MongoDB.local Bengaluru 2026 General Session

Thank you, Chirantan. Good morning, everyone. Come on, there's a lot of people in here. Good morning. All right. I'm excited to be back here in India. This marks my 15-year hiatus from being here, so this is my first time in Bengaluru as well. So thank you all for having me. But it's not just for the craft beer scene, which I hear is picking up here. It's really for the tech and startup scene which is moving faster than almost anywhere else on the planet. By any measure, India is one of the world's most important AI markets already, accounting for nearly 6% of the global market share and on track to grow tenfold by 2032. The adoption and momentum we're seeing in this region is absolutely remarkable. Many of you probably think of MongoDB as just another database company, and that's maybe fair. MongoDB serves as a system of record for tens of thousands of applications, nearly 75% of the world's largest companies. But that's only part of the story. Because AI changes what a database needs to be. It has to do a lot more than just store and serve your data. In the AI era, your database needs to be a system of intelligence with the ability to reason over your data in real time to drive outcomes that actually matter. This is what MongoDB has become. An intelligent data platform enabling you to build and run production AI applications. That is how many companies are using MongoDB. Today, Infosys, one of India's largest technology services companies, is helping its enterprises build AI-ready architectures. CoreAI, an AI-native unicorn serving organizations around the world, is building intelligent applications and agents on top of MongoDB. The reason these companies are seeing success is that we've solved two of the biggest challenges developers face with AI: trust and compliance. So let's start with trust. To trust AI applications, you need them to retrieve the right information in the right context at the right time. Wrong actions drive up costs and can have disastrous impacts on your business. You can't deploy an application if you can't trust it. And the key to trust is accuracy. Your model is only as good as what you feed it. If an LLM is working from incomplete, outdated, or poorly retrieved information, you get responses that are wrong. You need to get search and retrieval right. And that's exactly what our search and vector capabilities combined with Voyage AI embedding and rerankers help you do. The second challenge, compliance, is more important than ever given data sovereignty and privacy rules such as the Digital Personal Data Protection Act here in India. Being locked into a single cloud is no longer an option. It's more important than ever to have the flexibility to run your applications across multiple environments, whether those are multiple clouds or on-premises or some combination of the two. That's where MongoDB's freedom to run anywhere means you get the same experience wherever you need to be without sacrificing security, consistency, or control. So just to be clear, you need trust and compliance to deploy AI applications in production at scale. Today, we're going to take a deep dive into how MongoDB and our latest product releases allow you to achieve both. To get us going, let's cover what stops most teams before they even get to a deployment: accuracy. And there's no one better to walk you through this than MongoDB's Field CTO for AI, Pete Johnson. Pete, well done, sir.

Thanks, Ben. And thank you all for joining us here today. I'd like to start by telling you about a company that has a major presence here in Bengaluru. How many of you have heard of Lena AI? Okay, for those of you not familiar, Lena AI is a pioneering company that builds employee-facing pre-trained AI agents that handle HR, IT, and finance requests at enterprise scale. And their client list is a who's who of leading companies: folks like Coca-Cola, Etihad Airlines, and Abbott Pharmaceuticals. So the stakes are incredibly high. These agents are trusted with mission-critical business processes like payroll, benefits, access provisioning, and financial reconciliation. A wrong answer isn't just frustrating; it can create compliance issues or financial discrepancies. Lena's customers need to trust that the agents will produce the right results, and they can because Lena gets search and retrieval right. Lena leverages MongoDB as both its system of record and its system of intelligence to ensure accurate answers in the appropriate context. Now, accurate retrieval is the key to building trustworthy AI applications. And yet, this is where many companies struggle because they can't trust the application will consistently provide the right answer. When an agent gives a wrong answer, the first instinct is often to blame the LLM. But it turns out this is less of an LLM problem and more of an accuracy problem because it's the retrieval system that's getting it wrong. Accurate retrieval is critical for agents. To understand why, let's quickly ground ourselves in how AI agents actually work. Agents perceive what's happening around them, plan how to respond, then act on that plan. The loop then repeats. What makes that loop coherent across sessions and over time is agent memory. You see, agents by themselves are like goldfish: every event is a fresh start with no recollection of what worked before. Memory changes that. Memory is what allows agentic systems to learn and make smarter decisions as they go. And accurate retrieval is what makes memory work. Whatever an agent pulls from memory is what it reasons over, plans with, and acts on. Inaccurate retrieval systems don't just cause misinformation; it causes compounding negative effects across every step that the agent takes. So then the question becomes: what drives accuracy? Well, it's three things: your retrieval method, the quality of your embeddings, and your use of a reranker. We're going to go through each one. First, the retrieval method. Most people start with full-text search, which is exact keyword matching. It's fast and precise, but it only finds the specific words you search for. In contrast, vector search goes a step further by understanding a user's intent. So it can find information based on semantically similar meaning to the actual search terms. Let me give you an example: a query for 'car repair' can return results about 'fixing my vehicle' even without those exact words being in the query. Both full-text and vector search have real value, and hybrid search combines them, giving you the best of both worlds: the precision of keyword matching and the contextual understanding of vector search. Until now, implementing hybrid search meant running both queries separately and writing your own logic to merge those results, which you have to maintain your own combining logic over time. This is labor-intensive and manual. We wanted to find a better way, and we did. Today, we're announcing new hybrid search capabilities in MongoDB. Using one API call, you can now choose the best way to combine results, either based on blending by their relative ranks or by their relevance scores. Our new hybrid capabilities give you better retrieval with less complexity and less code for you to maintain. The second thing that drives accurate retrieval is embedding quality. Since vector search results are only as good as the embeddings behind them, an embedding model is used to convert everything from PDFs to images to audio or even code into an array of floating-point values that capture the meaning for software to process. And the key is this: embeddings work by proximity. The closer two embeddings are, the more similar their underlying data is, like the car-vehicle example I used before. Better embeddings mean more accurate, more relevant results. And that's exactly what Voyage AI brings to MongoDB. Voyage gives you industry-leading models to improve retrieval accuracy while lowering your cost of getting into production. Our Voyage AI models rank at the top of Hugging Face's Retrieval Embedded Benchmark, beating Google and Cohere AI. We have general-purpose text embedding models that convert pure text into embeddings. We also have multimodal embedding models that can vectorize text and images together while maintaining high retrieval accuracy. Now I have a question for you. What if your applications need to search really large, complex documents? Think legal contracts, technical documentation, or a financial report. These can be a challenge for many embedding models. If you've tried this yourself, you know what I'm talking about. The solution is to create chunks, breaking the document down into smaller, more manageable pieces. But it turns out that getting the size of the chunks right is difficult. If your chunks are too small, you lose the surrounding context. If your chunks are too large, storage costs start to climb and your precision can drop. So developers end up iterating across different chunk sizes over and over to try to find the right balance between storage cost and retrieval quality. We wanted to help you avoid all these iterations and make chunking easier. Well, our Voyage context model solves this exact problem. And today I'm happy to announce its newest release, the Voyage Context 4 embedding model. Each embedding encodes not just the content within the chunk, but the broader context surrounding it. This approach provides both precision and context to offer better retrieval quality at a lower storage cost. So, we've covered retrieval methods, introduced our new hybrid search capabilities, discussed our industry-leading Voyage embedding models and our newest contextualized chunk model. But here's one last thing you need in order to improve retrieval accuracy: ranking. Even with great retrieval methods and strong embedding models, you don't always get the best results at the top of your search to feed to the LLM. Rankers fix that. Rankers take the results and reorder them to improve accuracy and relevance. Think of it as your retrieval results casting a wide net, capturing fish from the sea, and the reranker as the hand that picks the best fish out of it. Until now, applying reranking meant extra steps: you run your query to get your candidate results, then you make a second API call to get your reranker. So, more code and more things for you to maintain. If only there was a way to do vector search and reranking all in one call. Well, now there is. Today, we're announcing the public preview of native reranking. Native reranking introduces a new $rerank stage to the MongoDB query language. You just add that to your queries and your results are reranked on the back end, all with a single API call. Now, to get all this capability, you could try to stitch together third-party search, embedding models, and rerankers yourself, and some people do this, but that means a lot of maintenance and manual integration work. Every time you iterate on your application, you have to check all those stitches of each individual component, and if something goes wrong, you have to troubleshoot each individual piece of a very complex puzzle. MongoDB solves this complexity issue, giving you a single, simpler solution. Now, to learn more about search, embeddings, and rerankers, you can join our workshop on building production-grade AI agents with MongoDB and Voyage, which takes place in this room a little bit later today. This will be an interactive session that will give you practical tools to help you build transformative applications. So, smarter retrieval methods, better embedding models, and more precise reranking all make agentic memory more accurate. This means better context, better decisions, and more trustworthy results. Your users won't know about hybrid search, Voyage embeddings, or the new $rerank stage. They'll just know that your product can be trusted when it matters. That's what a system of intelligence enables. And with MongoDB and Voyage AI, that's what you can build today. Thank you very much. Let's bring Ben back out.

Thanks, Pete. The ability to trust your AI applications is absolutely essential, but trust alone isn't enough. You also need to meet your organization's compliance obligations. And to do that, you need the ability to run anywhere. When we talk about compliance, it's not just about security or things like data residency rules or privacy policies. It also means operating within your organization's internal governance and infrastructure requirements. And you have to be able to use your company's approved cloud providers and AI services. So let's think about one of the companies I mentioned earlier: Infosys. They serve enterprise clients across a range of highly regulated industries. Their clients operate around the globe under different regulatory regimes with different data residency and infrastructure requirements. And just like here in India, these requirements are constantly evolving. These companies need the ability to adapt as needs change without constantly rebuilding applications or being locked into a single environment. And that's what MongoDB's run anywhere capability provides. With MongoDB, you can operate wherever and however the business requires. Whether you're developing on a laptop or running in your own data center, MongoDB is there. And if you're developing in the cloud, Atlas allows you to work natively on any of the hyperscalers. This is what's unique with Atlas: a single MongoDB deployment can span multiple cloud providers simultaneously. This gives organizations the freedom to architect around their own needs while also increasing resiliency and reducing dependency on any single provider. No matter where you build, it's the same platform, same API, same set of skills. You don't have to learn a new way of building as your requirements evolve. This flexibility also needs to extend to the AI services you choose. The AI ecosystem is evolving incredibly quickly. There are constantly new models, new infrastructure, and new services. Maybe one of your teams is using AWS Bedrock, maybe another uses Google Vertex AI. Six months from now, the best option for your use case may be something entirely different. MongoDB allows you to choose the services that best fit your needs. The ability to run anywhere also helps satisfy your privacy requirements. You might need tight control over your data to protect user privacy or intellectual property, ensuring it never leaves your security boundaries. And the need for flexibility extends to search and retrieval as well. Until now, teams running on-premises or in private cloud environments often had to stitch together separate systems to get the retrieval capabilities they required. Today, this ends. MongoDB search and vector search are now generally available as an option for Enterprise Advanced. The same retrieval capabilities Pete discussed are now available not just in Atlas but also in your own environment under your operational and compliance controls. That means you can build AI applications on whichever platform works within your compliance framework. To get started with MongoDB search and vector search on-premises or in private cloud, join the session at 3:00. And for more on how running anywhere can help with your data requirements or compliance rules, your data and your rules at 4:10, which is another session down the hall. This session will specifically address the Indian Digital Personal Data Protection Act. So let's step back. AI is moving fast, models are getting better, infrastructure is evolving, and new services continue to emerge every day. It's no longer enough for your database to simply store information. You need an integrated data platform that enables you to build applications that deliver on the promise of the AI era. A platform that gives you the most accurate search and retrieval possible so your applications can be trusted. A platform that gives you the freedom and flexibility to run anywhere your business requires. That is what MongoDB is: a system of intelligence that exists to make sure you have what you need to deploy AI in production at scale. But having the right foundation is only part of the story. The other part is how companies turn innovation into meaningful business outcomes. So to talk about that, please welcome MongoDB's Chief Customer Officer, Erica Valini.

Well, good morning everyone. I am thrilled to be back with all of you here in India, a place I very much love, and I'm happy to be back. I think this is my maybe 15th trip here, and every time I'm just immersed in the culture, and I'm so excited to be with everyone here today. So, a little bit about me. From two decades advising organizations through their change journeys to having a front-row seat to large-scale technological change at both ServiceNow and MongoDB, I have spent my entire career helping organizations navigate technological shifts: cloud transformation, data transformation, and now AI. And across all of those shifts, there have been two constants. The first is India, a country at the heart of where innovation starts and scales. A place that has shown the world what large-scale transformation actually looks like: how to fully embrace new technology and how to keep change at the front of the agenda. The second is the realization that technology shifts are never just about technology alone. They're about organizational change, adopting a new mindset, and having the courage to do work differently than it's ever been done before. Because while every single technology shift is a little bit different, they all arrive at the same exact question: How do you turn new innovations into meaningful outcomes? Now, earlier you heard from Ben and Pete, and they showed you the incredible innovations that we at MongoDB have to offer. And what I want to do now is use a few minutes to talk about the outcomes that we can now deliver together. And in my opinion, there is no better place to have this conversation than here in India. Because at MongoDB, we're not just selling in India, we're investing here. We're betting on the belief that India is not just going to adopt AI, you're going to shape the future of it. And this week in all my customer conversations has shown me just how true that belief is. Over the last few days, I've traveled everywhere from Hyderabad to Gurugram to here in Bengaluru, and I have heard incredible stories from our customers on how they're adopting AI to change the way that work happens. I had the privilege of spending time with Physics Wallah, a company that I'm sure many of you know very well, and they're using AI to create their AI Guru, which is an online agentic tutor that furthers their mission to drive down the cost of education so it is within the reach of every single student in India. And then I spent time with Zomato, and they talked about their latest advancements in AI, including the launch of Nugget, which is a new end-to-end customer experience that sits at the heart of their goal to bring fast retail to life. These are two very different companies with two very different missions, but both have had the courage to challenge the way that work gets done, to shift the mindset of both students and consumers alike, and to embrace AI not just as a technical capability, but as a completely new way of working. These two companies have figured out not just how to make the technology work, but how to make the technology work for them. And they're not alone on that journey. When I came to MongoDB just four months ago as the Chief Customer Officer, I had one primary goal: to make sure that we're giving our customers, all of you sitting here in the room, the best support we possibly can. And there was no doubt that AI has an enormous role to play in doing that. But how big of a role was really up to me. And the way I saw it, I had two different paths that I could choose from. The first, a standard efficiency...

Thank you. Thank you for inviting us. Super excited to be here.

Yeah. So I'm founder and CEO of Emergent. Emergent is an AI-native platform that allows anybody without any programming background or technical knowledge to actually build production-grade software just by chatting with our agents. We started our journey about a year back, launched about a year back. Today, one of the fastest-growing startups in the world with 10 million users, 12 million apps have been built, hundreds of thousands of apps deployed in production. And we started with this core insight that people who are closest to the problem should be building software themselves. Most of our users are business operators, new entrepreneurs who do not have a tech team, and they are using us today to build software that is close to the problem.

Right. So when we started our journey about two years back, we actually started off as a research lab building coding agents. And at that time, everybody was building copilots. That's the second stage that you saw in the slide. And we had this massive belief that soon you'll have agents which can actually do long-horizon tasks and take abstract goals and actually take them to completion. And we started building these coding agents. We became world number one on SWE-bench, which is the benchmark for all coding agents. And our belief was that as AI exponential continues, you'll have agents which can actually do these complex tasks. And we started thinking about how do we truly democratize software development and build things end-to-end for people.

Right, and we think that we're just getting started. The kind of complex things that we are seeing on the platform is amazing. Yeah, a true founder mentality. I know both of you were just getting started. I think we'll hear that. Jatendra, let's talk a little bit about your journey.

Good morning everyone. We started eight plus years ago. So in simple words, we build purpose-built AI agents for customer experience. Let me explain. So as we all reach out to customer support for anything and everything, millions of conversations happen through voice, chat, and emails. But what happens in the contact centers is only a few percentage of these calls, maybe 1 to 2% of calls, would be evaluated for quality, compliance, and all that. That's where we thought an opportunity to automate this. This is again 8 years ago where deep learning, speech-to-text, NLP, or NLU were getting really advanced and the accuracies were becoming at a stage where you can apply in real-life applications. We saw the opportunity and that was kind of a wow moment for us: we could really automate 100% of customer conversations through our platform and help the contact centers to do their workflows much more efficiently. It not only saves their time but also improves customer satisfaction and in some cases sales conversion rates. That was half of our journey. Then last two, three years we pivoted more into building agents. As Mukun was telling, we started with copilot use cases, we call companion agents. We've really seen the adoption. End of the day, we all use as developers cursor and all, but we really want to provide that same experience for our agents too. Human agents really love it. It not only saves a lot of time for them but also helps them to do mundane tasks like providing notes at the end of the call or fill some of the CRM at the end of the call. All of these things are done using AI.

Yeah, I think one trip me and my co-founders went to Philippines. I mean we know that contact center, Philippines is a hub of contact centers, millions of agents, and it's really operate the city is amazing and it operates only in the night. So just to work with the US, we went to a lot of these contact centers, observed these agents. The one thing that really hit hard for us is their life is very busy. They take call after call and put a lot of sticky notes across their desktop. That's where it really hit hard for us: okay, why can't we automate some of those things? And that's where my experience coming from a long background of research in speech and NLP, I really thought, and before that I was building a lot of these applications for consumer side like for phones and other digital production. I thought this is where we could really make the difference in life. And that was the aha moment for us. We could automate a lot of these processes and simple listening to more, getting actionable insights from these customer conversations.

Yeah, so MongoDB for us was a very deliberate choice. When we started building, our goal has always been that you come and describe what kind of software you want to build and we take the responsibility of planning, testing, debugging everything behind the scenes. So we started reimagining how a modern software development system should look like. It has to be end-to-end. We manage our own hosting, deployments, everything. Very early on, we knew that agents are going to do both because our users are going to be non-developers, so they won't understand the technical nuances. So when we started designing the system, we experimented with every single database that existed at that point and figured out two core insights. One was that because our users were building apps for the first time, their requirements were evolving as they go. For example, they'll start with a simple CRM, then they want to add payments, then they want to add a WhatsApp integration. Agents had to figure all of these things out on the fly. The software is evolving as we built. What we figured out with a lot of experimentation was that MongoDB really fit this need: if your software schema is evolving, you need a schemaless system. We saw a lot of our early competitors actually get stuck in this migration loop where database operations could get stuck. In fact, one of our competitive advantages early on was using MongoDB, which allowed us to pass through all of these migration and schema issues. We have been scaling on it since then. We believe internally that MongoDB is the best database for agentic coding. If your agents are coding, MongoDB is the most versatile, especially given the kind of use cases people are building on our platform: from manufacturing to shipping to a store owner trying to build a storefront. If you need this wide variety of schemas, and especially now a lot of people are building AI apps, they need RAG, vector databases, all of those things. MongoDB really helped us scale. It was a very deliberate choice and we think it was one of the core advantages we got early on. Now we are seeing a lot of our competitors also moving to MongoDB.

I also happen to be like first thousand customers of MongoDB. Like when it launched, I was one of the early developers, so excited we're building on that.

Yeah, for us, I think it's a natural fit. As Mukun said, the document model really fits our data because what we have is raw conversations and we extract a lot of insights, intelligence from these conversations. Also, a lot of these configurations keep changing as we build more and more AI models, more and more enrichments. So we don't have any fixed schemas to start with. I think we made a very early choice of MongoDB. We are more than six years working with MongoDB. The first reason is natural fit for the data we have, and the second is the flexibility. We don't want our engineers to spend too much time figuring out schemas but rather build the products or iterate faster. And we could really iterate much faster with MongoDB. For example, now we have more than 100 clusters and more than 50 terabytes of data on MongoDB, and it became a source of record for us. All our enrichments over the raw conversations, a lot of these configurations of different models and now different agents, all stored in MongoDB. It sits between our data layer and the orchestration layer nicely, and we scale enormously on MongoDB over the years. The other big part, as the previous speaker said, is trust and security. That's really key for us. So all three points: the natural fit, the flexibility to iterate faster, and the scale and security.

Yeah, I think the first one is providing reliable service and accurate service. Since we work with enterprise, these two really matter very much. With all the AI, both are difficult now because as you know, the LLMs produce nice text but they hallucinate at times. Also, most of these use cases are very generative. For example, in our case, we want to generate coaching notes for supervisors so that they can provide better coaching for their agents. The notes would be very fluent, but is it actionable? Does it really have all the important points where the agent can really improve the KPIs they're supposed to do? So how do you validate? There is no metric as such in the literature or public where you can take it compared to previous AI models which are supervised, where you can compare image recognition accuracy or speech recognition accuracy. But these models challenge you on how to measure the success of the models itself. That's one big thing, and every use case has its own nuances. So how do you measure? The other part is how do you take the feedback and improve these models. That's another big challenge for us. Then finally, scale and reliability. At a point, we have to go from one to 100x in a day or two. How do you handle all these systems? With the foundation data layer, but after that, how do you really provide both trust and accuracy at that scale?

Yeah, great question. One challenge we thought initially was okay, we will build our own foundation model for contact centers. We tried, we built, and we could show a big gap between us versus all different models. Then we realized that every time they release a new model, the gap is reducing. So we pivoted to models where we can really build smaller LLMs and right-size them for the right use case and still get advantage of these frontier models for very large use cases. This hybrid approach really worked for us both for scale and cost purposes. That's one big learning for us.

Yeah. I mean, I think our journey has been at two different levels. Level one has been just building the modern agent tech systems. When we started, a lot of people take a lot of the technology that is available like sandboxing, memory, all of these for granted. When we started, nothing existed. We had to pretty much invent everything on the go. We actually wrote our own sandbox technology. We pretty much invented memory disk snapshotting so that agents when they die can actually start from the same state. There was no database versioning at that point, so we had to invent a lot of those things ourselves. We ended up building the first multi-agent harness, the first memory system, first systems that actually scale test-time compute so you can parallelize agents, multiple agents doing the same problems. Being early had both the thrill to invent some of these things but also the pain that we had to build all of this technology ourselves in-house. That ended up being a strong advantage for us because we control every single layer of the stack and allowed us to innovate at every single layer to solve the actual problem. That was layer one. Layer two is actually that we stand for people who can come in, describe their app, and we take them to production. So we have to take care of hosting, security. How do you pipe production logs back to your dev agents so they can actually solve a production issue? One of the most loved features in Emergent right now is a customer who has deployed an app on Emergent gets a call from his customer and he can just go to the agent and type, 'Hey, my customer is facing this issue,' and our agents will go and fix that. I think we're in a very early journey of modernizing all of software development. What is important is having really strong evals, both offline and online versions, and then paying close attention to what parts of things you don't need to solve today because the next model is going to solve for it. Almost every time a new model comes out, we delete everything we have thought so far and reimagine what the world is going to look like. We have rewritten our system three times in 12 months. With every single model launch, we actually start reimagining the world. For example, Fable is a new class of models, so we start thinking about what multi-agent coordination is going to look like in this world. The pace is crazy right now.

Yeah, I feel that largely software engineering will be a solved problem. Most people will be able to build software that they want. People closer to the problems will be able to build software. For example, at Emergent, most of the software that we have built internally is built on Emergent. Our customer service folks are actually building customer support tools. Literally, the person who's answering the query is also prompting, 'Hey, I want this tagging feature in my tool.' So I think the next frontier is automating knowledge work. That's where the world is headed, where all of the context that lives in disparate software will come together and you'll be able to delegate most of work. We are building this for small businesses right now where you can offload a lot of your operations, marketing, sales automatically to agents just using WhatsApp and iMessage. That's the part we're really excited about. When we started the company, we started with the belief that AI exponential is going to continue, and at every single step, AI has been able to stay ahead of our imagination. So we are excited about that continuing.

Um, I think one of the copilot use cases really surprised us: how the customers use some of these knowledge retrievals and how they can really efficiently improve their operations as well as provide customer service. That's really amazing for us. The other use case we experienced recently is the end-to-end automation. Some of our agents are deployed in different enterprises in the US where they can not only take scheduling events, they can even process payments. That was really exciting to see: agents processing payments without any hurdles. That's really amazing to see.

Yep.

U, I think it's very hard. I'm in this space for quite some time, from doing research to building companies. So it's very hard to predict anything now, but it's always exciting to see new models coming. So wait for the new models, as you wait you build the foundations. One thing we learned is you have to understand your domain very well. As I said, how our users work day in, day out in their workflows, and understand your data. Feel for it. We handle a lot of voice calls: what kind of noises are there, what kind of nuances, how people break in their conversations, how emotional they are at the points when they explain a scenario or issue. So understand all the data and understand more the workflows in the contact centers and build that intelligence layer with strong data foundations, because that would be reusable with every model coming up. Models are getting better and better as we have seen, but your foundations will always stay. Also, you have to really build fast, as Mukun was telling. You have to almost rebuild every time your product. So you have to build systems in more like a Lego fashion so that you can really construct fast as innovation demands.

It's a lot of fun, right? I think it's the most fun that most people have had. A lot of people who work at Emergent work long hours, and I think it's just like a new toy. Every six weeks you get a new toy to play with with new capabilities, and people are so motivated to just solve these unsolved problems and reimagine how the world is going to shape up.

I mean, I would say that this is the best time to tinker. Coding has never been easier. You can spin up a lot of agents to build things. To me, it feels like the best moment to really go and chase the dreams that you have always put on hold and really go and start building things that you want to solve for yourself. The landscape, the market is changing really fast, and people who are able to bring solutions that actually have an outcome, as you said, are going to have crazy demand. Consistently, we're seeing that the demand for solving these problems is so high right now that if you just bring the right solution to the world, people will grab it.

Yeah, almost same thing as Mukun said. The best thing is you try it. Even in our company, I've seen one of our TA built his own applications to track all the recruitment and even he sorted the CVs because they know their domain better. So now AI is really helpful for them to work 10x or 100x efficiently. Don't have FOMO for these models. Models will keep coming, but try experiment with it. Another thing is the multimodal part. Models are really multimodal now, so you can really explore between speech to text to image. If you have any great ideas, try it. And as Mukun is telling, if you have good outcomes for these problems, you are the winner in this game.

Thank you.