise solution is maybe not as quick or as fast or coming into your fingertips as quickly as you want, what do you end up doing? You go and buy it yourself or go and build a solution yourself or purchase a solution yourself. And we will and are seeing that happen with AI as well, right? So if we don't, you know, keep up with the pace of innovation and speed that is necessary to get these agentic workflows and these AI infused into the day to day of every single employee, you'll see people creating their own personal accounts, you know, creating their own, you know, agentic capabilities outside of maybe your company enterprise, you know, secured environments. And that's exactly what is counter to what we want. So those two streams, I would say enterprise scale is definitely a stream.

And then also low code, no code, you know, democratized agent creation. is another avenue that we're looking at as well. No, that's really good to hear, and that's a pretty powerful story that you have. I want to follow up on one thing. I think the other day we talked about it. Before that, I want to ask you, I think I know the answer.

I'm looking at three different buckets today. One is AI curious, who are just seeing CHETGPT is everything. Then there are AI enthusiasts who are actually seeing the CHATGPT. They are seeing the other world like Claude than so many other open source model, KME 2.5 and so on. And then they are AI skeptic

where they think if AI touches anything, it's going to bring our world to the knees. Where do you put yourself in life? Yeah, I mean, of course, the majority of my, I would say, where I'm at is in the enthusiast bucket. But even being in the enthusiast bucket, there's a ladder, right? And there are so many runs on this ladder. And I by no means will say I'm at the top run.

I'm probably at the lower to mid, you know, run. Of course, I am, you know, exploring. I think just again, having that vantage point within Mongo and seeing how, you know, we are bringing to life a lot of these, you know, AI-infused capabilities. But then on the personal side, I had mentioned to you the other day, I, you know, got the itch to try to figure out, how can I, like, you know, start building my own stuff? And I built a lovable website and some AI-powered digital products. And I connected, you know, to a GitHub repo.

And I used Clod to help me, you know, do all of that, which, you know, I really was starting pretty fresh from being able to do that. So I think there's a ton of power and a ton of opportunity ahead. I do feel that, you know, even with the skeptics and maybe the folks that are still using chat GBT today, I believe that the evolution of where we're going to be, you know, with AI, you know, even a month from now is going to look vastly different. Like I have seen the timelines around the burst of this emerging technology is so much far more compressed than any other emerging tech that at least I've been privy to in, you know, the 15 plus years. You know, We saw the cloud boom and then the SaaS, you know, kind of boom.

And all of those still, you know, they took time. People are still on their cloud journey today, right, of, you know, migrating into, you know, cloud-based environment. So, but with AI, you know, what the landscape looked like, even with Claude, I would say, you know, what the landscape looked like just a month or a few months ago was vastly different that it is now. So, yeah, I think I would put myself in the enthusiast bucket,

but I still have a lot more to learn in that space. So I'm just excited to keep learning. No, you're humble and I see more humble people will say, yeah, I'm learning. And that's a good thing. If learning stops, because everything in your life, you have to learn, you know, even if you're a musician, you are an artist, you are a player, you're always picking something up from someone else. And that's a great thinking and the value that you have. So I appreciate that. I will also pick up on one more thing.

I think the other day we talked about, and I could be wrong, Glein's agent framework that you mentioned. What is it and how do you see it fit in what you're going to do? Yeah, so Mongo has adopted Glean as our enterprise search platform. And that's actually the other. You know, we talked about enterprise scale, agentic solutions. We talked about democratized, low code, no code, agentic solutions. Enterprise search was the other. I'll just give a little bit of a story before I kind of talk about the Glean agent platform.

But I, having grown up in this PMO space in many of my parts of my career, part of my remit was to implement a project portfolio plan of record, right? Like, let's, let's implement a PPM tool, a PPM system. We'll have one central, quote unquote, database for project and program tracking assets. Like, I always think projects are data objects and just like any data object, they have attributes associated with them, risks, issues, milestones, stakeholders, you know, resources, all the things that can inform the quote unquote health of a project or program. So in many, you know, experiences of my tenure, I have had this. remit of implement a system of record, implement a plan of record where we can pump all of our

project data and run analytics and run, you know, data-driven insights, obtain data-driven insights off of them. And I'm, you know, those have been successful. I've, you know, implemented multiple PPM tools. They've been successful. But even as a PMO leader, even with leading a team of, whether they're project program, you know, program managers, TPMs, as much as you can really drive to say that, you know, this solution is to help you do your job better. solution is to help you, you know, make life easier. There's always this administrative component, right? In the day to day of running a technical program, in full transparency, the TPMs and the project teams are not living in a PPM tool. They're living in docs and sheets and unstructured

data sources, right? Like, that is how they're making decisions. That is how they're documenting, driving decisions forward. They're not going and updating a, you know, PPM tool. So, like, what I found in any of those implementations is it becomes a little bit of an afterthought, even when do all the things to try to not make that the case, to try that be, have that be the forefront of driving program delivery, it still tends to be an afterthought. So what I see with Glean or with enterprise search capabilities is it solves that problem. Meet people where they are, right? People, and you and I both have different ways that we're using tools to help us do our job. And yes, we should have enterprise standard solutions and we don't want to talk the other day about

things like sprawl. And I'm also, I get the gift of leading our SaaS sprawl initiative. also. And so, yes, that is a real thing. You don't want to have this kind of hodgepodge of a bunch of different tools and different things that people are using. But what Enterprise Search does is it doesn't try to make this one-size-fits-all, whether it's, you know, it gave the use case about project data. There's so many other use cases for that. I'm able to just, hey, hey, give me the latest on Project 123. And it's searching the network for all of the different, you know, data sources that contribute to that. And it's translating it, using AI power and summarization to translate that into a meaningful output that I can then, again, whether I'm writing a, you know, QBR deck or giving a

strategic update to our senior, you know, leadership team, that's the power of enterprise search. So that's where we started with Glean. We have since kind of moved away from mandating the pumping of project information, at least from my, you know, my team standpoint, into a singular tool. The next stage of what we really want to do with Glein, and there is a dedicated team to focusing on this, is how can we release, you know, an agent moderation framework. across the enterprise so that we can build, again, build our own agents directly on the Glean platform. So that is sort of our next phase of the evolution. Oh, that is beautiful. And it just reminded me of one of my own stories. I worked with a large healthcare organization. And, you know,

they had several operating companies across the globe and we were managing different programs. And every year, we used to evaluate one of the PPM tools. So we had to do PMP and all that, but that's fine. But Then we are looking at PPM tools and guess what, none of the tool were really solving the purpose like you mentioned. You don't live in the tool. And all the PMS hated it. They didn't want it to go into the tool because they still have to see if you're entering the data there, somehow that should made into the people who are looking at it. But they don't look at it either. So then you are spending time in meetings and then saying the same thing that you entered over there.

And then they had a legit reason that, okay, I don't want to update it because I'm already providing. it in so many other ways. And I don't blame them. We started doing it in our organization from that point onwards. So in our consulting business, we are also using consultants and employees in India and other places. And those, you know, they have different cultures, as you already know, different time zones and things. So we started using different methods like, you know, maybe WhatsApp, teams chat, and started collecting it and started putting back into the system in the BPM tool itself. But it's more on the... No, I just imagine if you had agents to do all that for you.

Yeah. And those are happening as we speak, right? So they created their own unified box. I have my own unified box. And those are guiding us what to do. Not that we don't want to connect human to human, but that connection should be about solving problems rather than, you know, providing updates. And that's one of the thing I personally hated most,

that you are just sitting in a meeting where 10 people are providing. update, which you could probably read off and start taking actions. You are preaching to the choir. And I think about this a lot too. I mean, again, being in the PMO space, that is, this is the evolution from outputs to outcome-driven leadership, to value-driven leadership, to, you know, what is that focus on business outcomes and have that really underpin everything that we do. You know, I've always said the outputs, again, coming from the vantage point of program

delivery, the outputs are important because they inform the outcomes. You can't drive. Of course, an outcome is like, you know, I don't want to say it's a, you know, you have a strategic priority and intended impact on the business that you want to have. There's a set of work that has to be done. Bringing that down to earth and decomposing that strategic, you know, pie in the sky, you know, intended value outcome. You have to do a body of things. Yes, there's tasks and activities and things that have to be done to realize that outcome. But the value is in driving towards that outcome. If you can automate and build agents to focus on the outputs, to your point, then the shift in the conversation becomes very. different. You're not talking about administrative tasks and activities and tracking progress against those because in an ideal world, you have agents doing that part for you. You're really the shift in the, whether it's a meeting, a report out, an update from one of these reporting tools, you're able to look at a data-driven report that gives you decision velocity, right? That gives you, what do I need to do next to actually drive progress towards this outcome? And that's where I think AI is going to really change how many roles, as we already know, how many rules operate today.

And I think with that intended shift into focusing on the outcomes and accelerating decision velocity across the business. So, Priya, thanks for that great insight. One of the things we also like to hear from our guest is that how what you do in your job and your passion helps and solve the real world problem. Some of the things we already talked about, we personally work. in manufacturing and healthcare space. So we are focusing a lot more there.

What's your experience looks like? What customers at MongoDB ask for in terms of solving their own original problems and how you and AI is helping them? Yeah, no, thanks for the question. So yeah, as I shared earlier, like my passion, and I think why you and I had such a great call even earlier this week, too, you know, we're problem solvers. And so I think the crux of like the basis of how I

grew my career where I kind of evolved my passion for this space of IT strategy, IT transformation, program and portfolio leadership is rooted in that fundamental core principle of loving to solve problems, loving to help optimize. My husband kind of jokes and says I'm too type A for my own good. I can relate to that. I have someone at home also, same thing here. It's, you know, the relentless and ruthless. focus on optimizing everything down from like my own personal day to day to, you know, how we saw,

how we serve the business, again, working in IT, how we're driving solutions for our customers who are every, you know, employee of the business. Like, how can we ensure that we are delivering in the most efficient and effective way with, again, that relentless and ruthless focus on driving strategic value for, for, and customers. And I jokingly say, like, you know, it's, it tracks that I grew up in the space of project and program management. I am, again, I'm type A. I'm a Virgo, whatever, you know, you want to say, like, you know, kind of contributes to that particular persona. But it is, it's a persona for, I would say, people that really thrive in driving efficiency.

And I said, he says I'm too type A for my own good. He also says I'm almost to a fault, like, if it's not done in the most efficient way and it's like anything, even like cleaning the house or, you know, packing for, you know, trip and things like that. Like I try to just optimize and make, you know, as efficient as possible. anything that I do. So again, in how that translates in from a professional standpoint, again, that's where, you know, when I say I'm an AI enthusiast, it's because when you talk about driving efficiency and optimizing how you do things, like, that is what AI is for, you know, so I am going to continue to be a forever learner about how can I use the tools and the resources

available to me. Of course, security is very important. So I think, again, just being in in IT, you know, I don't think I'm the person that's going to just throw my likeness all over a bunch of different tools and see which sticks. Like, I do want to be thoughtful about what information I'm putting out into what tools that I'm using and how I'm using those tools and how my information is being used within those tools. So, again, I think taking a thoughtful approach to just, again, both professionally and personally, how can I use AI to make me better at my own job, at everything that I'm doing, at everything that I'm, you know, focusing on? and how can that translate into actual value that I'm driving, both for the enterprise as well as our customers? Now, that's a great talk. And one of the things you talked about being Taipei,

I'll divert it to a better place. What's the difference between the leadership thing AI does versus what you actually do it on the ground? Because sometimes, and this is a good thing with Taipei people, sometimes they are driving it in the way it needs to be versus what it's needed to be on the ground. So how do you use your personality type there to build better leadership within yourself

and within your teams also? How am I using AI specifically? Yeah, you and AI both. Ultimately, your AI will be your clone, right? I mean, at the end of it, it will be sort of an emotionless clone of what you do, but it will do it in a better way. So your personality would always reflect

in what you do in AI. Absolutely. And I think we talked about this earlier this week, too. Even with AI is exactly that. It is a partner. It can be even on this personal project that I tried really just with the intention of professional development, learning more about AI so that I can be better in my day-to-day job. From idea, from conceptual idea to product, AI was a partner.

The AI was in the passenger seat. I was in the driver's seat. And I think that always has to be the core. like you, yes, an AI is not going to be a replacement for your ideas. Your ideas are your own. There's a book that I read to my kids about similar. Like you want your ideas to be your own. That should never go away. And AI will never replace that. What AI can do is help polish your ideas, sharpen your thinking, give you know, a vantage point that maybe you didn't think about. But you also have to be mindful that that vantage point is based on whatever information could be scraped from the internet, you know, and surfaced into the LLM. So take it for whatever. it is, but always feel grounded in your own intuition, if that makes sense in terms of what your approach to leadership is, what your approach to human empathy, nothing replaces human experience. And that will always be the case. And I think having been in the industry now for, you know, almost 20 years, there's human experience that is valuable to me. It is part of my value. It is

part of the absolute value of the value that I drive. And I will always use that to, you know, really influence, you know, what my approach to leadership is, AI is just going to help me sharpen it. It will never replace, you know, me and my thinking. But what I will say is at the same coin, or like at the same time, I do believe that, you know, a few years from now, in every professional role, AI will in some shape or form, even for the skeptics, be infused in their day to day. into their workflows. And whatever that shape or form could be will look different.

But I do truly believe that with the accelerated timelines that we're seeing with the evolution of this emerging capability, it's going to be in the hands and in the workflows of every single professional across the world, I think. You said it well.

And you just mentioned skeptics. I think everyone, whether they don't know it or do not know it today, AI is already there. You know, YouTube is listening and then it's showing you the videos that you like. You go to and start to do a dream scrolling and then you see everything that you have done

and you're wondering why it shows because you actually clicked on something and is building the traces of it. So you are using AI or you are the consumer of AI without knowing it. So it's going to happen. One of the good stat I found, I forgot where I found it.

So call it conspiracy. We cannot define a fact here. But 3.3 months, AI is doubling. its capabilities and capacities both, right? Which means, and I did the math, it will be about million times better than what it is today by 2030. And that's a very short amount of time, you know,

just because how it is multiplying today. So I can only imagine what are the goods and what are the bats it can bring in. On that note, what surprised you the most with AI? I mean, we've both have been learning about it implementing in our lives. and what you didn't expect when you started implementing AI, meaning when you heard JetGPT,

I mean, we both have worked. I started AI in 1996-97 doing algorithms and things like that. A lot of things had to pre-plan. You had to have powerful computers. You have to figure out these machine learning algorithms and get a lot of testing data and everything. But now that all has been gone, or at least people think it's,

on. So what was your expectation and the unexpected thing that you saw with AI today? Yeah. So I mean, compute, you already kind of, you know, the compute aspect and how we keep up and keep up with the pace of the needs of AI from a compute angle, I think, is something that every infrastructure and technology company is having to grapple with. But another one that I would say, and I would say I wasn't, it's not that I am surprised that it happened, but, maybe I wasn't, but I'll tell, I'll share, I'll just share what, what it was. You know, when we talked about like the chat GBT coming out and the, you know, we created our own Mongo GBT. And then, you know, of course, Gemini, we have Gemini in our Google suite.

We have, you know, enterprise search. We have this, this notion of sprawl, right, agent sprawl is a real thing. And again, this is where I say, I can't say I didn't expect it because we saw it happen with, Fast sprawl. We saw it happen with tech sprawl, with process sprawl, with data sprawl. I mean, these are words that every scaling technology company has to figure out. And again, working in this space of program management and portfolio leadership, I have across all of these emerging techs, like have had to balance this. I've always had to play this balancing act. And me and my team and the value that we deliver, you know, I always say project management is a change management process. It was

actually born from an ITGC. Like it's a controls process, right? It's how do we ensure the controlled delivery of technology, right, in a controlled fashion and a risk-managed fashion? That was sort of at least my introduction into the space, you know, you know, 15, 17 years ago, project management was really, you know, kind of anchored on that being a very risk-managed, risk, you know, risk-focused delivery approach. But as this evolution has happened, you know, we talk about outputs and outcomes, you have to

balance process and risk and governance, dare I say the word governance, with speed, agility, value, you know, value acceleration, innovation. So how do you balance those two things? That's always been something that, you know, it depends, you know, it depends on the company that you're in, the risk appetite. There's pace of innovation, their pace for, you know, of growth. But that has always been a fundamental balancing act that PMO leaders have had to play. And so I think it's no different, that kind of, that balancing act is table stakes for the nature of my role, but how it shows up, right? What level of process? What level of governance versus, you know, how quickly are we trying to deliver this thing? And, you know, it, that's where I think it varies based on what you're

trying to deliver. And I would say with AI, because it's moving so quickly, like the old frameworks for governance don't really shape up anymore, right? Like, I'll give you an example. In the PMO space, we build intake and prioritization frameworks where we're looking at, you know, value and complexity and how long is it going to take to deliver this thing? Just the time it takes to ask those five questions, you can spin up a solution with AI. So, you know, I will say the AI agent sprawl, AI sprawl is a real thing. There is such a thing as AI governance. Now, is it going to look the same way as traditional program governance that, you know, we've typically overlaid on some of those, you know, historical technology solutions that we've delivered? No. But I think that was

you know, again, not an unexpected hurdle, just something that we are living the reality of that, right? And if we don't kind of put some baseline guardrails in place for what, what is considered low-code, no-code democratization creation that, you know, you don't need to run through some intake and prioritization effort. And you have the capability. Those don't need to go to some central tracker or central, you know, rule up versus what is this other body of enterprise scale solutions, AI-powered solutions that we're building that do need to follow because there is security implications. There is, you know, a mass change management effort, whatever the cases may be, you know, I think thinking of those two buckets in that way and thinking about AI governance as, you know, how can we ensure, again, we're delivering in the most efficient and effective way, but not losing sight of whatever the balance is needed for, you know, based on the company and risk appetite and growth pace that we have to, you know, drive towards. Now, this is a great conversation and it pointed me to a few things.

I'm looking at unexpected hurdles adapting AI. And that happens to every new technology. When you go to, let's say, an EV-based, you know, vehicles, then you have to have a charging issue. The blockchain came and a lot of fraudulent events started to happen, right? So every technology innovation will bring that risk. And you have to bypass or surpass that hurdle. Now, going back to governance, and I'm kind of putting all big companies,

tech companies on spot. Back then, you know, I'm 54. we came from a background where the software needs to be really stable, and then only you start using it. And then we see Google's and Microsofts, and I want to pick one company here, they started releasing beta versions. And if you relate our PMP head or project management head there,

that's not a good solution because you are having your customer tested out. But it goes in both ways because some customers are hungry. So they want to get their hands on it. But most solutions today are release in beta versions. They are not governed. So I would argue on AI side the same way, if those softas can be released in beta, AI can be released in beta.

Now, I understand that threat. Threat is much bigger. And that's why some of the organizations are going into responsible AI. What's your take on that? And that relates to tech sprawl that you are saying before. Yeah. I mean, I think you're 100% correct.

A beta is a, whether it's a POC, whether it's a pilot, Right? Like, I mean, even internally, we roll out solutions to a beta group, right? And that is the group that it's testing from the standpoint of the user experience as opposed to the stability of the platform. So I do think those two things, beta doesn't mean put something out into a production use case that hasn't been tested, right? What beta means is this is it's beta, right? We're not GA yet. We haven't, you know, general purpose this solution. We want to get user, it's really grounded in the user feedback. and the user-customer journeys that are utilizing, that's the specific solution. How is it actually helping them solve their problem? So I do think maybe thinking about beta from that standpoint is important.

With AI, like code generation is, that's what everybody wants. You know, everyone's using, wants to get their hands on cloud code and augment and, you know, cursor, like all these big tools that are out there. And they're very powerful. Again, as a non-coder, I've had dabbled with cloud code. And I'm like, wow, like I can build, again, digital products, build AI power. things. So again, that is becoming a democratized use case. I think it is something that companies are

going to have to think about operational rigor. It has to be table stakes, right? So we still want to look at things like percentage of defects post-production rollouts, even if those rollouts were 90% AI generated. There are still basic DevOps-style checks and balances in place that we do want to make are put in place even with beta versions. So, yeah, so that is, you know, security, reliability, durability. I mean, these are things that have to continue to be table states, even with AI-powered development. Well said.

And one of the other point I also wanted to touch base on is that AI comes with, this is the biggest hurdle with AI, which is hallucination. Yeah. How? What have you seen? How are you handling it? and what do you see in future needs to be fixed or will be fixed that you have a hope for? Yeah, hallucinations, and this is where it gets very much more technical that the engineers of my team within Mongo are super skilled at dealing with.

I will say we saw, at least from an end user standpoint, more hallucinations occurring with some of the earlier models. That's where the right reasoning, the right embedding, the way it's configured really does matter. and that's where even with beta, you have to test that, right? You have to test these things even before you, you know, release something into beta. And even within beta, that's where, you know, it depends on what those workflows are. Are you building agentic capabilities in that, you know, solution that are really going to completely remove the human from the workflow? If so, those various scenarios of testing to really test, you know, are those, what are those hallucinations that we're getting and how are those going to impact, you know, the workflow? in its journey, those have to be really tested. Now, if you're, the agents are pulling information

from various sources to give you an output that might not be actioned on, right? That's still the human that's actioning on that. That might be something you can test in a beta format, right? Like that, you could go before you go GA, get a few users, maybe, you did that with some of the internal agent solutions that we've created, ask it, certain questions, I'll go to market one, like pull up information about a certain account. Is that information the right information or is it hallucinating, right? So that's the kind of user feedback that we do want to get because those users are the ones that are closest to the knowledge and the accuracy of that knowledge. But it is a very real thing. You know, I with certain, you know, you see prompt libraries now all over LinkedIn. People are

suggesting ways to kind of contain the amount of hallucinations that you can get. I do think that a lot of the latest models, like with Opus, for example, I haven't really had that problem. It's become, you know, I've stored all my memory there. And, you know, really, of course, you have to review everything that an AI is generating for you, and it's been pretty spot on. So I've knock on what I've been. I haven't seen it as much there, but there are things that you can do in prompting that can contain the level of drift, if you will, with the data output. Yeah, no, that's very true. And one of the things I remembered or what is happening, I spend about 30% of my time learning and where AI is heading and where the world is adding. in technology. What do you see a big shift today?

So there are three things. One is prompt, prompt engineering. Second, which people are discounting a lot more is context engineering. And AI hallucinates a lot more because it's missing context,

because it only has a million token, the best model. The other ones only at like about 300, 400K in terms of token. And token are loosely coupled like words. So only those many number of words, will remember. So then people started to evaluate, and I build my own second brain using obsidian, a story for another day. But it started to remember everything. But now that

everything's still too much. So even that needs to be summarized. And when you summarize things, things are going to miss out. Right. So context engineering is one thing I see as a big challenge and it is being solved. The third one I see, which is the biggest of all, and Elon Musk of the world and others are trying to solve it as well. Infrastructure. I really see AI itself as an infrastructure. So recently I saw innovation quite tiny AI, not a plug to them. I mean, I'm still curious to see what they are doing.

But it's like a small hard drive. And it has all the AI models of the world sitting in it. And you don't have to pay for it. It's really powerful. And you can plug into your Macs. or any device. And so you are bringing your own AI. So that's an AI as an infrastructure.

I see that as a big shift, right? That's what is coming in AI. The third that is coming. So along with context, I should have said that, is memory. So memory management is a big deal. These are the three things which is coming that I see. And, you know, there are certain evolution or evolution or innovation that is happening.

So you see OpenClaw has been talked about. NVIDIA jumped onto her. So they created Nemo Claw. Then Claude Code has started producing a lot more features, such as remote dispatch and some of the other features. This is what I see coming. What do you see coming or what do you want to see it coming also is the question here. So from your own vantage point.

Yeah. No, I think that's a great question. Just the memory aspect alone, I think, is huge. like I will be the first to say even in my current workflows prior to Claude being, we do have, you know, cloud now rolled out, you know, across the enterprise for engineering at Mongo. But on a personal standpoint, too, like I had a running GBT from when I was using chat GBT. I had a running chat because that's how it had quote unquote context or memory, right? But then I hit my chat limit.

So then I would copy paste that chat and go to another chat. Right. And then when I created my own personal cloud account, I copy paste it all of that and I stored it as memory. So I had to kind of do my own like stitching to create a starting point of my, my helping my brain, my brain partner, which is now what I do use is primarily Claude. But yeah, so I do feel like exactly what you said, the context, creating context as infrastructure. That is your starting point. How do you create the right starting point? So you're hitting all the things that you said. You're minimizing hallucinations. You're starting from the right vantage point. It's indexing on the right priorities too, because what I've also noticed, again, just from a personal individual user, I'm giving it certain information. And then when I need to need help writing an output for something, it's not wrong, but it's prioritizing data points that I would have never from a, the indexing was it was indexing on things that I would not have considered pulling into whether it was a report, a memo or an update. It's because it didn't have the full context, right? So I 100% agree that is a place where I think there's huge opportunity in all of the models, right? Is A, you know, we were starting with certain perplexity and things like that where you can, you know, has access to all the models.

That's great. But using that to actually build a context layer infrastructure base that starts, that is your starting point for, you know, it has the memory. It has the full context. So you can truly use it as a comprehensive brain partner for your individual workflows. as opposed to trying to stitch together the way I did. Yeah. Now, this is a great solution, and that's a good tip for the listeners as well,

that you need to build your own context, and I'm building an obsidian, but you don't have to. You can simply put it in a Google Drive, a Word document or something like that. But maintain it, an AI, you can prompt it out so that whatever you talk to AI, it can start saving it as a transcript to it,

or you can copy, paste either way. and then keep feeding back so that there's a great example I use with my team also is that what is context, right? So when you go to someone's house and you say, they'll say, what do you want to eat? And you'll say anything. But then you have lactose intolerance,

you're a vegetarian, you have certain other things. You never give any context there. And now you're expecting someone to guess your mind. That's what the prompt is. You're giving one line to the prompt and you expect that that LLA model will guess it. It will guess it. But it has millions and billions and even trillions of data points for it.

And it will pick up something randomly that might apply to you. Or maybe which is majorly being used on an average. And that's the hallucination in a nutshell. Exactly. And that's such a great example because you just talked about, oh, the lactose intolerant, and vegetarian. These are binary zeros and ones. But how about what have you eaten for the last 10 days?

are so true. Yeah, like these, that contextual, historical trend analysis is also part of that. It should be part of the context in terms of what it's going to recommend to you to eat next. And yeah, I would say on that we're not using it as, at least I'm not using it as much anymore. But on the, we were talking about projects and kind of the project tracking. Notebook LM for a long time was, was a helpful.

I love that. Yeah. I love that. Green G sheets and G drive all the way. So when you can, instead of creating project folders with all the things, like just creating notebooks was super helpful because then I could ask it exactly. Like it had targeted context because we would use it for specific programs and specific strategic initiatives. So throw everything associated with that initiative into a notebook LM and make sure all of your

artifacts are there so that it did have very targeted specific context. But yeah, I think, you know, it still required the steps of going and creating the notebook ensuring that you're connecting all your data sources. I think the mass, like at least again on a personal level, like here's my Google Drive. here's my email, here's my calendar, here's all the chat history and all of the things that we've talked about. And that's really, like you said, the infrastructure layer of how to set context. If we can really optimize that, then it's just going to, it's going to bolster the productivity that, you know, gains that we are able to get from these tools. Great. No, I think you keep your eye on the things that how it is saving the world or helping the world, whatever we want to call it.

what would you say to the CIOs to prepare for now? I'm not even talking about future, but now, because they might be lagging behind from where we are today. What should they be doing? And especially these are skeptics, right? So some people straight out reject AI. And rejection on anything is not good, right? It's like earthquake is coming and nothing will happen in last 10 years or 20 years.

Nothing will happen. But it could destroy. So what would you do to prepare it? Similar thing for CIO. It's a storm and earthquake coming to you already here. What should they be doing to get started? Definitely if you already have an AI strategy. It's important.

This is not something to sit back and see how it evolves. You need to have an AI strategy. You need to think about agentic infrastructure. Is that something you want to build in-house? You want to go with one of the hypers. You stitch together a number of point solutions. Like these architectural considerations need to be factored into your AI strategy because that is going to inform five versus build decisions. I mean, everything.A.I.

is out there. Every company is, you know, kind of trying to, you know, hit the gold rush in terms of the, you know, AI products available to enterprises, to drive efficiencies in their, you know, in their ecosystem. Of course, the number of question that every CIA and every tech leader is asking is, how are you using AI to, you know, healthy enterprise? How are you using AI to healthy enterprise? Having that strategy defined, having that, you know, those technical architecture, decision points defined up front is going, that can inform every buy, every build, every, you know, solution creation, every project or initiative, AI powered initiative that gets prioritized. The fundamental of it is having that AI strategy aligned to both with, of course, at the CIO and,

you know, technical leadership level, but even with the rest of the, you know, the C-suite. And so that's what we did at, you know, Mongo, the first couple of years, I would say, we're definitely innovation center. Let's, you know, get our hands wet, figure this thing out. build the Mongo GBT, build that, you know, other things without really a clear strategy. And this year I feel, especially with our new CIO, you know, we have a very clear strategy of how do we want to propel AI across the enterprise. And so I think that's really the most important thing. Think about AI governance. What does governance actually mean to your organization, again, based on those things that we talked about risk versus about innovation and speed, how do we want to

balance those two things? And having some semblance of it doesn't have to be these bright, shiny, complex, robust frameworks. It's just a lightweight thing that you're going to use to assess power to emit, you know, AI initiatives and what your architectural approach is going to be to solving those, you know, those problems and building those AI solutions. Think about that now. Define that now because it's very quickly going to become too late. Yeah, it's on your head already and you just have to deal with it. Whether deal with adaption or deal with resistance, it's up to you. It's going to hit you very hard. But AI is a friend is what people should take out of this discussion right now. One of the thing I also want to

pick up on, you just mentioned that your organization is using Cloud Code. We are using it in a very siloed way because we're a small consulting shop. So I'm a little bit curious. Every strategy, you know, strategy has few components. You have strategy. You have strategic initiatives. Then you have phases, projects, workflows. And then you have measurement, KPIs. How are you measuring or how the organization is measuring, cloud code, performing, and comparing, because that's what most CIOs and CTOs would do. They will compare with the old world. This is how our developers were performing.

This is the KPI we had. What's the difference? Are you noticing anything? Is there anything in place that is giving you some differentiator? So we are in the maturity evolution. So I will say that. We're building out what some of those KPI should be. at the minimum baseline,

the percentage of AI-assisted PRs is something that we are going, we are starting to track now, right? Historically, we really didn't have that. How much of our code is generated purely from AI. And then what are those delivery like? What is the cycle time? How does the cycle time compare to what it was prior?

So these are, we're building, we're in the process of especially now reestablishing our engineering productivity, develop or productivity metrics for success are operational excellence metrics for success. And so these are some of the metrics that we're tracking. They, I think the story is still unfolding, right? Like time is now evolving to see how is this trend over time occurring. Are we seeing improvements in cycle time as a result of uptick in code generation, you know, code generated from AI? Are we seeing the complementary, you know, tread down in actual cycle time? Those two things, you know, those in

are what we're working towards now. Okay. Now, that's again a great insight. And I love this conversation. I know it will keep going into a lot more details if we keep unfolding. But two questions I don't want to miss, which is coming to your own personal side. And, you know, for me, three things are very important, you know, leading with empathy, motivations, and, you know, technology or solving business problems using technology. We talked a lot about it. Leadership, we talked about it. But then there are some personal side that you mentioned. The children's book story, I have a 15-year-old, and he tried a lot of different things. He also uses AI. I actually have a new company with him called Admit Shure.

So while he's preparing for getting into different colleges, I said, why not you start building? So we got funded by Google, and he started seeing those powers. he got trained a little bit. And he saw goods and bads of AI. And he can kind of coach to the people now. So I'm pretty proud of that. You have a similar story there. Let's talk about it, the children's book story you had.

Yeah. So my husband and I both work in tech. So we are very tech forward, I think, in this house. When we got our Google, Gemini was first introduced. I have a five-year-old and a seven-year-old. Both girls, they were having such a blast with asking Gemini, Build me a picture of a princess riding a banana holding a hot dog. And it would. It was random things. So we would have some fun with that. My seven-year-old is in second grade. They have

Chromebooks this year. So inherently the world is different, right? When we were growing up, my teachers wrote on chalkboards. We had notebooks and pencils and nothing, right? We didn't even have cell phones in the house at that point. So it is inherently a vastly different world. We are very conscious about how we allow technology into their hands. I mean, simply put, we don't even let them use their iPads or watch TV during the week. So there is that. We still are very, I would say, controlled and conservative when it comes to screen time and things like that. That being said, there is no reason why a seven-year-old can't enjoy the power of AI. And so the example that you were talking about the other day is my seven-year-old, she's a wonderful writer.

She loves to write. She loves to read. She's super creative. She writes her own little books. And so over Christmas break, she had written a book, a children's book. She even drew all the pictures for it. The story is about little two moons, a mommy moon and a baby moon eating. They make a star cake. It was a very sweet story. She drew all the pictures for it.

We sat together. We uploaded her pictures and her story into Gemini and asked it to basically create, lose her images as a starting point, but create them into. children's book ready images, keep the same style of the pictures and kind of round out, I would say, the language and that is from a grammatical standpoint. And if there's, you know, missing pieces from the sentencing structure, you know, fill those in. We did that together. I got her comfortable with giving it a prompt. She's also learning how to type in school. So, you know, prompting, you know, Gemini specifically on, you know, I want to add this. I want to add that. And we built a

children's book for her. And I'm in the process of getting it stitched together and printed. And, you know, we'll read it every night before bed. And so, yes, one could argue, oh, well, is that, you know, hindering creativity? And I say absolutely not. That is empowering her to think beyond the bounds of what, you know, she's able to put pen to paper on and giving her the tools to be able to prompt and to, you know, kind of find the way to take her own creation and create it, you know, a product, an actual physical product out of it is the journey that I'm helping to, you know, trying to instill in her at this age.

But, yeah, so I think, you know, it's here to stay. I think AI, using AI responsibly is super important. And again, even for the children, you know, like this is the, I think, the extent of what we want to be able to showcase to her at this time and, you know, have her, you know, develop her own, hopefully passion for technology the way her parents have. Yeah. No, that is great. And, you know, you're encouraging that. That's really good. A good example I also use about technology.

So if you're a furniture maker, you use advanced tools. that does not take away or scale your creativity. Same as when you use calculator, when we're studying, you know, you could do calculation on paper, fine. But that's a repeatable task that you could let a calculator do it. Same thing happened with computers.

So you have adopted technology all throughout wire resistance on AI. It's the same thing. It'll bring a good and bad in everything. But if you adopt a good thing and you look at the good things, it's going to help you. And I definitely encourage kids,

I, in fact, have a movement for disabled kids where I'm providing free education on data and AI. I have a school community on that too. And I really want to encourage everyone to learn AI. Learn with the caution that it's going to bring a lot of risk to you. So like anything else, you use a digital card online putting on Amazon, same risk you have, but in a different scale and proportion. So you need to be educated about those things and know how to respond if and one the risk gets triggered as opposed to ignoring it and acting like it's not going to happen.

Yeah. And everyone needs to learn and educate themselves and educate others. And that's the key for any technology, including AI. One closing thought and question, and this I ask everyone, what would you tell your past self, you know, your 16 year or 18 year or 21 whenever you were there? And I have a lot of story about myself too. I was, you know, my father used to pick me up, put me in school. So I got my wings when I started wearing braces.

So I didn't want you to study so much. I was good in studies. But I wanted to travel. I went into music and I started working in different places. But one thing I missed at that time is I was not too reflective and I was ignoring what is happening around me and not enjoying. So I tell my past self that I should be enjoying that journey that has happening a lot of great things has happened, but I kind of keep moving on. So what would be your thing?

That's such a great question, Dave. Thanks for asking it. I think always trust your gut and have trust and faith in your intuition is what I probably would have told myself. Because, you know, when I, and I won't share my age, I'm well-classed eating at first birdberg. Well, in any case, I'm older so that we'll settle down to that. Yeah. But it's, I remember it like it's yesterday, almost. That's the irony of the whole situation. It was not anytime reason, but I do remember it as if it was yesterday. And just with this, like we're talking about AI and technology, you know, my husband and I, we were recently saying, too, we are the last generation that remembers a life before all of this. Like, internet. We had a TV in our house. That was it. And it was this big boxy thing. And I had a big, you know, computer. No, when I grew up as when I was my daughter's age, no iPads, no cell phones, no any of that. So and we are the last generation, I think, to like have that preserve that memory. So by the time

I was 16, of course, I think I had a very big looking cell phone. But, but yeah, I think it's your, we're such sponges for knowledge. And that's a good thing. But I do feel like just reminding myself that, you know, your intuition, even and especially with AI coming in and all of these data is an abundance today and is quite difficult to navigate through all of it and know who you are and know what it is you want. We were talking about the other day. I'm getting influenced. You clicked on maybe one AI arbitrage ad on my Instagram and now I'm getting 15 of them

or your AI consulting business or, you know, AI arbitrage business. So, you know, yes, that that is happening. We're all getting influenced by, you know, the data around us and the channels around us. And those are so much more relevant today than they were, you know, when I, when we were, you know, 16 years old. So I think, you know, just the what I would have told myself back then, what I will tell my children, you know, because they are kind of the mirror of what I wish I could have told by my, you know, 16 year old self is just trust your intuition. Lead, you know, really truly believe in your, you know, it's, it's almost like manifestation mindset. Like think about your gut is what's going to tell you what your true path is and always trust that intuition. and lead that path with intention.

Don't just let the world influence you and bounce you around without really asking yourself your question intentionally, what is the path that I want to go on in my life? And just trusting that. No, this is so beautiful. And I also relate to that. And I just thought these are four amps that I look at. So my doom scrolling is about mindset, motivation, music,

and there was one more REM. So there are four amps that I constantly look at see. So I'm welcoming all that AI sending those stuff to music, right? So it's sending it to me and I love seeing those things and, you know, I love connecting with people. So thank you so much for connecting with me. You have been incredible. And where can people find you? I know you're on LinkedIn, but is there where people should look at and find you? And I'll provide a link to people on my blog if that's okay. That would be wonderful. I love, you know, similar to you like I'm I'm just excited to be in a space that has so much opportunity for learning.

You know, thank you for reaching out and connecting this. I'm so happy that we, you know, we were able to meet and we have each other in our network now. And so, yes, I'm available on LinkedIn. Feel free to share the profile. I started writing recently as well. So I have a substack and I've just been sharing insights, you know, that I'm passionate about. And I'm really, you know, I think entering a phase of my growth and my professional journey of, you know, continued learning. And yes, I want to continue learning from both experts, novices,

people that are in the journey with me, whether it's AI. I have a very strong passion for music as well. So we share that in common. But yes, thank you for this, you know, wonderful dialogue and just looking forward to, you know, staying connected. Thank you, Priya. That's Priya Odishi, Chief of Staff to the CIO and heard of ITPMO at MongoDB. If you're thinking about AI strategy in the enterprise, that is someone you want to follow.

She just mentions Substack. So do go ahead and follow her. Have a great day. Thank you. You have been listening to Think Yeah, podcast with Dave. Take one idea from this episode and turn it into action.