- Cerebral Valley
- Posts
- Building context-rich AI applications with Tecton ⚡️
Building context-rich AI applications with Tecton ⚡️
Plus: Co-Founder/CEO Mike on expanding into generative AI and agentic systems...
CV Deep Dive
Today, we’re talking with Mike Del Balso, Co-Founder and CEO of Tecton.
Tecton is an AI data platform designed to help companies get machine learning models into production faster and more efficiently. Drawing from Mike’s experience at Google and Uber—where he built Uber’s Michelangelo platform—Tecton focuses on solving the data challenges that hold back AI applications. The platform bridges the gap between data and models, providing companies with the infrastructure they need to productionize data, build accurate models, and integrate AI into their products seamlessly.
Today, Tecton is used by advanced tech companies like Coinbase, Block, and Atlassian to power real-time AI systems for use cases like fraud detection, personalization, and dynamic pricing. With a focus on making AI infrastructure accessible to both cutting-edge tech teams and larger enterprises, Tecton is positioning itself as a critical player in helping companies maximize the potential of their data.
In this conversation, Mike shares the story behind Tecton, the technical challenges they’ve tackled, and what’s next as they expand into generative AI and agentic systems.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with Mike 💬
Mike - welcome to Cerebral Valley! First off, give us a bit about your background and what led you to found Tecton?
I’m Mike, co-founder and CEO of Tecton. I’ve been in the machine learning space for over a decade now, so I guess I’m kind of a veteran at this point. My career started at Google, where I led the teams behind the AI that powers the ads auction. Every time you type a search into Google, there’s some pretty sophisticated AI running behind the scenes to figure out which ads to show you. This system operates at an enormous scale, in real time, with incredible reliability—Google’s basically never down. It drives hundreds of billions in revenue, so think of it as the most production-ready machine learning infrastructure, even before "mlops" became a thing. That’s what my team was working on.
It was an amazing learning experience. I wasn’t the main guy, but I was surrounded by super geniuses on that team. That’s actually the team that published the famous MLops paper, Machine Learning: The High-Interest Credit Card of Technical Debt. After that, I joined Uber, and the situation there was unique. Uber had all this data, and in business, data is usually seen as a competitive advantage—it’s your proprietary knowledge about the world. Uber was making progress in collecting all this data, but the real question was, what to do with it.
The first step is always visualization. You put it on a dashboard and see what insights you can gather. But the next step is taking action based on that data—not just looking at it on a PowerPoint slide. So, the goal became: let’s use machine learning to do cool things with this data, like route cars differently, detect fraud, or improve recommendations in Uber Eats to drive more revenue.
To do that, we had to democratize machine learning—make it accessible across all teams, from data scientists to engineers. We built Michelangelo, which became the central ML infrastructure at Uber. That platform was key to unleashing a wave of machine learning at Uber. It gave everyone a fast path to production, helped them build accurate models, and did it efficiently, without causing executives to stress over costs.
Within two and a half years, we went from just a handful of models to tens of thousands of models in production. That’s the dream scenario for any executive at a big company—you want AI deployed everywhere. And we made that happen.
After we published a blog post about Michelangelo at Uber, we had a lot of people reach out to us saying they had the same challenges and wanted help. It became really clear that there was a huge demand for the kind of transformation we’d gone through at Uber. Companies have data, they know there’s something valuable in it, but they struggle to actually build that into their products and improve the customer experience. It’s a common problem that a lot of businesses face.
So we reflected on what we got right at Uber. The key insight was that an AI application, the kind that impacts a product and users, isn’t just about the model. Back then, everyone was focused on model training and serving, but those quickly became commodities. The real bottleneck was the data—whether it was accessing it, transforming it, or integrating it into production. Almost all the challenges we faced revolved around data.
At Uber, we centralized and automated that entire workflow—from ideation and prototyping to production and monitoring—through our AI platform, Michelangelo. That’s where the idea for Tecton came from. Initially, we called it a feature store, then expanded to a feature platform because it solved a broader set of data problems. The reality is, companies like Google or Facebook have hundreds of top engineers building these systems, but most other companies can’t do that. So we set out to build that platform for everyone else—bridging the gap between data and models, getting AI systems into production faster, and doing it in a way that’s both accurate and cost-effective. Does that make sense?
Give us a top level overview of Tecton - how would you describe the startup to those who are maybe less familiar with you?
The idea behind Tecton is really about eliminating data bottlenecks for teams working on AI. We aim to help them get into production as quickly as possible while building accurate models, and do it in a way that’s cost-efficient so they don’t end up in trouble for overspending—which surprisingly happens a lot.
What Tecton provides, at its core, is a workflow for everyone involved in an AI application. That includes data scientists, ML engineers, production engineers, and even those focused on data governance and compliance. It’s a platform that allows teams to transform, manage and retrieve data as context for models, while handling the governance, sharing of data, and centralization. Context in an LLM application can include user information, documents returned from a vector database in a RAG system, or even conversation history, all fed into the prompt. For predictive machine learning (ML), context is represented as engineered features and embeddings. But beyond just the workflow, we also efficiently manage the infrastructure to support that process.
Companies like Coinbase and Block are running their real-time AI on Tecton. So when you send a payment, for example, the data that powers the models making decisions about fraud or approval is sent to the Tecton infrastructure. We calculate all the relevant signals and context for that decision and deliver it to the model. It all happens in under 100 milliseconds, at scale, for hundreds of thousands of requests per second. And we do it in the most efficient way across the market. We support both predictive ML use cases like recommendation systems and fraud detection.
We also support real-time pricing, dynamic pricing, credit decisioning, and generative AI applications. And in generative AI, the challenges are pretty similar to machine learning—just with a few variations. Jeff Bezos has a quote I like, where he says something like, "We’re stubborn on vision, but flexible on details." That’s how we approach it. We’re all about connecting data to the model and making that seamless, but we’re flexible on how we get there.
In generative AI, for example, you’re not just working with structured data—you’ve got unstructured data, too. And when you’re transforming that data, it's not just about handcrafted features. You might want to generate embeddings or allow signals to be defined without writing code. Think about product descriptions: you might ask an LLM if something violates a policy. That’s hard to code in Python but easy for an LLM to evaluate and generate a signal for.
At the retrieval layer, the challenge is similar, but with a twist. A predictive model always asks for the same data—for example, the same ten signals for fraud detection. But with an LLM, it might need different data each time, depending on the question. So, we’ve built a retrieval system that lets agentic systems decide which data they need. We have a tools interface that acts as a resource—essentially, the model can request data like user behavior or product information, and we deliver it. There's also functionality for similarity search and more.
The big idea is managing data in one platform for both predictive ML and generative AI, and people even use it for things like rules engines. This way, customers have one system to control all the data across their AI applications and one integration service for all their data sources.
How do you measure the impact that Tecton is having on your key customers, such as Coinbase or Block? Any customer stories that you’d like to share?
We have a few different types of customers. One group is advanced tech companies like Coinbase and Block, who need high-quality decision-making for things like catching fraud. For these companies, missing fraud can cause serious problems, so having real-time AI systems that can bring in data faster and catch more fraud is crucial. Tecton helps them implement signals and use data they couldn’t integrate into their AI systems before, all at a speed they weren’t able to achieve before.
Another group of customers is Fortune 100 companies, like banks and large insurance companies. These companies have a lot of engineers but are less sophisticated in AI. They just want to get AI into production without spending years on a single model. Before Tecton, it took them months just to get one feature into production. Now, it's much easier, and they’ve gone from two models to 65-70 models in a year.
Lastly, we have customers like Atlassian, who use Tecton to improve user experience. Atlassian makes B2B SaaS products, and they use Tecton to inject personalization and smart recommendations across their products, like Jira and Confluence. So when you're using these products and you see auto-completes or tagging recommendations, that’s powered by Tecton’s intelligent data processing.
There’s been an explosion of interest in agentic workflows and multi-modal AI. How has that shaped the way you’re thinking about building Tecton?
I think it's going to be huge. We're actively working with partners on this right now, and if you're trying to figure out agents, it might be worth reaching out to Tecton. The key thing to consider is how important the problem of context is for agentic systems. Agents make multiple iterative decisions, not just one, and they need to do that with a deep understanding of your business and your customer.
There are different types of knowledge agents need: factual (e.g., what company is this, what's the price), procedural (e.g., how to process a refund), and personal (e.g., who is this customer, what did they do last week). Without that context, it's impossible for these systems to be useful. What we're building at Tecton is a way for companies to manage the context and knowledge that agents have access to during workflows, with proper governance and compliance.
This is what separates agents from being the "dumb" chatbots of the past and moves them towards the dream of truly intelligent, personalized AI assistants. It’s all about delivering the right context, and that’s what we focus on at Tecton—assembling and delivering that context to make agentic systems smart and effective.
How do you plan on Tecton progressing over the next 6-12 months? Anything specific on your roadmap that new or existing customers should be excited for?
There are two main priorities for us right now. First, strengthening our foundation—our platform is amazing, but there’s still so much more we can do to enable additional capabilities for the market. Second, diving deeper into agents and LLMs. No one has really figured it out yet, and the best practices just aren’t there. We're all about productionizing AI, but when you consider that only about 5% of AI use cases are in production, there's so much more to learn as more of these use cases roll out.
The space is moving quickly, so it’s important for us to stay ahead and work closely with customers. We just had a big launch that significantly expands our platform for ML teams, especially around GenAI capabilities to support LLMs and agents. We're also expanding beyond just providing tech and software—we’re starting to work with companies on full AI success programs. A lot of them come to us not knowing what to do, so we offer a more hands-on experience, evaluating their situation and guiding them through the right architecture and design. That's been crucial for many in navigating this fast-moving market.
Lastly, tell us a little bit about the team and culture at Tecton. How big is the company now, and what do you look for in prospective team members that are joining?
We’re hiring across the board—looking for top engineers who care about both product and users, and who are smart, hardworking, and humble. We're also seeking sales engineers and solution architects. Our team is technically impressive, and working here is a unique career opportunity. It's not just about working at an AI company, but actually helping some of the best AI teams in the world build their AI.
At Tecton, you get deep exposure to AI infrastructure across various industries and companies, gaining a broad understanding of what it looks like to build AI in different environments. It’s the kind of experience you can’t get anywhere else. On top of that, we really value a strong culture—it’s a team of great people who genuinely care about each other and the work we’re doing.
Conclusion
To stay up to date on the latest with Tecton, learn more about them here.
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.