<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Zeta Decoded on Medium]]></title>
        <description><![CDATA[Stories by Zeta Decoded on Medium]]></description>
        <link>https://medium.com/@zeta-decoded?source=rss-41cca7aaedfd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*LnofCrLtEpSachXgT5tHQg.png</url>
            <title>Stories by Zeta Decoded on Medium</title>
            <link>https://medium.com/@zeta-decoded?source=rss-41cca7aaedfd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 03:15:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zeta-decoded/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Becoming AI-Native: My Internship Building Software at Zeta]]></title>
            <link>https://medium.com/@zeta-decoded/becoming-ai-native-my-internship-building-software-at-zeta-8ee1288af7c9?source=rss-41cca7aaedfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ee1288af7c9</guid>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-development]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai-native-development]]></category>
            <dc:creator><![CDATA[Zeta Decoded]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 19:05:50 GMT</pubDate>
            <atom:updated>2026-09-08T19:05:50.238Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cBmF6oYQgTpQCFoJ0N5XXg.png" /></figure><p>By San Vu</p><p>Before this summer, my understanding of AI in software development was that you would:</p><ol><li>Open the AI chat.</li><li>Type out your problem in the best way possible (and hope it understands).</li><li>Work with whatever answer you get.</li><li>Then, repeat and wait until the day you hit your token limit!</li></ol><p>By the end of my internship as a Developer Experience Engineer within Zeta’s AI Enablement Team, I realized that using AI wasn’t just about asking it to write code but utilizing AI natively within the software development lifecycle. This meant specs before code, agents with context and guardrails, and a newly defined paradigm for engineers.</p><h3>My AI-Native Start</h3><p>The AI Enablement Team at Zeta helps to drive AI-native development adoption across the organization while also building internal developer tools that enable teams to operate at the highest velocity and quality. This meant to succeed in my internship, I had to both deepen my understanding of AI and assess how AI would fit within the steps of the SDLC.</p><ul><li>I started with the fundamentals: What are MCP servers? Models? Agent modes? Skills? Harnesses?</li><li>I learned about developer operations: How do engineering teams document, code, test, and review? How do we track developer productivity?</li><li>Then zoomed out to the larger picture: Where are we now and where do we want to be? What metrics are we solving for?</li></ul><p>Our team moved fast. Priorities shifted, we shipped often, and we were constantly testing and iterating. AI-native was no longer just a chat in the IDE, it was a system that:</p><ul><li>Generated specs and plans from human-curated requirements before large diffs.</li><li>Searched for dev context, repo guides, skills, and standard docs so that agents made more informed decisions by default.</li><li>Used harnesses and guardrails to ensure reliability.</li><li>Verified quality using different models/agents and self-correcting feedback loops.</li></ul><p>I worked across the stack, building an analytics dashboard for our internal AI-native platform as well as a maturity model for how AI-ready a repository is — supporting the transition to spec-driven workflows for cross-functional teams.</p><h3>AI-Native Spec Driven Development</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zxJfmJgXSn6kUHOhktQgew.png" /></figure><p>By week five, I had shipped the entire analytics dashboard feature.</p><p>As an intern, I had expected to start on a small bug fix, do a whole lot of reading, waiting, and maybe finish a dashboard as my stretch goal for the summer. But with an AI-native organization and a strong mentor, I was able to own a much larger slice of the problem and expand to directing product decisions alongside system design.</p><p>When given a big feature like the analytics dashboard, I wrote a v1 proposal on requirements and aligned it with my manager. Then, I worked with my mentor on the implementation plan. We divided the large task into smaller tasks. For each task, I created a Jira ticket which would state the requirements, acceptance criteria, and tests to make sure it worked correctly.</p><p>For each ticket, I had my developer workflow. With AI as the code writer, I had to take a step back and wear the multiple hats of a product manager, UI/UX designer, architect, etc.</p><ol><li><strong>Plan</strong>: One or more agents would help think through the feature before we started coding. Not just what to build, but how it should work for users and fit into the existing system. This included the user flow, UI/UX, technical approach, performance, maintainability, and acceptance criteria.</li><li><strong>Build</strong>: An implementation agent would take that plan and turn it into code, while following the project’s coding patterns and adding the necessary tests and documentation.</li><li><strong>Verify</strong>: Verification agents would review the implementation against the original plan and test that it worked as intended. They helped catch gaps, identify regressions, suggest refactors, and verify the user experience, security, performance, and reliability. If something was off, we repeated the build-and-verify loop until it was ready.</li><li><strong>Deploy</strong>: Finally, we opened an MR with the code, documentation, and verification evidence. After the required AI-assisted and human reviews, we merged and deployed the change, then monitored it to confirm it was working correctly.</li></ol><p>Throughout the process, I would still have to debug the AI agent’s output with my mentor, but this is where mentorship and technical depth mattered most. I learned that the agent could build at a granular level, but it still had a hard time abstracting and anticipating scalability, UX, and edge cases that only deep human experts can envision.</p><h3>Reflecting on my AI-Native Journey</h3><p>AI doesn’t replace the need to learn to code in software engineering. Instead, it shifts-left the SDLC by focusing on system design, product trade-offs, performance, and API design prior to building. Just because AI can ship code faster doesn’t necessarily mean it’s better. AI may be the collaborator at every step of the workflow but at the end of the day, I am responsible for the outcome. It’s my responsibility to continue to work on my technical depth and make better system design decisions as I build.</p><p>Thank you to Zeta Global and the AI Enablement team for the opportunity and for the perfect project to work on. I really valued the mentorship, ownership, and enjoyed working with everyone across Zeta!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ee1288af7c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Zeta’s Lakehouse Journey: A Composable, Scalable, and Federated Architecture]]></title>
            <link>https://medium.com/@zeta-decoded/zetas-lakehouse-journey-a-composable-scalable-and-federated-architecture-df0ab5f19c3a?source=rss-41cca7aaedfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/df0ab5f19c3a</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[lakehouse-architecture]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-lakehouse]]></category>
            <dc:creator><![CDATA[Zeta Decoded]]></dc:creator>
            <pubDate>Tue, 06 Jan 2026 15:53:49 GMT</pubDate>
            <atom:updated>2026-01-06T15:53:49.710Z</atom:updated>
            <content:encoded><![CDATA[<p>By <a href="https://www.linkedin.com/in/ananthdurai/">Ananth Packkildurai</a></p><p>As a <a href="https://zetaglobal.com/resource-center/forrester-ai-cross-channel-marketing-hubs-report/">leader in the AI Marketing Cloud</a>, Zeta powers mission-critical, agentic marketing workloads for some of the world’s most sophisticated enterprises. Today, <a href="https://zetaglobal.com/">44 percent of the Fortune 100</a> rely on Zeta’s platform to activate customer intelligence and deliver autonomous marketing outcomes. At this scale, the quality, readiness, and adaptability of our underlying data assets directly influence the business value we can generate for our clients. Ensuring those assets are AI-ready, clean, governed, semantically consistent, and immediately consumable is fundamental to our mission.</p><p>Our landscape expanded significantly with the acquisitions of <a href="https://zetaglobal.com/news/zeta-global-to-acquire-liveintent-in-a-highly-accretive-transaction/">LiveIntent</a> and <a href="https://zetaglobal.com/news/zeta-global-to-acquire-marigolds-enterprise-business/">Marigold</a>. These additions brought tremendous strategic value but also introduced a much broader set of data sources, schemas, processing patterns, and operational models. As a result, we found ourselves operating in an inherently heterogeneous ecosystem where traditional warehouse-centric or single-format lake approaches would impose significant friction. We needed an architecture capable of absorbing diverse data assets, normalizing them efficiently, and making them usable across analytical, activation, and AI workloads, without compromising scale, cost efficiency, or governance.</p><p>The Lakehouse paradigm emerged as the architectural model best suited to address these needs. It provides a unified data layer that brings transactional guarantees, schema enforcement, and governance to low-cost open object storage. This combination allows us to integrate data from disparate systems, preserve fidelity, maintain consistency, and expose it through a single, high-performance analytical surface. Most importantly, it enables us to build durable, reusable, AI-ready data assets that power both real-time activation and advanced AI workloads across our expanding customer base.</p><h3>Composable Data Platform: Our Core Design Principle</h3><p>To support AI-driven marketing workloads at scale, we designed our foundational data platform as a set of modular, independently replaceable components.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/469/0*VFQzU_AEkDSzWCMu" /></figure><h4>The core layers include:</h4><ul><li><strong>Object storage</strong> is the universal persistence plane. Parquet is the standard columnar file format.</li><li><strong>Lakehouse table formats</strong> provide transactional guarantees, schema evolution, and time-travel semantics.</li><li><strong>Metadata catalogs</strong> governing discoverability, versioning, and governance across the organization.</li><li><strong>Multiple compute engines</strong>, such as Spark, Snowflake, Clickhouse, Trino, and others, are consuming these assets for diverse analytical and AI workloads.</li></ul><p>Decoupling these layers ensures that each component can evolve without imposing downstream disruption. This composable design allows us to bring new engines online, adopt emerging table formats, enhance governance, or introduce specialized optimizations, all without re-architecting the entire platform. It is fundamental to supporting AI workflows that include streaming inference, real-time activation, large-scale batch computation, and exploratory analytics.</p><h3>Why We Chose S3Table?</h3><p>We selected Apache Iceberg as our Lakehouse table format because it offers strong transactional guarantees, schema evolution, time-travel semantics, and an open specification supported by a broad and growing ecosystem. Iceberg’s emphasis on hidden partitioning, metadata scaling, and cross-engine interoperability aligns well with our composable platform strategy, where multiple compute engines: Spark, Flink, Trino, and others, must operate seamlessly over shared datasets. Iceberg provides a robust foundation for long-term compatibility and avoids the lock-in concerns associated with proprietary table formats.</p><p>Once we aligned on Iceberg as the table format, we needed a scalable and operationally efficient mechanism to manage table state, metadata updates, and multi-tenant lifecycle operations. After evaluating multiple catalog and maintenance solutions, we chose S3Table paired with AWS Glue Catalog as the control plane for managing Iceberg tables. Our prior work, <a href="https://aws.amazon.com/blogs/storage/how-zeta-global-scales-multi-tenant-data-ingestion-with-amazon-s3-tables/">published on the AWS Storage Blog</a>, demonstrated how S3Table enables Zeta to scale multi-tenant ingestion pipelines with predictable performance and cost efficiency. In comparative testing, S3Table offered simpler operational semantics, faster metadata access paths, and more predictable behavior under high-concurrency workloads than alternative catalog-backed Iceberg implementations. This combination of Iceberg for table format and S3Table + Glue for management gives us the flexibility of an open Lakehouse with the operational stability required for our AI-driven, multi-tenant platform.</p><h3>Multi-Account Federated Catalog Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yGX0SYcaScZPctIH" /></figure><p><strong>Federated Multi-Account Lakehouse Design</strong></p><p>With the table format and management plane established, the next challenge was enabling secure, consistent access to these Iceberg tables across multiple AWS accounts without duplicating data or centralizing everything into a single environment. Unifying data from multiple business units and historical acquisitions into a single, centralized catalog often introduces significant delays in time-to-market. Each environment brings its own account boundaries, governance models, and operational constraints, and forcing full consolidation before delivering value slows down product integration and AI feature rollout.</p><p>To avoid that bottleneck, we adopted a multi-account federated catalog architecture. This approach allows each account, producer, governance hub, and consumer to maintain autonomy over its own storage, metadata, and lifecycle policies while still enabling unified discovery and governed access to shared datasets. Rather than replicating or migrating datasets across accounts, data remains in the producer account, and consumers gain secure, on-demand access through well-defined catalog associations and permissions. This pattern aligns with our composable platform philosophy and minimizes both duplication and operational friction.</p><h3>Future Work</h3><p>As we continue to evolve our Lakehouse foundation, we are investing in three major areas to further strengthen scalability, governance, and developer productivity:</p><h3>Establishing a Unified Data Product Framework</h3><p>We are formalizing a unified data product development framework that standardizes how datasets are modeled, published, validated, and governed across domains. By defining consistent patterns for data creation and management, the framework promotes shared semantics, reduces unnecessary duplication, and improves interoperability across our AI and analytics ecosystem.</p><h3>Deepening Lineage and Observability with OpenLineage Integration</h3><p>Integrating OpenLineage across ingestion, transformation, and consumption layers will provide end-to-end visibility into data flows. The OpenLineage integration enables proactive governance, faster incident resolution, improved auditability, and richer operational insights for all pipelines.</p><h3>Engineering the Next 10x Leap in Platform Scale</h3><p>As customer adoption accelerates, we are scaling the infrastructure by an order of magnitude and engineering the next wave of performance and capacity improvements across storage, catalog services, and compute orchestration.</p><p>If you’re interested in building the next generation agentic marketing platform, apply on our <a href="https://zetaglobal.com/about/life-at-zeta/#jobs">careers</a> page.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df0ab5f19c3a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>