<?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 Surabhi Gupta on Medium]]></title>
        <description><![CDATA[Stories by Surabhi Gupta on Medium]]></description>
        <link>https://medium.com/@surbs?source=rss-76c152c0322b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*69A4SM8NSEX_X15J.JPG</url>
            <title>Stories by Surabhi Gupta on Medium</title>
            <link>https://medium.com/@surbs?source=rss-76c152c0322b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 13:33:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@surbs/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[The four levels of AI-driven engineering: Our approach at Klaviyo]]></title>
            <link>https://klaviyo.tech/the-four-levels-of-ai-driven-engineering-our-approach-at-klaviyo-12e50b9f04a1?source=rss-76c152c0322b------2</link>
            <guid isPermaLink="false">https://medium.com/p/12e50b9f04a1</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[engineering-leadership]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[developer-productivity]]></category>
            <dc:creator><![CDATA[Surabhi Gupta]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 20:07:51 GMT</pubDate>
            <atom:updated>2026-03-12T20:07:51.751Z</atom:updated>
            <content:encoded><![CDATA[<p>AI adoption in software engineering is no longer a question. Every engineering leader I speak with is roughly operating in the same way: their teams are using AI daily to write code, debug, and explore unfamiliar systems.</p><p>What’s changed more recently is the scale of what these tools can take on. Over the last several months, AI agents have crossed a meaningful threshold. They can handle coordinated, multi-step work and recover from errors without constant human correction. The constraint is no longer the tools themselves. It’s how we structure our workflows, how we provide context, and how we enforce quality that matters. Unlocking the next step requires changing how we work.</p><p>Everyone seems to be figuring this out in real time. Below, I have outlined how we are thinking about this shift at Klaviyo. Not as a finished playbook, but as a framework which we hope is useful as the industry experiments and learns together. I believe the more stories we exchange and build in the open, the better we’ll all be.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VgbBFJaOawtKfYCrQP4enQ.png" /></figure><h3><strong>A framework for AI evolution in engineering</strong></h3><p>One useful way to understand this shift is by looking at how work flows through the system as AI becomes more central. At Klaviyo, we’ve been thinking about this progression in four levels of AI-driven maturity.</p><p><strong>Level 1</strong></p><p>AI accelerates individual tasks like boilerplate, test scaffolding, and exploration. Engineers remain the primary author, and the structure of the work stays the same.</p><p><strong>Level 2</strong></p><p>Engineers delegate meaningful pieces of work to AI, but they remain deeply involved throughout the process. AI proposes changes, and engineers iteratively guide, edit, and correct them. Most of the cognitive load still sits with the engineer. Reviews are detailed and line by line, and work still ships one PR at a time. Each task may be completed faster, but the same amount of human effort and review is required every time.</p><p><strong>Level 3</strong></p><p>Engineers’ roles shift from producing code to designing intent and verifying outcomes. Engineers define the goal, constraints, and context up front. Agents generate the implementation. Engineers remain accountable for correctness at merge time, focusing on security, performance, system behavior, and long-term maintainability.</p><p>At Level 3, autonomy increases, but accountability does not change. Review remains the default control point. This is where system-level throughput meaningfully increases without sacrificing trust.</p><p><strong>Level 4</strong></p><p>At Level 4, the unit of work expands from individual changes or features to larger cross-system builds. Engineers define product and system intent and direct agents to execute work across services and teams, with AI handling coordination and implementation. Engineers remain responsible for goals, trade-offs, and overall system coherence.</p><p>Level 4 represents a real future direction for the industry. We include it here because, while this isn’t our current target, naming where things are heading shapes how we build toward Level 3 today. It requires stronger guarantees around testing, observability, and self-correction than we have today.</p><h3><strong>Our target: Level 3</strong></h3><p>Our focus is clear. By the end of H1 2026, we intend for all our engineering teams at Klaviyo to operate at Level 3 as the expectation, with roughly 90% of our code being AI-generated.</p><p>We are moving quickly on purpose. The sooner our teams begin working in this model, the sooner we can uncover where it doesn’t yet fit our current reality, whether that’s missing context, tooling gaps, or review processes that need to evolve. Working through them early lets us improve the system once rather than stretching the adjustment out over time, and ultimately build an engineering workflow that moves faster as the default.</p><p>Although some of our teams are already operating at level 3, changing how the entire engineering team works is not trivial. To get there, we are supporting our teams across three areas:</p><ol><li><strong>Agent-first foundations (harness engineering): </strong>We are building shared templates, skills, and domain-specific context so teams aren’t rebuilding the same scaffolding repeatedly.</li><li><strong>Limiting SDLC bottlenecks:</strong> We are tightening the integration between planning and coding, improving the speed and reliability of testing, and reducing human waiting time throughout the pipeline.</li><li><strong>Measurement and feedback</strong>: We are building visibility into adoption, throughput, and quality so we can identify what works, spread it quickly, and correct what does not.</li></ol><h3><strong>How we’re asking engineers to work differently</strong></h3><p><strong>Start work with an agent.</strong> For most tasks, express intent to an agent and review the outcome. Repeated manual workflows are good places to begin.</p><p><strong>Encode context once, improve it continuously.</strong> Capture standards, constraints, and workflows in shared agent context. When an agent produces a poor result, update the context so the system improves for everyone. Treat failures as inputs to system improvement.</p><p><strong>Maintain the same quality bar.</strong> AI-generated code must meet the same standards of correctness, clarity, and maintainability as human-written code. If a change can’t be confidently explained or validated, it’s not ready to ship.</p><p><strong>Keep human ownership explicit.</strong> Agents generate code, but humans own outcomes. Every merged change has a clear owner accountable for its behavior in production.</p><p><strong>Ship one small feature, agent-first. </strong>Pick a well-scoped customer problem and take it from intent to merged PR with an agent doing the implementation. Start where you can define “done” clearly before you begin. Any rough edges are signals about what context is missing. Fix those, and the next one goes smoother.</p><h3><strong>What this means for builders</strong></h3><p>What changes most at Level 3 is not the importance of technical depth, but where it’s applied.</p><p>Systems thinking, knowing how to decompose ambiguous problems into well-defined tasks that agents can execute independently, becomes essential. So does product thinking, using a deep understanding of customer workflows to guide technical decisions. And above all, ownership remains fundamental. Regardless of how code is produced, engineers remain accountable for how their systems behave in production.</p><p>The organizations that invest in these skills and move deliberately through these levels, rather than hoping productivity gains happen organically, will build a meaningful and durable advantage. The goal isn’t just to ship faster. It’s to redefine what great engineering looks like.</p><p>If you’re interested in joining us on this next wave of engineering and innovation at Klaviyo, you can see our open roles <a href="https://www.klaviyo.com/careers">here</a>.</p><p><em>A huge thank you to</em><a href="https://www.linkedin.com/in/fnthawar/"><em> Farhan Thawar</em></a><em> for inspiring me to think about this framework.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12e50b9f04a1" width="1" height="1" alt=""><hr><p><a href="https://klaviyo.tech/the-four-levels-of-ai-driven-engineering-our-approach-at-klaviyo-12e50b9f04a1">The four levels of AI-driven engineering: Our approach at Klaviyo</a> was originally published in <a href="https://klaviyo.tech/">Klaviyo Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Klaviyo’s BFCM Scale in 2025]]></title>
            <link>https://klaviyo.tech/klaviyos-bfcm-scale-in-2025-99e4bc402b5c?source=rss-76c152c0322b------2</link>
            <guid isPermaLink="false">https://medium.com/p/99e4bc402b5c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[bfcm-2025]]></category>
            <category><![CDATA[scale]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ecommerce]]></category>
            <dc:creator><![CDATA[Surabhi Gupta]]></dc:creator>
            <pubDate>Tue, 16 Dec 2025 19:09:41 GMT</pubDate>
            <atom:updated>2025-12-16T19:39:42.857Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fWNoBfyGg-rsQqR60x6HiQ.png" /></figure><p>Every Black Friday and Cyber Monday (BFCM), millions of shoppers make decisions driven by marketing messages, real-time data, and personalized experiences powered by Klaviyo. For our customers, BFCM isn’t just a revenue event, it’s the most important four days of the year. For Klaviyo Engineering, it’s our annual reliability championship.</p><p>Thanks to months of preparation, a strong engineering culture, and systems we’ve spent years evolving, this year’s BFCM was smooth, predictable, and record-breaking.</p><p>Here’s a rundown of the weekend and a behind the scenes look at how it all came together.</p><h3>First, The Numbers</h3><ul><li>Our customers drove over <strong>$3.8 billion</strong> in total Klaviyo Attributed Value, a 27% YoY increase</li><li>Nearly <strong>20,000</strong> customers hit their best sales day ever</li><li><strong>22.7B</strong> total messages sent over BFCM 2025, a 25% YoY increase</li><li>Our event pipeline ingested over <strong>900 million</strong> events per hour during our peak hour</li><li>Our flows engine processed over <strong>2.8 billion</strong> steps</li><li>Over <strong>4.2B</strong> inbound webhooks processed</li><li>Our segmentation system handled <strong>125 billion</strong> updates to segment memberships. This is actually <em>lower</em> than our total in 2024, but that’s because we made significant investments in the pipeline to reduce overall load and only update segments when necessary</li></ul><h3>Why BFCM Matters at Klaviyo</h3><h4>Our success is defined by our customers’ success.</h4><p>Every metric we track, from event throughput to email send rates, rolls up to one goal: helping brands grow. That’s why our north star is <strong>Klaviyo Attributed Value (KAV)</strong>, a metric that captures the revenue customers generate directly from Klaviyo.</p><p>In 2024, <a href="https://medium.com/klaviyo-engineering/our-bfcm-2024-scale-eb8716c101c8">we publicly shared </a>that Klaviyo customers drove over $3 billion in KAV. This year, they topped that record with more than <a href="https://www.klaviyo.com/bfcm/holiday-shopping-trends">$3.8B in total Klaviyo-attributed revenue for BFCM 2025</a>, a 27% YoY increase!</p><h4>BFCM is a make-or-break moment for brands</h4><p>During BFCM, our customers send campaigns at their highest volumes, rely on live segmentation and AI to react instantly to shopper behavior, and trust Klaviyo to deliver messages quickly, accurately, and reliably.</p><p>If we fail, they fail. That’s why we treat BFCM as a full-fledged product: planned collaboratively, tested thoroughly, and executed with precision.</p><h3>How We Prepared for BFCM 2025</h3><p>BFCM prep is a team sport, we give guidance and support as a leadership group and have people tasked with the larger perspective, in the end each team owns their success inside of that framework. This year we had a few people who were asked to focus on BFCM prep, so before we continue, it’s worth shouting out John Meichle, Patty de Castro, and Chris Chiodo.</p><p>For 2025, we modified the approach we’ve taken in past years, while keeping many aspects that have historically worked for us. Let’s take a look at how these two approaches work together.</p><h4><strong>Bottom-Up: Every Team Owns Its Scale</strong></h4><p>Our BFCM readiness starts at the ground level. Each team validates not only the capacity of its own services, but also the cross-service data flows that matter most during peak traffic. Things like campaign sending, on-site data syncing, checkout events, segmentation, flows, and more.</p><p>In 2025, we introduced a quarterly BFCM readiness scorecard to bring further consistency, transparency, and accountability to this process. Teams measure and report on both service-level and team-level metrics across a wide range of categories. And while measurement and reporting generally are considered table-stakes for an engineering organization, we view it as critical in reinforcing Klaviyo’s culture of ownership, continuous learning, and data backed process improvements.</p><h4>Top-Down: Central Coordination and Standards</h4><p>To complement the bottom-up preparation, Engineering leadership sets the structure, clarity, and communication needed to keep the entire organization aligned. This direction ensures teams stay focused on the highest-priority work required to win the weekend. It includes a set of rituals and systems such as:</p><ul><li><strong>Guidance in defining scale targets</strong></li><li><strong>Creating, enhancing, and socializing the BFCM readiness scorecard process</strong></li><li><strong>A cross-engineering review process for scorecards and enumerated risks</strong></li><li><strong>A centralized risk registry capturing all BFCM risks</strong></li><li><strong>Tracking readiness across engineering pillars, teams, and services</strong></li><li><strong>A support forum for any BFCM-related preparation, technical or process</strong></li><li><strong>A clear timeline for code and infrastructure chill, freeze, and thaw</strong></li><li><strong>Adoption of a revamped incident-response system</strong></li><li><strong>All hands meetings to celebrate wins and highlight areas of risk</strong></li></ul><p>This structure establishes consistency and removes ambiguity. Teams know exactly what “ready for BFCM” means, and how readiness is measured.</p><h3>Always Be BFCM-ing</h3><p>The most notable changes in 2025 centered on improving the efficiency and precision of our preparation process. We kept the core components that have historically worked well, but refined how we executed them to reduce unnecessary overhead and ensure all work directly supported BFCM-scale readiness.</p><p>One of the largest shifts was in our approach to scale-out testing. In previous years, these tests ran periodically throughout the year and were treated as coordinated, engineering-wide maintenance windows. For 2025, we streamlined the process: we ran fewer — more targeted — tests, aligned them much closer to BFCM, and integrated them directly with team-owned load tests. This ensured that scale-out work validated the things that matter most: connection limits across datastores, system stability at projected BFCM scale, and the accuracy of our capacity assumptions. All without consuming cycles earlier in the year when load patterns were less representative.</p><p>We also significantly expanded the use of combined load tests across engineering. With Klaviyo composed of many independently deployed applications, each generating distinct downstream workload characteristics, combined load testing provided a realistic picture of platform-wide behavior. In 2025, we coordinated the largest set of teams to date to generate production load simultaneously. This indirect load surfaced system interactions and bottlenecks that would not appear in isolated tests. By pairing these efforts with scale-out tests, we ensured infrastructure was already sized to BFCM targets, allowing teams to focus on application-level and architectural constraints rather than basic capacity validation.</p><p>This year also introduced <a href="http://cortex.io/">Cortex.io</a> as the foundation of an internal service graph. We began using Cortex to centralize service metadata, readiness scorecards, ownership, and dependency mapping. This complimented the manual, ad-hoc mechanisms we previously relied on and gave teams a single source of truth for operational context. BFCM served as the first major test of this system and demonstrated its value in accelerating decision-making and cross-team coordination.</p><p>Finally, we formalized a model that combines bottom-up system ownership with structured top-down program coordination. Engineering pillars became directly accountable for their own readiness, with senior technical leaders driving verification within their domains and escalating cross-pillar needs to the core BFCM team. Given the scale and complexity of Klaviyo’s architecture, now spanning thousands of services, this distributed ownership model has become essential for maintaining depth of expertise while still providing unified program visibility.</p><h3>The Payoff: Reliability in Action</h3><h4>Fewer Incidents, Faster Response, Smaller Blast Radius</h4><p>Preparation didn’t replace vigilance — it empowered it. We were confident in the alerting and dashboards we built and we watched them intently throughout BFCM. What we saw was the payoff of many months of work across Klaviyo.</p><p>Under our revamped incident process, which is set up to catch and escalate potential issues before they impact our systems or customers, we had 32 internal incidents that had no customer impact. Teams were proactively catching issues <em>over an hour before</em> they would have triggered paging or caused customer impact. This is exactly the behavior our refined incident framework was designed to encourage.</p><p>We did have one external incident declared on our status page during BFCM: 18 minutes related to services which didn’t impact our Marketing Product, it was fully resolved with no lasting impact. A clear measure of the successful planning and preparation of the entire organization!</p><h3>What Engineering Actually Does During BFCM</h3><p>The culture of BFCM at Klaviyo is one of the most unique parts of the weekend.</p><p>As graphs spike and throughput curves rise, engineers are monitoring dashboards, watching queues and traffic patterns, and validating autoscalers. We’re also grabbing treats from the snack cart stocked with customer products, watching our annual turkey-themed internal dashboard, and celebrating the day’s numbers and wins together. Side note — we run an annual contest at Klaviyo to design the turkey for our turkey themed dashboard, and this year’s winning bird was Alan Turking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*I-EWnHLIg6gafmDun__o_Q.png" /><figcaption>Alan Turking — our BFCM mascot for 2025</figcaption></figure><p>This year we added a dedicated Slack channel for real-time “engineering highlight reels”: screenshots, metrics, and dashboards from teams proud of how their systems were performing under unprecedented scale. It became a running celebration of both technical achievement and teamwork. BFCM is serious engineering work, but it’s also a moment when the entire organization comes together to take pride in the foundations they’ve built.</p><h4><strong>Looking Ahead to BFCM 2026</strong></h4><p>As we celebrate this year’s performance, our focus has already shifted to 2026. We’re continuing to invest in our AI-powered platform, expand on the features we offer, and strengthen the culture that makes BFCM possible.</p><p>The bar rises every year and that’s exactly what makes Klaviyo Engineering such an exciting place to build.</p><p>Interested in joining us? <a href="https://www.klaviyo.com/careers/search-jobs?team=Engineering">Take a look at our open roles here</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99e4bc402b5c" width="1" height="1" alt=""><hr><p><a href="https://klaviyo.tech/klaviyos-bfcm-scale-in-2025-99e4bc402b5c">Klaviyo’s BFCM Scale in 2025</a> was originally published in <a href="https://klaviyo.tech/">Klaviyo Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Our BFCM 2024 Scale]]></title>
            <link>https://klaviyo.tech/our-bfcm-2024-scale-eb8716c101c8?source=rss-76c152c0322b------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb8716c101c8</guid>
            <category><![CDATA[high-scalability]]></category>
            <dc:creator><![CDATA[Surabhi Gupta]]></dc:creator>
            <pubDate>Mon, 16 Dec 2024 20:15:04 GMT</pubDate>
            <atom:updated>2024-12-16T20:15:04.420Z</atom:updated>
            <content:encoded><![CDATA[<h4>20 billion API requests, 37 billion events…</h4><p>You don’t get to be big and successful without learning from failure along the way. I joined Klaviyo as CTO in September of this year. Just shy of eight years earlier, Klaviyo went down during the single most important period to its customers — Black Friday to Cyber Monday. For a period of hours, emails from our customers to their customers failed to send. In the midst of everyone scrambling to diagnose and fix, I’m told, a customer showed up at our then nearly-impossible-to-find office to tell us how upset he was. That’s how critical our software is to our customers.</p><p>That pain became a rallying point for the company. We would put in place the technical and human systems to be prepared for every future BFCM. We would test and test our systems ahead of time to ensure they could scale well beyond anticipated load. We invested in observability. We made sure there were backup systems for everything including all our internal tools. We created incident command structures and communication plans and started traditions like our <a href="https://klaviyo.tech/turkey-dash-1293a58b4bd2">turkey dashboard</a>.</p><p>Every year the numbers get bigger. We have systems that handle more volume in minutes than they handled during the entire BFCM period in 2016. I’m proud of our team for our <a href="https://status.klaviyo.com/">100% uptime</a> during BFCM 2024. A few numbers from Thanksgiving to Cyber Monday that touch on the scale we operate at here:</p><ul><li>Our customers drove over<strong> $3 billion</strong> in total Klaviyo Attributed Value.</li><li>Over <strong>15,000</strong> of our customers experienced their best sales day ever.</li><li>Our messaging systems sent over<strong> 420 million </strong>emails, texts, and push notifications during our <strong>peak hour</strong>.</li><li>Our <a href="https://help.klaviyo.com/hc/en-us/articles/20920631501979#h_01HGV07KDFKBNJH8KJ545KRFES">event</a> pipeline processed over <strong>37 billion</strong> events.</li><li>Our <a href="https://help.klaviyo.com/hc/en-us/articles/115005237908">segmentation system</a> handled over <strong>250</strong> <strong>billion</strong> updates to segment memberships.</li><li>Our APIs handled over <strong>20 billion </strong>requests.</li><li>Our <a href="https://help.klaviyo.com/hc/en-us/articles/115002774932">flows engine</a> processed over <strong>2 billion</strong> steps.</li><li>Our analytics database ingested over <strong>half a petabyte</strong> of data.</li></ul><p>Thank you team and congrats to our customers!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/881/1*lXOLOQb1ehOYO8-wzk7Zkg.png" /><figcaption>AB introducing me to the company at our weekly “By The Numbers” meeting on October 10</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb8716c101c8" width="1" height="1" alt=""><hr><p><a href="https://klaviyo.tech/our-bfcm-2024-scale-eb8716c101c8">Our BFCM 2024 Scale</a> was originally published in <a href="https://klaviyo.tech/">Klaviyo Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>