<?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 Stuart Caborn on Medium]]></title>
        <description><![CDATA[Stories by Stuart Caborn on Medium]]></description>
        <link>https://medium.com/@stuartcaborn?source=rss-e74b72f5cae3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*SxOqOAeRyd6_FFcF.</url>
            <title>Stories by Stuart Caborn on Medium</title>
            <link>https://medium.com/@stuartcaborn?source=rss-e74b72f5cae3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:48:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stuartcaborn/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[Integrating AI development into our playbooks]]></title>
            <link>https://tech.loveholidays.com/integrating-ai-development-into-our-playbooks-d882832faf1e?source=rss-e74b72f5cae3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d882832faf1e</guid>
            <category><![CDATA[extreme-programming]]></category>
            <category><![CDATA[claude-skills]]></category>
            <category><![CDATA[dora-metrics]]></category>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[lean]]></category>
            <dc:creator><![CDATA[Stuart Caborn]]></dc:creator>
            <pubDate>Fri, 06 Mar 2026 11:39:07 GMT</pubDate>
            <atom:updated>2026-03-06T17:40:27.913Z</atom:updated>
            <content:encoded><![CDATA[<h3>Integrating AI development into our playbooks (loveholidays AI diary #6)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hyGaXkHlU9Wdv-1gt4Lz-g.png" /></figure><p>A year ago, we told ourselves we didn’t have a playbook for the agentic world. We were wrong. Code committed with AI assistants now accounts for nearly 70% of our output, and in the same period, our code health has improved every quarter.</p><p>In 2025 we published how we were <a href="https://tech.loveholidays.com/how-loveholidays-is-solving-the-technical-excellence-paradox-933987956758">solving the technical excellence paradox</a> at loveholidays. The Boost team settled into a rhythm, working with our teams to boost technical excellence and we were comfortable with our tried and trusted playbooks. In the first half of the year we made good progress. The health of our code hotspots was rising and a nascent <a href="https://tech.loveholidays.com/from-firefighting-to-foresight-a-new-culture-of-software-reliability-affcf565f9b3">zero surprises culture</a> was tackling our Sentry alerts, steadily driving them down.</p><p>Then AI exploded. We had to regroup and work out what this meant for loveholidays.</p><p>We’ve spent the past year experimenting, measuring, and learning. Code committed with AI assistants increased steadily, passing 50% by the end of 2025, and we’re now approaching 70%. In the same period, we’ve managed to consistently improve code health quarter on quarter, using both quantitative and qualitative measures. The DORA <a href="https://dora.dev/research/2025/dora-report/">2025 State of AI-assisted Software Development report</a> confirmed what we were seeing: AI amplifies what’s already there. Strong practices get multiplied. Poor practices get multiplied as well.</p><p>So in June 2025 we told ourselves that we didn’t have a playbook for the agentic world. Nine months on, we’re revisiting our old favourites. XP, Theory of Constraints, Lean. We found we didn’t need to rip them up and start afresh. The values hold, even as the practices evolve. Making quality visible, identifying bottlenecks, and reducing waste in the value stream, these aren’t new ideas.</p><p>Let’s look at what it means for a codebase, team, or function to be “AI-Ready.” Rather than inventing a new framework, we are simply applying our enduring principles to three adapted areas of practice: guardrails, process, and knowledge.</p><h3>AI-Ready Guardrails</h3><p>The XP principles of making quality visible remain fundamentally unchanged; what’s adapted is who, or what, needs to see it.</p><p>In January 2025, we’d initially focused on raising the bar for code quality, to benefit the humans. Getting there had meant making quality visible. Dashboards, Slack alerts, regular reviews around code health. Teams set their own targets. The Boost team provides visibility of progress and guidance when things slip. When the cost of repair is too high, our engineers rebuild.</p><p>Agentic coding presents a different challenge. <a href="https://arxiv.org/pdf/2601.02200">CodeScene’s research</a> also shows that code health is important for LLM success, and that the return is non-linear, much as it is for humans. Agentic coders are more successful at modifying and refactoring higher quality code and also complete with a lower token spend. For AI to succeed reliably, code health needs to be above 9.75 and this is what we are now beginning to define as AI-Ready code health.</p><p>Around July 2025, we started embedding CodeScene tooling instructions in our CLAUDE.md files to drive agentic behaviour, with code health as a guardrail. We dogfooded on our own team’s repo and it hit 10.0 health within days of adopting the pattern. We also began to observe this in other teams following this approach. Overall hotspot code health at loveholidays as a whole went from 6.0 to over 9.75 in twelve months, as did our average code health along with the qualitative health feedback from our team sentiment checks. As you can see from the chart below, we stumbled a little as we learned how to apply quality guardrails, but as more engineers adopt them quality continues to improve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ky-_Ag_MOjbES64i-ZTR6w.png" /></figure><p>Today over 80% of our teams have what for humans would be considered healthy code above 9.0. A third of them are what we consider AI-Ready with hotspot health above 9.75.</p><p>Code health, however, is just one signal that can help our agents be more productive; a growth area for us in 2026 is systematically encoding our human development guardrails into our agentic development: test coverage, linting and TDD alongside code health.</p><p>Code quality guard rails are the robot equivalent of the refactor signal in TDD. The quality gate provides feedback to the agent, as a pair might do, that it should tidy up the code.</p><h3>AI-Ready Processes</h3><p>The Theory of Constraints posits that it is the bottlenecks in a system which control overall throughput; a system is only as fast as its slowest part. This principle hasn’t changed, but AI has the potential to change where the bottlenecks might lie.</p><p>The idea of a 10x programmer is exciting, but we need to remember that cutting code is only one part of the development process. Agentic output is only as valuable as the delivery pipeline’s ability to keep pace with it. When the build pipeline is the bottleneck the 10x speed gain to code is lost as it meets the limit of CI/CD pipe throughput. Pipelines quickly become inefficient when you fold in queues and wait times for code review and builds. Then there is rework waste. If your agent is submitting poor quality PRs into the pipeline, you will loop inefficiently through a waste cycle of patching problems found in CI.</p><p>To reduce overprocessing waste, we have aggressively <a href="https://tech.loveholidays.com/build-baby-build-276e445f539e">optimized our build times</a>. We ran 4.4x more builds than the previous year while making them 3x faster. Reducing this processing overhead ensures that when an agent triggers a build, the system provides a validation signal with minimal ‘toil’ waste, keeping the focus on value-added coding rather than pipeline overhead.</p><p>Our deployment data suggests we’re currently keeping pace with respect to avoiding rework waste. We have had an elite DORA level of deployment frequency for many years and this continues to hold steady as we scale up agentic assistance. We also saw the P90 of our lead time to deploy a change reduce from 138 minutes to 44. In addition to this, the build success rate gap between AI-assisted and human commits has closed recently, and is passing human quality for the teams which have AI ready code.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cp4krNX4n0o-bmgJLdb88g.png" /></figure><p>As we drive down build times and improve commit accuracy, we expect our bottleneck to move to PR review handoffs. Better data on process and cycle time is a focus for 2026.</p><h3>AI-Ready Knowledge</h3><p>We need to revisit Lean’s principle of ‘right information, right time’ to prevent defects or overprocessing. For humans, this might have been process documentation and tribal memory; for agents, missing context defaults to guessing, errors and rework.</p><p>Guardrails and processes provide speed and safety. But without organisational context, your AI assistant is guessing. And guesswork leads to rework or overprocessing. If your agents have an accurate understanding of your business and processes, then they will make more focussed and accurate decisions most of the time.</p><p>We’re now experimenting with a couple of complementary approaches to help with this.</p><p>Democratise our data: we’ve <a href="https://tech.loveholidays.com/loveholidays-ai-diary-1-27bb7d8cea14">built an MCP server</a> that gives agents access to our <a href="https://tech.loveholidays.com/a-data-mesh-for-the-rest-of-us-12e2c10ac128">data products</a> with business context attached. Team mappings, use cases, domain knowledge, whatever the agents need to know. This reduces overprocessing — agents can pull the data that they need on demand, and rework where they guess based on training rather than act on hard local data.</p><p>Usage of our internal insights tooling is growing within loveholidays, and now serves more than three times the questions in Q1 2026 as it did in Q3 2025.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yzRr4-0gwiQZ833Vufgxgw.png" /></figure><p>Compose specialised <a href="https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview">agent skills</a>: we are taking our first steps to operationalise knowledge that would otherwise live in docs or tribal memory. We have early examples of skills in the wild which are starting to teach agents how to run pipelines, onboard systems, and expand our data products. This is less processing for the agent and the humans to build context sufficient to solve their problems.</p><p>Skills and our data products need not be bounded by the technical domain of the organisation. An example at loveholidays is our approach to compliance and auditing; a process which in many organisations is opaque and adds toil with negligible value. We are encoding our policy goals and reasoning into skills and data products. The goal is to streamline these processes and extract value rather than hinder teams. Encoding this knowledge into an AI ready format has potential to lower cognitive load and toil on our teams and increase consistency across the organisation. We are in the early stages of measuring impact of encoding process, expect more on this later in the year.</p><h3>Conclusion</h3><p>This validates our engineering culture. The values in our playbooks hold, even as our practices evolve.</p><p>If you are looking for a place to start, get your guardrails in place. We’ve found that high-quality feedback is a catalyst for agentic productivity. Once you have that, look at your processes: if your AI is bottlenecked on your pipeline, you are bleeding out your gains. The playbooks still work; we just have to be disciplined enough to identify our constraints and adapt our practices.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d882832faf1e" width="1" height="1" alt=""><hr><p><a href="https://tech.loveholidays.com/integrating-ai-development-into-our-playbooks-d882832faf1e">Integrating AI development into our playbooks</a> was originally published in <a href="https://tech.loveholidays.com/">loveholidays tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Escaping our joined-at-the-database monolith using event sourcing]]></title>
            <link>https://tech.loveholidays.com/escaping-our-joined-at-the-database-monolith-using-event-sourcing-eb46c3397768?source=rss-e74b72f5cae3------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb46c3397768</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[conways-law]]></category>
            <category><![CDATA[event-sourcing]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[data-mesh]]></category>
            <dc:creator><![CDATA[Stuart Caborn]]></dc:creator>
            <pubDate>Fri, 27 Sep 2024 09:23:21 GMT</pubDate>
            <atom:updated>2024-09-27T09:23:21.355Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Written by </em><a href="https://medium.com/@stuartcaborn">Stuart Caborn</a><em> and </em><a href="https://medium.com/@luke_redroot">Luke Williams</a></p><p>It is common in the software industry to build monsters.</p><p>Business success takes the lead and pragmatic design choices race ahead of the drive for simplicity. Accidental complexity creeps in steadily. Code, which began life small and well formed, grew fast as the company became more successful, taking on responsibility for more and more critical business activities. So what do you do when the time arrives to evolve that system into something better?</p><p>We faced this challenge within some of our checkout systems and have been on a journey to migrate towards a better architecture for a few years now. As the company has grown, we increasingly find the need to decouple our systems. This is so that multiple teams can deliver value rapidly with minimal interdependency. When systems are shared between teams it introduces dependency and coupling and the same goes for shared databases. Where once we had a single team and a single code base we now need a much cleaner separation of concerns.</p><h4>Naming our monster</h4><p>The book <a href="https://teamtopologies.com/book">Team Topologies</a> defines the concept of a joined-at-the-database monolith as “different applications and/or services linked to the same database schema, making it difficult to modify, test, and deploy them independently”.</p><p>Deep within our checkout process lives an old complex service with a runtime database. While this structure served us well initially, as loveholidays has grown and moved to smaller independent teams, Conway’s Law inevitably has had a vote. New systems have been created and sub systems have been extracted over time. One of the problems we have is that some of these additions to our architecture maintain a link to the original database schema in our data warehouse. In some cases systems have integrated with the runtime database itself.</p><p>We had built a “joined-at-the-database monolith”, and this system was now hurting us in a number of ways:</p><ul><li>The database structure was old and poorly understood, which increased the complexity of analysing purchasing activity on the site.</li><li>The Checkout delivery team could not make significant changes to their system without risking damage or unexpected side effects to other delivery teams or offline reporting systems.</li><li>The system complexity had increased the chance of new changes causing rework, bugs and incidents.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xeR4q4eiadHd9Bjq" /><figcaption>External dependencies on our schema not only in the runtime db but also in the warehouse</figcaption></figure><p>The replication of our schema into the warehouse is a twist on the joined-at-the-database monolith anti-pattern. Not only were we joined in the live schema, it had replicated itself into the analytical world where it was virally linked to many reports downstream creating a ‘joined-at-the-data-warehouse monolith’! It was crucial to cut this link to allow the runtime system to be improved.</p><p>We have work in place to sever the classical external system dependencies but this post will focus on our approach to remove the coupling from our analytical data.</p><p>Data Products are self-contained units of data that serve a specific business purpose and are owned by the team that is responsible for the systems that represent that section of the domain, living within a <a href="https://tech.loveholidays.com/a-data-mesh-for-the-rest-of-us-12e2c10ac128">data mesh infrastructure </a>supported by our Scaling teams. They are like building blocks for a data warehouse, each with a defined interface and functionality. In the context of our checkout system, we want to build a basket data product that contains everything the business may need to know about our checkout activity.</p><p>In line with our engineering principles we would build a checkout data product learning from our previous <a href="https://tech.loveholidays.com/event-sourcing-at-loveholidays-fcec7557abd0">event sourcing experiences</a>. We identified three goals:</p><ul><li>Decouple the analytical data source from the runtime system structure.</li><li>Deliver a new data product for analysis while preserving our existing data flows as is.</li><li>Provide a runtime visualisation of the checkout flow to help engineers and non-engineers to inspect user behaviour and outcomes of automated decisions.</li></ul><h4>Thinking big, delivering incrementally</h4><p>Although our ambition is ultimately to replace the incumbent systems entirely, we focused on the three goals above first in order to achieve value quickly.</p><p>Up first, we focused on severing the link between the application database and all offline, analytical consumers. This allowed us to free up the Checkout team to improve the system without impacting other teams.</p><p>The starting point was to mine the system for useful events within the system we aim to retire. Whenever an interesting activity takes place, such as a user selecting a different hotel room, we publish an event. We built an event processor to consume and store these events. The event processing also outputs a projection from all the events seen for a given basket streaming to BigQuery via <a href="https://cloud.google.com/bigquery/docs/change-data-capture">change data capture</a> into our data mesh. To maintain the compatibility with the old data warehouse, the new data product is reduced back down to recreate our legacy schema. We use <a href="https://www.getdbt.com/blog/what-exactly-is-dbt">DBT</a> to manage this process. By building our legacy schema from the new data product we achieve the separation from the runtime database, and also provide a migration path to our new schema as the DBT script provides a working example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qjQ4WPMAPeGh1v_o" /><figcaption>Breaking the link between the runtime schema and it’s mirror in the warehouse</figcaption></figure><h4>Iteratively building the checkout data product</h4><p>The first incremental delivery was a subset of the basket data product which could replicate our legacy schema. By delivering this we immediately had decoupled our teams from the runtime database itself, even if they still were bound to the concepts of this schema.</p><p>By building out enough of a data product to mirror what we had already, we learned a lot about our incumbent system, and also began to iterate and learn what the data product really needed to look like. During this build we discovered and fixed a number of issues in the data, some of which were bugs, others were inconsistent coverage across the package holiday components which we sell.</p><p>We learned that we could expand our view of the basket to encompass more sophisticated events that provide deeper insight into the purchasing process and also to provide the teams supporting the selling operations the information that they need to better understand and support their systems. As we iterated, we expanded the detail of our events and added new events to capture previously lost information such as user selection choices. Other teams who were trialling our prototype product also fed back their ideas around improvements to the shape or gaps in the information which we could fill to make the data even more useful.</p><p>And so we built baskets v2! This is a cleaner and more meaningful view of the activity which we see in the checkout and our analysts are now beginning to integrate directly with this and start the journey to move away from the old setup. Through event sourcing, we have been able to improve the visibility of the behaviour of important checkout business processes. Our data product also provides a deeper understanding for engineering and commercial teams, such as our revenue management team.</p><h4>Visualising the Checkout Flow</h4><p>We like observability here at loveholidays. In fact, we love it. From day 1 we built a basic web UI to allow our developers to easily view the events for a given basket and the projection at runtime which we streamed to BigQuery for the data product. We knew this would be useful for the engineers developing the new data product, but also for all engineers and non-engineers who work closely with the checkout flow. The ability to quickly inspect data and algorithmic decision outcomes around a checkout flow, beyond what was just visible in the consumer-facing UI is valuable. It means that feedback cycles around issues or optimisations are greatly reduced when compared to the system before which essentially required you to access a read-only database replica.</p><p>While the view of the events and the projection were useful, we wanted to accelerate this feedback cycle, and decided to visualise the history of a basket with a simple timeline UI too. The timeline shows each stage of the checkout, including user selections, and we can easily visualise how the basket and the associated component cost and margin varied over time. This is built by transforming a subset of the projection we had previously built.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lo7KJf8ue97n_iS6" /><figcaption>Visualising the basket through time</figcaption></figure><p>This was immediately useful to engineers and non-engineers alike, and via feedback from using the timeline, helped us spot further improvements, including visibility on optional page components and better signposting for interesting changes like errors or cost movements.</p><p>Since the timeline was derived from the basket projection itself, any additional events or projection changes we made to help this runtime UI also made its way into the data product in our BigQuery warehouse, meaning you can spot an interesting checkout flow and very quickly look for similar ones in BigQuery.</p><h4>Embracing the data mesh</h4><p>We are steadily expanding our data mesh at loveholidays to provide quality data to inform business decision making. This is an ongoing collaboration between our scaling teams, our product delivery teams and our data analysts. Our new checkout data product gives us a clearer picture of how customers buy from us, similar to what our booking data product tells us about customers who book with us.</p><p>By moving away from simply collecting data and instead focusing on a more complete view, we’ve been able to get a better understanding of our customers. We’ve made sure to name our data clearly so it’s easy to understand, and we’ve added more events to track what customers are doing, which helps our engineering teams experiment and learn more. This improved data foundation allows us to explore things like how customers change their flight options, the different pricing strategies we use, how well our suppliers are pricing things, and how effective our processes are for managing inventory and removing duplicates.</p><p>Building the checkout data product demonstrated that innovation doesn’t always equate to building something completely new and we’ve mined a wealth of valuable information from one of our oldest systems. Moreover we now have a definition of the events which the successor system should provide in order to service the analytical requirements of loveholidays.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/354/1*IV8uftMk6dVPaldUNVJC7w.png" /></figure><p>If you’d like to work with us as we solve challenging problems and grow, <a href="https://careers.loveholidays.com/pages/tech-engineering?owner_type=Department&amp;preview=true">we’d love to hear from you</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb46c3397768" width="1" height="1" alt=""><hr><p><a href="https://tech.loveholidays.com/escaping-our-joined-at-the-database-monolith-using-event-sourcing-eb46c3397768">Escaping our joined-at-the-database monolith using event sourcing</a> was originally published in <a href="https://tech.loveholidays.com/">loveholidays tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>