<?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 OCTAVE - John Keells Group on Medium]]></title>
        <description><![CDATA[Stories by OCTAVE - John Keells Group on Medium]]></description>
        <link>https://medium.com/@octave-jkh?source=rss-1f7723ee7a3e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Ty2EzQnvSO68GSJjLbnC_g.jpeg</url>
            <title>Stories by OCTAVE - John Keells Group on Medium</title>
            <link>https://medium.com/@octave-jkh?source=rss-1f7723ee7a3e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 07:08:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@octave-jkh/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[Your Model Isn’t Broken, Your Data Moved: A practical guide to drift detection]]></title>
            <link>https://medium.com/octave-john-keells-group/your-model-isnt-broken-your-data-moved-a-practical-guide-to-drift-detection-894d6d4e922e?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/894d6d4e922e</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-drift]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-11-18T07:18:55.689Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vmSE5b80xDga4wCW9PK5uA.png" /></figure><blockquote>By Darshi Kodithuwakku | Data Science and Engineering Associate</blockquote><p>Imagine a retail team with a customer segmentation model. It sorts customers into high-, medium-, and low-value groups based on how much they spend, and marketing uses those tiers to decide who gets the premium offer. At launch, it works well. A year later, nothing has crashed. Every pipeline is green, and the dashboards load on time. But the “high value” group is now full of fairly ordinary customers, and the premium offers are going to the wrong people.</p><p>Nobody got an error, because the model never actually failed. The world it learned from quietly moved on.</p><p>This is the quiet side of running models in production, and it is easy to miss. Google researchers made a similar point in their ML Test Score paper: real-world ML systems need ongoing testing and monitoring, because they run into problems that never appear in offline experiments (Breck et al., 2017). Today, many everyday business decisions run on live models, making data and model reliability a core part of our job as data professionals. In this article, I’ll walk through that using one specific problem: drift.</p><h4><strong>What drift actually is</strong></h4><p>A model learns its patterns from past data. Drift simply means the data it sees today no longer looks like the data it learned from. There are two main kinds, and the difference is worth getting clear.</p><p><strong>Data drift</strong> (also called covariate shift) means the inputs change. Statistically, the distribution of the features, P(X), has shifted (NannyML, n.d.-a). In our example, prices go up, and a promotion lifts basket sizes, so average spend rises across the board. The old cut-off for “high value” now catches far more people than it used to.</p><p><strong>Concept drift</strong> means the relationship between inputs and outcomes changes: P(Y|X) (Gama et al., 2014; NannyML, n.d.-a). Here the numbers can look almost the same, but what they mean has changed. Say your biggest spenders used to be your most loyal customers, but now a discount-hunting crowd spends heavily and then disappears. On paper, the segment looks unchanged, yet it no longer tells you what you think it does.</p><p>Here is the way I remember it:</p><p><strong><em>Data drift changes what goes in. Concept drift changes what it means.</em></strong></p><p>There is also <strong>label shift</strong>, where the mix of outcomes changes, for example far more customers churning than before (NannyML, n.d.-a). Worth knowing about, but the first two cover most of what you will see in practice.</p><p>Drift usually comes from something ordinary: a new market, a pricing change, a seasonal pattern, or an upstream change like a team renaming a field or changing how spend is calculated. If you want the formal treatment, Gama and colleagues’ survey is the standard academic reference (Gama et al., 2014), and Lu and colleagues offer a more recent overview of detection and adaptation methods (Lu et al., 2019).</p><h4><strong>How to detect it</strong></h4><p>The core idea is simple. Keep a reference dataset (what the model was trained or validated on) and compare it regularly with a current dataset from production. Tools like Evidently work exactly this way, and they pick a suitable statistical test based on the column type and data volume (Evidently AI, 2026a).</p><p>Here are a few practical methods:</p><p>• <strong>Kolmogorov–Smirnov (KS)</strong> test compares two numeric distributions, such as last year’s spend and this month’s (Evidently AI, 2026a).</p><p>• <strong>Population Stability Index (PSI)</strong> is popular in finance and marketing. It scores how much a variable has moved across bins.</p><p>• <strong>Chi-square test</strong> works for categorical features, such as region or product type (Evidently AI, 2026a).</p><p>• <strong>Prediction drift</strong> checks whether the model’s outputs have shifted, for example, whether it suddenly puts 40% of customers in the high tier instead of 15%.</p><p>• <strong>Performance monitoring</strong> tracks accuracy or business results, but only once the true outcomes arrive (Evidently AI, 2025).</p><p>That last one is the catch. Labels usually arrive late. You only find out who stayed loyal weeks or months later, so until then, feature drift and prediction drift are your early warning signs (Evidently AI, 2025). NannyML’s approach goes a step further by checking for multivariate drift, meaning changes in how features move together, not just one column at a time (NannyML, n.d.-b).</p><h4><strong>A practical example</strong></h4><p>The example below simulates a customer-spend feature from two periods: a reference period used when the segmentation model was built, and a current production period after prices and customer behaviour changed. It then runs a two-sample KS test, which compares whether the two samples appear to come from the same underlying distribution (SciPy, n.d.).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/532/1*S2N46UhnIASnYKfDB20l1w.png" /></figure><p>The summary table gives you the business-side view: how many customers fall into each period, what typical spend looks like, and how the 90th percentile has moved. The KS statistic then tells you how far apart the two distributions are, from 0 (identical) to 1 (completely different). In this simulated case, the current distribution is shifted upward, so the test flags a change.</p><p>Don’t stop at the p-value. With thousands of rows, even a trivial shift will give you a tiny one. So look at how big the change is as well, and ask whether it is large enough to affect real decisions. A test can tell you the distribution changed. It cannot tell you whether you have a business problem. An alert is a reason to investigate, not a reason to retrain automatically.</p><h4><strong>What to do when the alert fires</strong></h4><p>The worst reflex is “drift detected, so retrain.” If the real cause is a broken data feed, retraining just teaches the model the bug. So slow down and work through it in order:</p><p>1. Check the data first. Look for pipeline bugs, schema changes, delayed loads, a one-off campaign or plain seasonality. Very often the real answer is hiding here: a late load, a renamed field, a changed business rule, or a campaign that temporarily reshaped customer behaviour.</p><p>2. Measure the impact. Compare segment sizes, revenue, conversion or retention before and after the shift. Some drift turns out to be harmless.</p><p>3. Work out the type. Is it just the inputs moving, or has their link to outcomes changed? The second is usually more serious.</p><p>4. Pick a response that fits. You could recalibrate thresholds, add features, retrain on recent data, rebuild the segments, or fall back to simple rules for a while.</p><p>5. Write it down and monitor. Record what changed and add a check, so you catch it sooner next time.</p><p>The research on adapting to drift backs this up: there is no single fix, and the right strategy depends on the type of change and how quickly it happens (Gama et al., 2014; Lu et al., 2019). A drift alert should start an investigation, not end one. Depending on what you find, the answer might be fixing a data issue, collecting labels, retraining, recalibrating thresholds, or even changing the decision process itself.</p><h4><strong>Where AI agents can help</strong></h4><p>Look back at steps one to three. Someone has to query the data, slice it by region or segment, compare distributions, check upstream tests and write up what they found. It is useful work, but it is slow and repetitive.</p><p>This is where AI agents could help. Imagine an agent that picks up a drift alert, inspects the affected features, compares the reference and current data, checks recent data-quality results, and drafts a short note like: “Spend rose mostly in the new regional market, starting the week of the promotion.” A person then reads that summary and decides what to do.</p><p>I’d keep one boundary very clear: <strong>the agent investigates; it does not decide</strong>. Retraining, deploying and judging business impact still need a human’s approval. This is still an emerging way of working, not a settled practice, and it only works well if the agent has access to clean, well-governed data. The tools for detection already exist. What we need next is not more alerts, but faster, better-evidenced diagnosis, so teams spend less time piecing together context and more time deciding what to do.</p><h4><strong>The evolving data professional</strong></h4><p>Not long ago, our job more or less ended when the model was deployed. Now deployment is where the work starts. Because businesses depend on live models, data professionals are increasingly designing systems that notice when things change, explain why, and keep decisions trustworthy over time.</p><p>The model in our example was never broken. The data moved, and our job is to notice. As more decisions run on live models, the people who matter most will be the ones who can spot that movement early and explain it clearly.</p><h4><strong>References</strong></h4><p>Breck, E., Cai, S., Nielsen, E., Salib, M., &amp; Sculley, D. (2017). The ML test score: A rubric for ML production readiness and technical debt reduction. <a href="https://research.google/pubs/the-ml-test-score-a-rubric-for-ml-production-readiness-and-technical-debt-reduction/">https://research.google/pubs/the-ml-test-score-a-rubric-for-ml-production-readiness-and-technical-debt-reduction/</a></p><p>Evidently AI. (2025a, January 9). What is data drift in ML, and how to detect and handle it. <a href="https://www.evidentlyai.com/ml-in-production/data-drift">https://www.evidentlyai.com/ml-in-production/data-drift</a></p><p>Evidently AI. (2025b, January 25). Model monitoring for ML in production: A comprehensive guide. <a href="https://www.evidentlyai.com/ml-in-production/model-monitoring">https://www.evidentlyai.com/ml-in-production/model-monitoring</a></p><p>Evidently AI. (2026a, January 7). Data drift. <a href="https://docs.evidentlyai.com/metrics/explainer_drift">https://docs.evidentlyai.com/metrics/explainer_drift</a></p><p>Evidently AI. (2026b, January 7). Data drift. <a href="https://docs.evidentlyai.com/metrics/preset_data_drift">https://docs.evidentlyai.com/metrics/preset_data_drift</a></p><p>Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., &amp; Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4), 1–37. <a href="https://doi.org/10.1145/2523813">https://doi.org/10.1145/2523813</a></p><p>Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., &amp; Zhang, G. (2018). Learning under concept drift: A review. IEEE Transactions on Knowledge and Data Engineering. <a href="https://doi.org/10.1109/TKDE.2018.2876857">https://doi.org/10.1109/TKDE.2018.2876857</a></p><p>NannyML. (n.d.-a). Glossary. <a href="https://nannyml.readthedocs.io/en/v0.13.1/glossary.html">https://nannyml.readthedocs.io/en/v0.13.1/glossary.html</a></p><p>NannyML. (n.d.-b). Monitoring a tabular data model | NannyML Cloud. <a href="https://docs.nannyml.com/cloud/model-monitoring/tutorials/monitoring-a-tabular-data-model">https://docs.nannyml.com/cloud/model-monitoring/tutorials/monitoring-a-tabular-data-model</a></p><p>SciPy. (n.d.). ks_2samp — SciPy v1.18.0 manual. <a href="https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.ks_2samp.html">https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.ks_2samp.html</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=894d6d4e922e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/your-model-isnt-broken-your-data-moved-a-practical-guide-to-drift-detection-894d6d4e922e">Your Model Isn’t Broken, Your Data Moved: A practical guide to drift detection</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Analytics Paradox: Why Better Predictions Don’t Always Lead to Better Decisions]]></title>
            <link>https://octave-jkh.medium.com/the-analytics-paradox-why-better-predictions-dont-always-lead-to-better-decisions-df8200263218?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/df8200263218</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[analytics]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-10-28T04:31:03.360Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w_xLq0D90Sh5M60muGIVpw.jpeg" /></figure><blockquote>By Chanuri Sarathchandra | Analytics Delivery Associate</blockquote><p><em>A model can find the right customer and still fail to change what happens to them.</em></p><p>A retail company once spent six months improving a churn model. Accuracy went from 78% to 91% — a genuine, well-earned technical win. The team was confident retention would follow.</p><p>Eighteen months later, retention had moved less than one percent.</p><p>The model wasn’t the problem. Many of the customers it flagged as “highest risk” had already mentally checked out — nothing was going to change their mind. The customers who <em>could</em> still be won back sat further down the list, where nobody was looking. The company had built a better prediction and, almost by accident, a worse decision.</p><p>This is the analytics paradox, and it’s worth naming out loud: a model can get measurably better at prediction without making the business measurably better at anything.</p><p>It shows up more often than most teams admit, and it’s rarely because the model was wrong.</p><h3>Prediction Isn’t the Same Job as Deciding</h3><p>Take a churn model, as an easy example — though the same gap shows up just as often in pricing models, demand forecasts, or recommendation engines, anywhere a prediction is meant to trigger an action. Its job is to flag customers who are likely to leave, so the business can step in — an offer, a call, some form of retention effort. Improve that model and every metric can move in the right direction: better separation between churners and non-churners, better risk ranking, fewer people missed entirely. On paper, it’s a clearly better model.</p><p>And yet retention can barely move. Not because the prediction got worse, but because “knowing who is going to churn” and “doing something that stops them from churning” are two almost separate problems wearing the same trench coat.</p><p>A model can be 90% accurate and still generate very little value, if the customers it flags aren’t reachable, if the intervention doesn’t change their behaviour, or if the business can’t execute against the list consistently. Prediction and decision-making look like the same activity from a distance. They aren’t.</p><p>A prediction is a statement about the world: this customer has a 90% chance of leaving. A decision is a choice about what to do given that statement, and decisions come with constraints predictions never have to deal with — budget, headcount, eligibility rules, the cost of an offer against what the customer is worth.</p><p>Add those constraints in, and the customer with the highest churn probability stops being the obvious person to target. Some of the highest-risk customers are, frankly, already gone no matter what anyone does — the model is right about them, but nothing in the playbook is going to change their mind. Customers a little further down the risk list are often the ones who move when a business reaches out, simply because they’re still on the fence rather than out the door already.</p><p>That’s the gap between predictive power and decision value, and no confusion matrix will ever show it to you.</p><h3>Correct Isn’t the Same as Actionable</h3><p>There’s a second, quieter problem: even a well-prioritised, correct prediction is only as useful as the business’s ability to act on it.</p><p>If a retention team can only work through a few hundred accounts a week, flagging a few thousand doesn’t help much beyond the first few hundred. If the only lever available is a discount, and the discount costs more than the customer’s expected future value, “retaining” them is a loss dressed up as a win. If the list lands after the customer have effectively already left, accuracy stops mattering.</p><p>A highly accurate prediction of something the business can’t change, can’t afford to change, or finds out about too late isn’t worth much more than a rough guess would have been. The real question is never just who’s going to churn. It’s closer to: who’s going to churn, who can still be influenced, and what can realistically be done about it before it’s too late.</p><h3>Did It Work — or Would They Have Stayed Anyway?</h3><p>Even when the targeting and the intervention are right, there’s still an uncomfortable question to answer afterward: did the campaign work, or would those customers have come back regardless?</p><p>Retention going up after a campaign feels like proof. On its own, it isn’t. Customers return for all sorts of reasons that have nothing to do with the offer sent to them — seasonality, a competitor’s price change, a product they’d been waiting for. The only real way to separate what happened to the targeted customers from what happened because they were targeted is to hold back a comparable group and watch what happens to them too.</p><p>That’s simpler to say than to do properly. Test and control groups need to look alike before a campaign starts, not just after — if the test group was already trending better beforehand, a gap afterward isn’t the campaign talking, it’s the trend continuing. Getting this step wrong is a quiet way to convince a team an intervention worked when it didn’t, or the other way round.</p><h3>Beyond Accuracy: What’s Actually Worth Measuring</h3><p>So, if model accuracy alone isn’t the signal, what is?</p><p>It helps to break the question into a simple sequence:</p><ol><li>Was the right thing predicted?</li><li>Were the right customers picked to act on?</li><li>Did the business follow through on that action?</li><li>Did their behaviour change?</li><li>Would it have changed anyway, without any intervention?</li></ol><p>After all that — did the value created outweigh the cost of doing it?</p><p>A model can score brilliantly on step one and still fall apart by step four or five. When that happens, no amount of model improvement fixes it — the problem was never in step one.</p><h3>Start With the Decision, Not the Model</h3><p>The instinct, when a use case underperforms, is to go back and improve the model. Sometimes that’s the right call. Often, though, the model was never the weak link. The weak link was targeting people who couldn’t be moved, offering something unconvincing, being unable to execute the recommendation at the volume needed, or having no real way to tell whether any of it worked.</p><p>Squeezing another couple of points out of accuracy doesn’t fix any of that.</p><p>A more useful starting point is the decision rather than the prediction — asking what decision is actually being improved, what information would make that decision better, whether that information can be predicted, what action follows from it, whether the business can realistically execute that action, and how success will be measured afterward, before getting anywhere near model architecture. That doesn’t make the modelling less important. It just makes it a means to something, rather than the whole project.</p><h3>Where the Value Actually Shows Up</h3><p>Predictive performance is going to keep improving — that part isn’t really in question. But it’s worth remembering that a better model doesn’t automatically hand a business a better outcome. The value shows up further down the chain: in whether the prediction can be acted on, whether the business can execute that action at scale, and whether anyone can tell, afterward, that it made a difference.</p><p>A better prediction is useful. A better decision is valuable. Whether it changed anything is the part the business will remember.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df8200263218" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Prediction to Action: Why a Model Is Only Part of the Decision]]></title>
            <link>https://medium.com/octave-john-keells-group/from-prediction-to-action-why-a-model-is-only-part-of-the-decision-856689f98f3d?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/856689f98f3d</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[predictive-analytics]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 04:31:02 GMT</pubDate>
            <atom:updated>2026-10-21T07:55:49.598Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QxlG2JtiClHt7hfYM3FG_w.jpeg" /></figure><blockquote>By Sanduni Ubayasiri | Data Science and Engineering Associate</blockquote><p>Predictive models can help businesses answer many useful questions. They can estimate which customers are likely to buy, which products may run out of stock, or how much demand to expect. Once a model produces a prediction, the next step may seem simple: act on the customers, products, or situations with the highest scores.</p><p>But a high score does not always point to the best action. A customer who is very likely to buy may have made the purchase even without an offer. In that case, the prediction may be accurate, but the discount may not create any additional value.</p><p>This brings us to an important distinction: a good prediction and a good decision are not always the same thing.</p><p>The offer that shouldn’t go to the obvious customer</p><p>Say you have one promotional offer left, and two candidates:</p><ul><li>Customer A: 85% likely to buy</li><li>Customer B: 45% likely to buy</li></ul><p>Who gets the offer? Every instinct says A.</p><p>But ask a different question: what happens <em>without</em> the offer? If Customer A buys regardless, the promotion doesn’t create a sale, it just quietly eats your margin. Customer B, less “predictable” but more persuadable, might be the one the offer changes.</p><p>This is the gap between two questions that sound similar but aren’t. The prediction answers things like which customer is likely to purchase, which item may run out of stock, which product is at risk of expiry, or how much demand to expect next week. A decision answers a different set: who should receive an offer, how much inventory to order, which item to discount and when, or where to send limited resources.</p><ul><li>Prediction: What’s likely to happen?</li><li>Decision: What should we do about it?</li></ul><p>A model answers the first, but the business needs the second. The score is an input to a decision, not the decision itself.</p><h3>“Likely to respond” isn’t “likely to be influenced”</h3><p>This shows up constantly in retail personalisation. A standard response model ranks customers by purchase likelihood after an offer. The top of that list is usually full of loyal repeat buyers: easy to predict, hard to move, because they were already going to buy.</p><p>A more decision-focused question is: whose behaviour is most likely to change because of the promotion? Rather than predicting only the outcome, the aim is to estimate the difference the offer itself might make, compared with doing nothing.</p><p>Conceptually, customers fall into a few different groups. Some are likely to purchase with or without an offer. Some may be persuaded by it. Others are unlikely to purchase either way. There may even be customers for whom unnecessary communication has a negative effect, sometimes called a “sleeping dogs” effect, where reaching out reduces the chance of a purchase. A pure purchase-probability model doesn’t distinguish between these groups, because it was never built to answer that question.</p><p>That doesn’t make the model wrong. It just means it may be answering a different question than the one the business needs solved.</p><h3>Probability isn’t profit</h3><p>Here’s a simple example that makes the point concrete. To value an offer properly, you need two numbers per customer: how likely they are to buy <em>with</em> the offer, and how likely they’d be to buy <em>without</em> it. The difference between the two tells you what the offer is actually worth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/0*UHf5_rO7u27pkenH" /></figure><p>Customer A looked like the safer bet on raw purchase probability alone. But because A was very likely to buy anyway, the discount mostly just gives away margin on a sale that would have happened for free, so sending the offer loses money. Customer B has a far lower chance of buying overall, but because that chance jumps so much <em>with</em> the offer, sending it creates real, positive value.</p><p>This is also why picking a model threshold based on accuracy alone is a trap. Missing a real stockout is not the same cost as investigating a false alarm. An unnecessary discount and a missed markdown-before-expiry are different kinds of expensive. The right cutoff depends on the decision, not the ROC curve.</p><h3>Reality has a budget</h3><p>Even the “correct” decision must survive contact with constraints. A model might flag 100,000 good candidates for a campaign that only has budget for 10,000. Now the question isn’t “who’s a good target,” it’s “which 10,000, out of everyone good, creates the most value given what can actually be afforded, stocked, and legally contacted.”</p><p>This is where prediction hands off to optimisation and business rules. The model estimates <em>what might happen</em>. The rules define <em>what’s allowed</em>. The constraints define <em>what’s possible</em>. All three must agree before anything ships.</p><h3>The model doesn’t run the store</h3><p>A technically perfect recommendation can still die in the last ten feet: the product is flagged as available but isn’t on the shelf, the offer lands a day late, or the person acting on it gets a recommendation with zero context on why to trust it</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=856689f98f3d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/from-prediction-to-action-why-a-model-is-only-part-of-the-decision-856689f98f3d">From Prediction to Action: Why a Model Is Only Part of the Decision</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Being the Youngest in the Room]]></title>
            <link>https://medium.com/octave-john-keells-group/being-the-youngest-in-the-room-14d089bbd8e7?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/14d089bbd8e7</guid>
            <category><![CDATA[career-development]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[human-resources]]></category>
            <category><![CDATA[work]]></category>
            <category><![CDATA[personal-growth]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 05:01:01 GMT</pubDate>
            <atom:updated>2026-09-30T07:07:19.914Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XxgfljOHMhJaVgXI9qkQmw.jpeg" /></figure><p>A few months ago, I found myself nervously waiting to be called in for my very first interview.</p><p>Just three weeks earlier, I completed my Advanced Level examinations. I was stepping into a world that was completely unfamiliar to me. I had no corporate experience, no extensive portfolio, and very little understanding of what professional life would actually look like. What I did have was curiosity — a curiosity to explore what existed beyond the classrooms I had spent years in and to experience the realities of the corporate world.</p><p>That curiosity led me to OCTAVE. To my surprise and excitement, I was offered the opportunity to join the team as an intern.</p><p>For a brief moment, everything felt like it had fallen into place.</p><p>Then came my first day.</p><p>As I walked into the office that morning, I realised something almost immediately: I was the youngest person in the room.</p><p>The excitement quickly gave way to a whirlwind of questions.</p><p>What if I’m too young for this? Will I be good enough? What if I make mistakes?</p><p>But looking back, I realise that being the youngest person in the room did not determine the value I could bring. It simply meant I had more opportunities to learn.</p><h4><strong>Settling Into the Unknown</strong></h4><p>Stepping into my first corporate role meant entering an environment where almost everything felt unfamiliar. From attending meetings and understanding professional communication to managing responsibilities and meeting deadlines, it was all very new.</p><p>I was suddenly in a space where the decisions made had a real impact; collaboration looked different, and expectations extended beyond simply completing a task. It required me to understand how every contribution would eventually connect to the output and success of a team bigger than myself.</p><p>At times, it felt as though everyone around me knew exactly what had to be done while I was still trying to find my footing. However, over time, I’ve come to understand that growth isn&#39;t an overnight transformation. It is a process built through small improvements, experiences, and lessons learned along the way.</p><h4><strong>Trusting the Process</strong></h4><p>There were moments when I would overthink even the simplest questions before asking them, constantly wondering, “Am I doing this, right?” I distinctly remember reading through every email I sent at least ten times, only to hit “Send” and immediately spot the mistake I had somehow missed all ten times.</p><p>But one of the most important lessons I learned was that no one expected me to get it all right on the first go or have everything known. What mattered was the willingness to learn, adapt, and improve. My team and the people around me were not looking for perfection. They valued the eagerness to learn, accountability, and the effort put into becoming better each day. It was a matter of showing up and taking up the challenge, contributing, and being reliable.</p><p>Instead of trying to have everything figured out, I focused on listening carefully, learning continuously, following through on responsibilities, and improving with every experience.</p><h4><strong>Learning Through Imperfection</strong></h4><p>The biggest mindset shift I experienced was changing the way I viewed mistakes.</p><p>Mistakes are inevitable, especially when you are stepping into unfamiliar territory. They are not signs of failure; they are instances that reveal where there is room to grow. What matters is the way in which you respond to them. They’ve taught me the importance of taking ownership, being open to feedback, and having the humility to recognise that every experience — whether successful or challenging — has something valuable to teach.</p><p>Looking back, some of my biggest lessons came from the moments when I did not get everything right the first time. Those experiences taught me resilience, encouraged me to keep learning, and reminded me that growth comes from being willing to try, make mistakes, reflect, and try again.</p><h4><strong>The People Along the Way</strong></h4><p>Beyond the skills and knowledge, one of the most valuable gains from my journey at OCTAVE was the people I had the opportunity to learn from.</p><p>Beginning my first corporate experience gave me the opportunity to learn from individuals who had travelled paths I was only starting to discover. Every interaction became an opportunity to learn — whether it was understanding how to approach a task, how to communicate effectively, or simply observing how people collaborate and support one another.</p><p>I realized that it is not only about the work you deliver. It is also about the relationships and connections you build along the way.</p><h4><strong>A Culture Built on Trust</strong></h4><p>OCTAVE is more than the work it does — it is defined by the culture it has built from day one. From the very beginning of my internship, one phrase I heard repeatedly was, “OCTAVE is built on trust.” As I experienced the workplace firsthand, I realized just how true those words were.</p><p>The trust that was placed in me, despite being the youngest person in the room, played a significant role in helping me grow more confident in my abilities. It gave me the space to learn, take ownership, and gradually step outside my comfort zone. Surrounded by people who were always willing to guide and support me, I came to understand that growth happens fastest where trust and encouragement go hand in hand.</p><p>When I first started at OCTAVE, the anxiety of being unable to deliver to the expected standard was constantly at the back of my mind.</p><p>Today, I see it differently.</p><p>This being my first corporate experience allowed me to learn with an open mind. It allowed me to ask questions, seek guidance, and absorb knowledge from people with different perspectives and levels of experience. It reminded me that every person, regardless of their position or experience level, has something valuable to contribute.</p><p>The biggest lesson I have learned is that age does not define capability.</p><p>What truly matters is the mindset you bring with you — the willingness to learn, the courage to step outside your comfort zone, and the discipline to keep improving.</p><p>Six months ago, I walked into OCTAVE with curiosity but very little certainty about what awaited me. Here I am today with a better understanding and a greater appreciation for continuous learning, and the confidence that growth comes from embracing experiences that challenge you.</p><p>Being the youngest person in the room was never a limitation. It was a reminder that every experienced professional was once in my shoes — starting somewhere, learning something new, and trying to find their place.</p><p>And perhaps that is the most valuable part of the journey: you get the opportunity to learn, grow, and become the person you aspire to be</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14d089bbd8e7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/being-the-youngest-in-the-room-14d089bbd8e7">Being the Youngest in the Room</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI and Authenticity: Redefining How Brands Connect with Audiences]]></title>
            <link>https://medium.com/octave-john-keells-group/ai-and-authenticity-redefining-how-brands-connect-with-audiences-d90ad038a32c?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d90ad038a32c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[branding]]></category>
            <category><![CDATA[digital-media]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[authenticity]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 05:01:02 GMT</pubDate>
            <atom:updated>2026-09-23T09:22:05.191Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*POHfWBlxPul0RDPKgXdToA.jpeg" /></figure><blockquote><strong>By Tirani Perera — Employer Branding Intern</strong></blockquote><p>AI has changed how quickly we can create content. Perhaps the more important question is whether it is also changing how audiences connect with brands. One interesting shift is that consumers no longer evaluate only the message; they evaluate its credibility. A recent brand campaign featuring a well-known influencer sparked conversation not about the product’s benefits, but about whether the influencer genuinely used it. Many commenters referenced the influencer’s association with another brand. Whether or not those perceptions were accurate, the reaction itself is telling. It suggests that audiences increasingly value authenticity over polished promotion.</p><p>Keeping this insight in mind, we turn to the use of AI in brand content. Do audiences perceive the use of AI as a lack of ownership of the creativity and art of a brand, or does its use push a brand’s standing as an innovative organisation?</p><p>AI has undoubtedly transformed the way marketers work. What once took hours of research, drafting and editing can now be completed in minutes. Brainstorming campaigns, structuring articles, generating social media captions, and analysing customer sentiment have all become significantly more efficient. Rather than replacing creativity, AI has the potential to remove much of the repetitive work that often stands in the way of it.</p><p>A brand is more than the content it produces. It is the experiences it creates, the values it represents, and the consistency with which it delivers on its promises. These are qualities that cannot simply be generated through a prompt. They are developed over time through real interactions with customers, employees, and communities.</p><p>As AI becomes more accessible, the competitive advantage is no longer the ability to create content quickly. Nearly every organisation now has access to similar tools capable of producing articles, graphics, and marketing copy in a matter of seconds. If everyone can produce content at the same speed, speed itself ceases to be the differentiator.</p><p>Instead, audiences begin looking for something that technology cannot easily replicate: a genuine perspective.</p><p>Consumers are becoming increasingly adept at recognising generic content. While AI-generated copy may be grammatically correct and structurally sound, it often lacks the nuances that come from lived experience. Stories about overcoming challenges, lessons learnt from failed initiatives or insights gained through years of industry experience are difficult to manufacture because they are inherently human.</p><p>This is particularly evident in thought leadership.</p><p>The purpose of thought leadership has never been to publish the highest volume of content. Its value comes from offering perspectives that are informed by experience, curiosity, and continuous learning. AI can help organise ideas, improve clarity and accelerate the writing process, but the thinking itself still needs to come from people.</p><p>That does not mean brands should avoid AI altogether. In fact, organisations that refuse to embrace new technologies may find themselves falling behind competitors who are able to work more efficiently and respond to changing market demands faster.</p><p>The challenge, therefore, is not whether to use AI but how to use it responsibly.</p><p>Brands should consider AI as a collaborator rather than a replacement. It can assist with research, provide creative inspiration, summarise complex information, and automate repetitive tasks. These capabilities allow marketers to dedicate more time to understanding their audiences, refining their messaging and building stronger relationships, activities that continue to require empathy, judgement, and human insight.</p><p>Perhaps the most successful brands will not be those that hide their use of AI or those that rely on it entirely. They will be the organisations that strike the right balance between technological capability and human authenticity.</p><p>As AI continues to reshape marketing and communications, authenticity may become one of the few differentiators that cannot be automated. While technology can help brands communicate faster, only people can create trust, credibility, and emotional connection that audiences remember long after the content has been published.</p><p>In an era where everyone has access to the same technology, the brands that stand out will not necessarily be those creating the most content. They will be the ones that remain authentically human. AI may help brands communicate faster, but it is genuine human stories, experiences and perspectives that build trust, foster connection, and leave a lasting impression.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d90ad038a32c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/ai-and-authenticity-redefining-how-brands-connect-with-audiences-d90ad038a32c">AI and Authenticity: Redefining How Brands Connect with Audiences</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is Your Organisation Prepared for a Data-Driven Future?]]></title>
            <link>https://medium.com/octave-john-keells-group/is-your-organisation-prepared-for-a-data-driven-future-85b6872c8daf?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/85b6872c8daf</guid>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 07:54:35 GMT</pubDate>
            <atom:updated>2026-08-31T07:54:35.684Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*24HIiNb72h5vDKIjtmm1yw.png" /></figure><blockquote>By Kimaya Kariyawasam | Lead — Customer Acquisition</blockquote><p>Every leadership team wants to make faster, smarter decisions using data. Yet in practice, this is often harder than it sounds.</p><p>Reports take days to produce. Different departments arrive at the same meeting with different numbers. And when a simple question comes up — <em>“How many customers did we lose last quarter?”</em> — no one can answer it with complete confidence.</p><p>Most organisations aspire to be data-driven, but only a few are truly prepared for it.</p><p>So what does it actually mean to be <strong>data-ready</strong>? More importantly, where does your organisation currently stand?</p><h3>Why This Matters Now</h3><p>Every organisation today is under pressure to move faster and make better-informed decisions. New tools and the growing wave of AI-powered solutions promise to help.</p><p>However, these tools are only as good as the data behind them.</p><p>Feed them messy, incomplete, or unreliable information, and you’ll simply get <strong>fast, confident, wrong answers.</strong></p><p>Simply put, data readiness is no longer a <em>nice-to-have</em>. It’s quickly becoming a fundamental requirement for staying competitive.</p><h3>What “Data Ready” Really Means</h3><p>Setting aside the technical details, data readiness comes down to five practical areas.</p><h3>1. Data Quality</h3><p>Is your data accurate, complete, and consistent?</p><p>Or does it contain duplicates, missing information, and numbers that don’t quite add up?</p><h3>2. Accessibility</h3><p>Can the right people access the information they need when they need it?</p><p>Or is it locked away in spreadsheets, hidden inside systems only one person understands, or buried behind layers of approvals?</p><h3>3. Ownership and Governance</h3><p>Is it clear who is responsible for each dataset?</p><p>Are there simple guidelines for how data should be managed, used, and protected?</p><p>Or does every team handle it a little differently?</p><h3>4. Culture</h3><p>When decisions are made, does the team check the data first or rely mainly on instinct?</p><p>Do people trust the numbers they’re looking at, or do they quietly work around them?</p><h3>5. Connected Systems</h3><p>Do your tools and platforms share information seamlessly?</p><p>Or does data live in separate silos — one system for sales, another for finance, another for operations — with no easy way to bring everything together?</p><p>If a few of these areas feel like they’re still a work in progress, that’s completely normal.</p><p>Very few organisations score highly across all five dimensions.</p><p>The goal isn’t perfection. It’s understanding where you stand today so you can confidently decide what comes next.</p><h3>A Quick Self-Assessment</h3><p>Rather than guessing where your organisation stands, take a few minutes to find out.</p><p>We’ve put together a short, interactive assessment that evaluates the five key dimensions of data readiness:</p><ul><li>Data Quality</li><li>Accessibility</li><li>Ownership</li><li>Culture</li><li>Connected Systems</li></ul><p><strong>Take the interactive assessment </strong><a href="https://anishkakariyawasam-jpg.github.io/Octave-Data-Readiness-Assesment/"><strong>here</strong> </a>and receive your results instantly.</p><h3>Turning Insight Into Action</h3><p>Wherever your organization lands, the good news is that progress doesn’t require a complete overhaul.</p><h3>If you’re ahead of the curve</h3><p>Focus on extracting even more value from what you already have.</p><p>This could mean automating routine reporting, enabling self-service analytics, or exploring how AI can build on your existing data capabilities.</p><h3>If you’re somewhere in the middle</h3><p>Identify one area to improve first.</p><p>In many organizations, strengthening data quality or improving accessibility delivers the fastest and most visible impact.</p><h3>If you’re just getting started</h3><p>Start with the fundamentals.</p><p>Choose one important dataset, clean it up, assign clear ownership, and ensure the people who need it can access it easily.</p><p>Steady, focused improvements are almost always more effective than trying to fix everything at once.</p><h3>The Bottom Line</h3><p>Data readiness isn’t a destination — it’s an organizational capability that develops over time.</p><p>Some organizations are just beginning the journey. Others are well down the path.</p><p>Most fall somewhere in between, and that’s perfectly normal.</p><p>What matters isn’t being perfect.</p><p>What matters is knowing where you are today, so you can take the next step with confidence.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85b6872c8daf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/is-your-organisation-prepared-for-a-data-driven-future-85b6872c8daf">Is Your Organisation Prepared for a Data-Driven Future?</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hidden Cost of Poor Governance: Projects Fail Long Before Delivery — the key is to identify the…]]></title>
            <link>https://octave-jkh.medium.com/the-hidden-cost-of-poor-governance-projects-fail-long-before-delivery-the-key-is-to-identify-the-e3df6145c9a9?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3df6145c9a9</guid>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[project-management]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 19 Aug 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-08-19T04:31:03.659Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ux50UYPYCqSgtVbtBpwhfA.png" /></figure><h3><strong>The Hidden Cost of Poor Governance: Projects Fail Long Before Delivery — the key is to identify the signs?!</strong></h3><blockquote>By Vishva Karunaratne | Head — PMO</blockquote><p>Teams remain busy.</p><p>Meetings continue.</p><p>Reports are submitted.</p><p>Milestones appear on schedule.</p><p>Yet beneath the surface, the project is quietly laying out the foundation of a potential failure.</p><p>Most projects don’t fail during execution. They fail much earlier — during planning!</p><p>By the time stakeholders realise a project is over budget, behind schedule, or unable to deliver the expected outcomes, the underlying causes have often existed for months.</p><p><strong>The Misconception Around Project Governance</strong></p><p>Ask ten people what project governance means, and you’ll likely receive ten different answers. Some believe governance is a reporting structure.</p><p>Others associate it with project documentation, status meetings, communication plans, or steering committee presentations.</p><p>While these are all components that support governance, <strong>they are not governance itself.</strong></p><p>Project governance is the framework through which decisions are made, authority is assigned, accountability is established, risks are escalated, priorities are agreed upon, and changes are controlled throughout the project lifecycle. Most importantly, not the full power and authority of the project is provided to one stakeholder — to avoid risks, power and authority should be distributed through various mechanisms.</p><p>Key questions to consider:</p><ul><li>Who has the authority to make which decisions?</li><li>What decisions require executive approval?</li><li>Should decisions be evaluated by a higher authority within the client organisation or different parties?</li><li>How are risks and issues escalated?</li><li>How are competing priorities resolved?</li><li>Who is accountable when project objectives conflict with business expectations?</li><li>How are changes evaluated before they become commitments?</li></ul><p>Without clearly answering these questions, projects do not have governance — they merely have administration.</p><p>The absence of true governance often goes unnoticed because projects continue moving forward.</p><p>Unfortunately, activity without governance often becomes expensive motion rather than controlled delivery.</p><p><strong>Governance Is Not Meant to Be Redesigned Periodically.</strong></p><p>One of the most damaging practices seen in struggling projects is the continual alteration of governance itself.</p><p>Ways of working and processes change due to changes in decision makers, reporting lines, approval authorities, escalations, or evolving success criteria.</p><p>A governance structure should provide stability throughout the project.</p><p>This does not imply that governance is inflexible. Business priorities can change. Market conditions may shift. Regulatory requirements may emerge. New stakeholders may become involved. Despite all these changes, governance structure should remain static with controlled mechanisms in place to cater to these ever-altering changes.</p><p>If governance must change, it should be:</p><ul><li>Formally assessed</li><li>Approved through the appropriate governance forum</li><li>Documented</li><li>Communicated to all stakeholders</li><li>Embedded into project controls</li></ul><p>Otherwise, every governance change introduces uncertainty, confusion, and inconsistent decision-making — which erodes confidence overall.</p><p>Projects need stable rules if they are expected to deliver stable outcomes.</p><p><strong>Planning Is Much More Than Creating a Project Schedule</strong></p><p>Perhaps the greatest misunderstanding in project management is the belief that planning simply means creating a Gantt chart, estimating effort, assigning resources, and developing a budget.</p><p>That is scheduling.</p><p>Planning is far broader.</p><p>Effective project planning establishes a shared understanding of the key questions:</p><ul><li>What is being delivered</li><li>Why it is being delivered</li><li>How success will be measured</li><li>What assumptions exist</li><li>What risks are acceptable, who makes decisions</li><li>How changes will be governed</li><li>What constraints exist</li></ul><p><strong>Most importantly — how the organisation will respond when those assumptions inevitably change.</strong></p><p>Many established bodies, such as The PMBOK® Guide (PMI), PRINCE2, ISO 21502 — Project, Program and Portfolio Management, and even Systems Engineering in NASA/INCOSE, consistently emphasise progressive elaboration and invest sufficient effort in establishing objectives, governance, scope, risks, stakeholders, and decision-making before execution. This is a principle that spans methodologies: predictive, agile, hybrid, government, engineering, and international standards.</p><p>Organisations often believe they are saving time by shortening planning.</p><p>In reality, they are postponing the work until execution — where every unresolved decision becomes significantly more expensive.</p><p>A project with fifty planning documents may still be poorly planned if stakeholders have different interpretations of the project’s objectives or governance arrangements. Conversely, a lean Agile project may produce relatively few artefacts yet demonstrate excellent planning because its objectives, governance, risks, decision-making authority, and priorities are clearly understood by everyone involved. However, poor Agile measures could also lead to the same issue.</p><p>The key you are looking for is — <strong>planning maturity is measured by clarity, alignment, and informed decision-making, not by the volume of documentation.</strong></p><p><strong>Planning Maturity!</strong></p><p>Consider a familiar scenario.</p><p>The project begins before scope is fully understood because delivery pressure exists.</p><p>Requirements continue evolving because stakeholders were never aligned.</p><p>The project manager attempts to accommodate changes informally to maintain momentum.</p><p>Different stakeholders begin interpreting project objectives differently.</p><p>To patch up the current issues, ways of working in adopting this become the norm.</p><p>Developers build features that another business unit later rejects.</p><p>Testing uncovers defects that are requirement misunderstandings rather than technical issues.</p><p>Delivery milestones begin slipping.</p><p>Executive pressure increases.</p><p>Teams work longer hours.</p><p>Confidence begins to decline.</p><p>The project manager spends more time managing expectations than managing delivery.</p><p>Eventually, meetings become centred on explaining delays rather than solving root causes.</p><p>None of these issues occurs because the team lacks capability.</p><p><strong>They occur because the project started without sufficient clarity, decision-making discipline, and governance.</strong></p><p>This pattern appears regardless of whether the project follows Predictive, Agile, Hybrid, or any other delivery methodology.</p><p><em>Methodologies cannot compensate for weak governance.</em></p><p><em>Framework cannot replace disciplined planning.</em></p><p><em>Technology cannot overcome poor decision-making.</em></p><p>Key signs of identifying the issue:</p><ul><li>One of the earliest indicators of a project in difficulty is not a missed milestone — it is when key stakeholders provide different explanations of the project’s objectives.</li><li>Projects rarely lose control overnight. They lose control incrementally, one undocumented decision and one unmanaged change at a time.</li><li>When a project manager spends more time negotiating expectations than managing delivery, the root cause often lies in governance rather than execution.</li><li>If every steering committee meeting results in new priorities without reassessing the project’s objectives, governance has shifted from providing direction to creating instability.</li><li>A project that frequently revisits decisions made during initiation is often revealing weaknesses in planning rather than changes in business needs.</li><li>The first casualty of poor governance is clarity. Once clarity is lost, confidence, accountability, and delivery performance often follow.</li></ul><p>Introduction of a project alignment index helps teams to a common mental location:</p><p>Ask five stakeholders separately:</p><ul><li>Why does this project exist?</li><li>What problem are we solving?</li><li>What does success look like?</li><li>What isn’t included?</li><li>Who approves of change?</li></ul><p>If every answer is different…</p><p>Planning isn’t complete.</p><p>That is a practical tool that people can immediately apply.</p><p><strong>The Hidden Cost Is Not Measured in Currency Alone.</strong></p><p>When governance is weak, and planning is incomplete, the financial consequences are usually visible.</p><p>Budgets increase.</p><p>Schedules slip.</p><p>Resources remain allocated longer than expected.</p><p>However, the less visible costs are often more damaging.</p><p>Teams gradually lose confidence.</p><p>People become reluctant to raise risks because previous concerns were ignored.</p><p>Project managers become administrators rather than leaders.</p><p>Business stakeholders lose trust in delivery teams.</p><p>Delivery teams lose trust in business stakeholders.</p><p>Every new change creates frustration instead of improvement.</p><p>Eventually, the organisation begins questioning project management itself.</p><p>The unfortunate irony is that project management is often blamed for problems it was never empowered to prevent.</p><p>A project manager cannot compensate for unclear governance.</p><p>A delivery team cannot succeed without stable decision-making.</p><p><strong>So, what does success look like?</strong></p><p>Planning is not an ‘x’ number of documents.</p><p><strong>Planning is an organisational alignment.</strong></p><p>The realisation of what kind of governance structure suits the organisation is based on how it&#39;s structured, the definition of authority within the organisation, and how decisions are made!</p><p>Planning is complete when the team understands what they are delivering, why they are delivering it, who will make decisions, what assumptions are being made, and how inevitable change will be managed and, most importantly — with requirement clarity!</p><p>Let’s consider Complexity Drivers.</p><p>Ask ten simple questions.</p><ul><li>How many stakeholder groups exist?</li><li>How many business units are involved?</li><li>How many delivery teams?</li><li>Are external vendors involved?</li><li>Are regulators involved?</li><li>Is the scope well understood?</li><li>Are requirements likely to evolve?</li><li>Does the project affect critical operations?</li><li>Does failure have significant financial or reputational consequences?</li><li>How many system integrations exist?</li></ul><p>Every “yes” increases planning.</p><p>Notice that project size isn’t even the biggest factor.</p><p><strong>Complexity is</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/777/1*8CnB3fHCsarX8Ko5ommz6w.png" /></figure><p>Hence, should complexity and alignment meet to achieve this maturity?</p><p>Key takeaways:</p><ul><li>Ensure projects have adequate mature governance achieved through adequate mature planning</li><li>Measuring project complexities helps achieve adequate governance and planning</li><li>Understanding the organization which the project serves helps achieve the level of maturity required</li><li>Embarking on a project alignment index</li><li>Planning need not be sequential; it can be distributed among the project at different stages to reduce the risk of failure.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3df6145c9a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Guardians of Production: A Behind-the-Scenes Look at LiveOps Engineers]]></title>
            <link>https://octave-jkh.medium.com/the-guardians-of-production-a-behind-the-scenes-look-at-liveops-engineers-31ab8463bfea?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/31ab8463bfea</guid>
            <category><![CDATA[dataops]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[live-ops]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[governance]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 29 Jul 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-07-29T04:31:03.798Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Qucrp6sKkzVQFuPsm6pbTw.png" /></figure><blockquote>By Anjalika Abeywardana | Senior LiveOps Engineer</blockquote><p>When people think about data platforms and advanced analytics solutions, they often focus on dashboards, machine learning models, data pipelines, and insights delivered to the business. However, behind every successful production environment is a dedicated LiveOps team working to ensure everything runs smoothly, securely, and reliably.</p><p>LiveOps teams play a critical role in managing production environments, supporting users across the organisation, handling production releases, enforcing governance standards, and ensuring that business-critical systems remain operational at all times.</p><p>In this article, let’s look at the day-to-day responsibilities of a LiveOps team, the challenges they face, and how their work helps maintain the stability of modern data platforms.</p><h4><strong>Managing the Production Environment</strong></h4><p>The production environment is where business-critical workloads run. Data pipelines process information, dashboards refresh for decision-makers, reports are generated, and users depend on the platform to perform as expected.</p><p>The primary responsibility of a LiveOps team is to ensure that this environment remains stable, secure, and available.</p><p>This involves monitoring production jobs, investigating failures, coordinating recovery efforts when issues occur, and working closely with Data Scientist teams to resolve production incidents. LiveOps engineers continuously monitor platform health, track resource utilisation, validate critical production data, and ensure that scheduled workloads are executed successfully.</p><p>They also play an important role in supporting disaster recovery and business continuity activities, helping organisations remain resilient in the event of unexpected disruptions.</p><p>While many users only see the final output, there is often a significant amount of operational work happening behind the scenes to keep services running without interruption.</p><h4><strong>Handling Production Releases</strong></h4><p>Every new feature, enhancement, bug fix, or platform improvement eventually needs to be deployed to production.</p><p>Before any release takes place, the LiveOps team carefully reviews deployment requests, validates that all required approvals have been obtained, assesses potential risks, and ensures that deployment procedures follow organisational standards.</p><p>Release activities often involve coordination between multiple teams, including Data Engineers, Data Scientists, BI Engineers, Architects, and business stakeholders.</p><p>Beyond deployment approvals, LiveOps engineers coordinate release schedules, manage deployment windows, support platform upgrades and migrations, oversee rollback procedures when required, and perform post-release validation checks to ensure systems continue to operate as expected.</p><p>The goal is simple: deliver new functionality while minimising risk to the production environment.</p><h4><strong>Governance and Compliance</strong></h4><p>One of the most important responsibilities of LiveOps is to ensure that production standards are consistently followed.</p><p>Not every request can be approved immediately, and not every change can be implemented without proper review. Governance processes exist to protect the integrity, security, and reliability of the platform.</p><p>Whether reviewing deployment requests, validating configurations, maintaining operational documentation, supporting audit requirements, or ensuring that operational procedures are followed, LiveOps teams act as custodians of the production environment, helping prevent issues before they occur.</p><p>Whilst governance may sometimes seem restrictive, it ultimately enables organisations to scale safely and maintain trust in their systems and data.</p><h4><strong>Access Management</strong></h4><p>Providing users with the right level of access is another key responsibility for LiveOps.</p><p>New team members join projects, responsibilities change, and users frequently require access to different environments, workspaces, datasets, and tools.</p><p>Each request must be carefully reviewed to ensure that permission aligns with security policies and business requirements.</p><p>In addition to access provisioning, LiveOps teams often manage user onboarding and offboarding activities, maintain service accounts, perform periodic access reviews, and ensure that credentials and permissions are managed securely.</p><p>The objective is to provide users with the access they need to perform their work effectively while maintaining strong security controls and protecting sensitive data.</p><h4><strong>Supporting Platform Users</strong></h4><p>LiveOps teams work closely with a wide range of stakeholders across the organisation.</p><p>This includes supporting Data Scientists in building machine learning models, Data Engineers in developing pipelines, BI Engineers in creating dashboards, and business users in consuming data products.</p><p>Support requests can range from troubleshooting failed jobs and resolving access issues to assisting with deployments, providing operational guidance, and helping teams navigate platform best practices.</p><p>In addition to these responsibilities, LiveOps engineers act as a dedicated support function for Data Scientists, Data Engineers, and BI Engineers. They assist with production-related issues, deployments, enhancements to existing pipelines and dashboards, and change requests, helping teams deliver and maintain reliable solutions within the production environment.</p><p>Because LiveOps teams interact with so many different users and technologies, they often serve as a central point of coordination across the platform, bridging the gap between technical teams and business stakeholders.</p><p>In many ways, LiveOps engineers become the “all-rounders” of the platform, wearing multiple hats and stepping in wherever support is needed, helping to keep both the technology and the people behind it moving forward.</p><h4><strong>Continuous Improvement and Operational Excellence</strong></h4><p>While maintaining daily operations is a key responsibility, LiveOps teams are also focused on continuously improving the platform and the processes that support it.</p><p>This includes identifying opportunities to streamline operational workflows, improving monitoring and alerting mechanisms, maintaining operational runbooks, standardising procedures, and reducing recurring issues through process improvements.</p><p>LiveOps engineers also contribute to performance optimisation and cost management initiatives, helping ensure that the platform remains efficient, scalable, and sustainable as business demands grow.</p><h4><strong>Incident Response and Problem Solving</strong></h4><p>Despite careful planning and rigorous processes, production issues can still occur.</p><p>When they do, LiveOps teams are often among the first responders.</p><p>Investigating failures, coordinating communication, identifying root causes, and restoring services quickly are all part of the role. These situations often require collaboration across multiple teams and the ability to make informed decisions under pressure.</p><p>Beyond immediate resolution, LiveOps engineers often conduct Root Cause Analysis (RCA) activities, identify recurring patterns, drive preventative improvements, and track operational service metrics to improve platform reliability over time.</p><p>Strong operational processes and technical expertise are essential for ensuring that issues are resolved efficiently while minimising business impact.</p><h4><strong>Looking Ahead</strong></h4><p>As platforms continue to grow and evolve, so do the responsibilities of LiveOps teams.</p><p>The industry is already seeing the rise of coding assistants, intelligent monitoring tools, automated approvals, and agent-based solutions that can help streamline operational processes. Many routine tasks that once required significant manual effort are becoming increasingly automated, helping teams work more efficiently and respond more quickly.</p><p>However, production operations involve much more than reviewing requests and following predefined rules. Behind every deployment, incident, access request, or production issue are business requirements, technical dependencies, governance considerations, and stakeholder expectations that often require human judgment and collaboration.</p><p>LiveOps engineers work with a wide range of stakeholders, including Data Scientists, Data Engineers, BI Engineers, Architects, project teams, and business users. Resolving production challenges often requires bringing together multiple perspectives, understanding the broader business context, and coordinating across teams to reach the right solution.</p><p>While automation can assist with identifying issues, reviewing configurations, or enforcing standards, it cannot fully replace the communication, collaboration, and domain knowledge required to manage a production environment effectively. A production incident is rarely solved through technical analysis alone; it often requires conversations, decision-making, prioritisation, and alignment between multiple teams.</p><p>As operational tools continue to evolve, the role of LiveOps engineers will likely shift toward higher-value activities, with a greater focus on governance, stakeholder management, problem-solving, and cross-functional collaboration. Technology may change, but the need for people who can connect teams, understand business processes, and make informed operational decisions will remain as important as ever.</p><h4><strong>Conclusion</strong></h4><p>LiveOps is much more than simply managing production systems. It is about ensuring stability, enabling innovation, protecting governance standards, supporting users, and helping organisations deliver reliable data products and services.</p><p>Whilst much of the work happens behind the scenes, LiveOps’ teams play a vital role in keeping platforms running smoothly and ensuring that business users can trust the systems they rely on every day.</p><p>In many ways, LiveOps serves as the bridge between development and production, balancing operational excellence with business needs to ensure that the right changes reach the right environment at the right time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=31ab8463bfea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Tools Got Better. Did We?]]></title>
            <link>https://octave-jkh.medium.com/the-tools-got-better-did-we-4469c998b959?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4469c998b959</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[growth]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-07-22T04:31:03.611Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sCM2Gi7ThFmeL3RENdWnuA.png" /></figure><blockquote>By Yumna Fazal | Analytics Delivery Associate Level II</blockquote><p>There is a particular kind of quiet confidence that comes with being the person in the room who can pull things together quickly. In the point of view of an analytics delivery associate, that has always been part of the job description. You are the one translating between the data scientists and the business stakeholders, packaging complexity into something digestible, building the narrative that makes a use case land. For a long time, doing that well meant putting in the hours. A stakeholder update required sitting with the numbers. A pitch deck for a new POC required real thought about what the business needed to hear. There was a natural friction in the process, and that friction, it turns out, was doing a lot of useful work.</p><p>Then the tools got significantly better, and the friction largely disappeared.</p><p>Today, I can prompt my way to a first draft of a business update in minutes. I can feed a set of rough notes into an AI tool and get back a structured narrative with a logical flow and clean language. I can generate five different angles for a pitch before I have even had a second coffee. The blank page is no longer the obstacle it once was. And I would be lying if I said that was not genuinely useful, because it is. But it has also surfaced a question I think a lot of people in non-technical roles inside technical fields are quietly sitting with, even if they have not quite put words to it yet.</p><p>If the production of output is now cheap and fast, what exactly is my value?</p><h4><strong>When Everyone Has the Same Superpower</strong></h4><p>The honest answer to that question starts with acknowledging something slightly uncomfortable. The productivity gains that AI tools offer is not unique to you. They are available to everyone on your team, across your organisation, and across every organisation in your industry. The person at the competing firm pitching a similar use case to a similar business stakeholder is working with the same tools. The analyst at the next desk drafting their own version of a progress update has access to the same capability.</p><p>What this creates, almost inevitably, is a kind of convergence. The outputs start to look and sound similar. The slide structures follow the same logic. The language carries the same polished, slightly neutral tone. The executive summaries hit the same beats in the same order. Not because anyone is copying anyone else, but because everyone is drawing from the same well and asking it roughly the same questions.</p><p>In my role, this matters more than it might in other fields. A large part of the job is building credibility with business stakeholders who may not fully understand the technical work happening underneath. That credibility is not built through volume. It is built through the sense that the person presenting to them has genuinely thought about their problem, understands their constraints, and has made deliberate choices about what to include and what to leave out. The moment a stakeholder starts to feel like they are reading something that could have been written about any project, in any organisation, by any team, you have lost something that is very hard to get back.</p><p>The risk of AI in non-technical delivery roles is not that it makes us redundant. It is that it makes us interchangeable.</p><h4><strong>The Thing AI Does Not Know</strong></h4><p>There is a distinction worth making here between articulation and insight, because AI is exceptionally good at one of them.</p><p>Give an AI tool a messy set of notes about a use case and it will return something that reads clearly, flows logically, and sounds professional. It will smooth the edges. It will add structure. It will produce the shape of a well-considered argument. What it cannot do is know whether that argument is the right one for this stakeholder, in this moment, given the politics sitting underneath the surface of the project. It does not know that the head of the business who will be in the room is still skeptical about the last initiative and needs to see something different from the standard progress narrative. It does not know that the POC framing that worked three months ago has since been undermined by a failed pilot in another part of the business. It does not know that the most important thing you could put in the update this week is one honest sentence about a delay, because trust is more valuable right now than polish.</p><p>That knowledge comes from being in the room. From picking up on what the business is not saying as much as what it is. From the accumulation of context that does not live in any document and cannot be fed into a prompt.</p><p>This is what critical thinking looks like in an analytics delivery role. Not the academic version, but the practical, applied version. It is the ability to look at what the AI has produced and ask: is this true to the situation? Is this what this audience needs? Does every slide here actually earn its place? Would someone reading this feel genuinely informed, or just professionally handled?</p><p>Those questions require judgment. And judgment, at least for now, is still entirely yours.</p><h4><strong>Scaling Back as a Signal</strong></h4><p>One of the more counterintuitive shifts that AI has brought about is that restraint has become a differentiator. When generating more content requires almost no additional effort, the decision to generate less becomes meaningful. A tightly constructed five-slide update, where every slide has a clear purpose and nothing is there to fill space, says something about the person who built it. It signals that choices were made. That there was enough understanding of the audience and the material to know what could be cut.</p><p>In contrast, a twelve-slide deck where eight of the slides are variations on the same point, smoothly written and professionally formatted, starts to read as exactly what it is. The shape of thinking without thinking itself.</p><p>This is something that stakeholders are beginning to develop an instinct for, even if they cannot always articulate it. There is a texture to work that has been genuinely considered. A specificity to the examples. A willingness to take a position rather than present every possible angle with equal weight. A sentence here or there that could only have come from someone who has been close to the problem for long enough to develop a real perspective on it.</p><p>That texture is not something AI produces naturally. It is something you bring to the process, and it is most visible when you use AI as a thinking tool (rather — the only thinking tool) rather than a production tool. The difference is in how you approach the work. Instead of opening a blank prompt and asking for a draft, you might start by asking the AI to identify the tensions in your raw notes, or to argue against the framing you are considering, or to tell you which part of your narrative is doing the least work. You use it to pressure-test the thinking before you get anywhere near the output.</p><p>The draft it produces at the end of that process will be materially better. But more importantly, so will your understanding of what you want to say.</p><h4><strong>What this Means Going Forward</strong></h4><p>There is a version of the next few years where the people who thrive in non-technical roles are the ones who become the most prolific users of AI tools. More updates, more decks, more POC narratives, more artefacts moving through the pipeline at higher velocity. And there is a different version where the people who thrive are the ones who used AI to slow down in the right places. Who spent the time the tools gave them back on deeper stakeholder conversations, on genuinely understanding what a business unit is trying to achieve before proposing how analytics can help, on being the person in the room whose judgment could be trusted precisely because it was evidently theirs.</p><p>The second version is harder to measure on a dashboard. It does not show up in the number of documents produced or the speed at which updates went out. But it is the version that builds the kind of professional reputation that compounds over time.</p><p>The tools got better. That is genuinely exciting. But the question it raises is not how to use them more. It is how to use the space they create more wisely. In a field where the technical work is already being done, the non-technical professional has always been there to bridge a gap. To make the complex legible, the ambiguous navigable, and the uncertain somehow worth committing to.</p><p>AI can help with the first part of that. The legibility. The structure. The polish.</p><p>The rest of it, the navigation, the judgment, the human read of a room, is still the job. And if anything, it matters more now than it did before.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4469c998b959" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Privacy & Security — Protecting the System]]></title>
            <link>https://medium.com/octave-john-keells-group/privacy-security-protecting-the-system-445746504ce0?source=rss-1f7723ee7a3e------2</link>
            <guid isPermaLink="false">https://medium.com/p/445746504ce0</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[OCTAVE - John Keells Group]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 04:31:04 GMT</pubDate>
            <atom:updated>2026-07-08T05:25:43.214Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2AXkz0A3AwQ21c3s9wJCjA.png" /></figure><blockquote>By Anuja Dassanayake | Data Scientist Level II</blockquote><p>Data is the foundation of everything OCTAVE builds. It comes from a wide spectrum of Group operations: demand signals generated at retail outlets, guest behaviour patterns within hospitality properties, financial transaction histories across banking and services, and policy and operational data from insurance, logistics, and other business units. The breadth of information flowing through the John Keells Group is one of our greatest analytical assets and also one of our most significant responsibilities.</p><p>The privacy and security risks associated with AI systems are broader than most teams initially appreciate. The data layer is familiar territory: how information is collected, stored, and accessed. But protecting data in an AI context means thinking about the full lifecycle. From the moment raw information enters a pipeline, through model training, to how a deployed system responds to queries. Vulnerabilities exist at every stage.</p><p>At the data layer, the fundamentals matter more than they are often given credit for. Encryption at rest and in transit is the baseline, not the ceiling. PII masking and tokenisation ensure that sensitive attributes, names, national identity numbers, and account details are either removed or replaced before data reaches analytical pipelines. This limits the blast radius if something goes wrong downstream. Role-based access controls and data minimisation, building models on only the fields they genuinely need, reduce exposure further. These interventions are unglamorous, but they are what separates a data infrastructure that contains a breach from one that amplifies it.</p><p>The model itself introduces a category of risk that data-layer controls alone do not address. Research has demonstrated that large models can inadvertently memorise and reproduce fragments of their training data under certain conditions. In contexts where models are trained on sensitive customer or financial information, this is a concrete exposure. A model that has memorised specific personal details is, in effect, a leaky data store, one that may not reveal itself until it is queried in just the right way. Inference-based attacks go further still. Techniques like membership inference allow an adversary to determine whether a specific individual’s data was included in a training set, without accessing the data directly. Model inversion attacks can reconstruct approximate versions of training inputs from model outputs alone.</p><p>On the defensive side, differential privacy offers a mathematically grounded framework for training models in ways that limit how much any individual record can influence the output. Federated learning addresses a different surface: rather than centralising data for training, model updates are computed locally, and only the gradients are shared, keeping raw data where it originated. Both involve real tradeoffs in model performance or infrastructure complexity, but those tradeoffs are increasingly accepted when the sensitivity of the underlying data demands it.</p><p>Deployment introduces its own considerations. Access logging and anomaly detection on model endpoints are becoming standard as firewall rules on application servers, because a model API left unmonitored is an attack surface like any other. Adversarial inputs crafted to cause misclassification and prompt injection in language model applications are areas where the industry is still developing consistent defensive practices.</p><p>What is clear from working across the JKH Group’s industries is that privacy and security cannot be retrofitted. The decisions made early, about what data is collected, how it is masked and encrypted, what fields models are trained on, and how endpoints are monitored, determine the exposure that teams will be managing for the life of the system. The time to think carefully about these questions is before the first pipeline is built.</p><h4><strong>Further Reading</strong></h4><h4><strong>For Business Readers</strong></h4><h4><strong>Microsoft Security Insider</strong></h4><h4><strong>Data Governance and Security Guide</strong></h4><p>Written for business and security leadership, this covers how to build observability across structured, unstructured, and AI-generated data; how to assign ownership across CIO, CISO, CDO, and CPO roles; and what a Classify-Label-Protect-Manage governance model looks like in practice.</p><p><a href="https://www.microsoft.com/en-us/security/security-insider/emerging-trends/ai-security-guide-data-governance-and-security">https://www.microsoft.com/en-us/security/security-insider/emerging-trends/ai-security-guide-data-governance-and-security</a></p><h4><strong>For Technical Readers</strong></h4><h4><strong>Blockchain Council</strong></h4><h4><strong>Privacy-Preserving AI: Differential Privacy vs Federated Learning vs Secure Enclaves</strong></h4><p>A 2025–2026 technical guide comparing the three dominant privacy-preserving architectures, explaining the tradeoffs between each, and outlining how they are increasingly being combined in production deployments for layered protection against reconstruction and inference attacks.</p><p><a href="https://www.blockchain-council.org/ai/privacy-preserving-ai-differential-privacy-federated-learning-secure-enclaves/">https://www.blockchain-council.org/ai/privacy-preserving-ai-differential-privacy-federated-learning-secure-enclaves/</a></p><h4><strong>Databricks Blog</strong></h4><h4><strong>ACM Computing Surveys — <em>A Survey of Privacy Attacks in Machine Learning</em></strong></h4><p>A peer-reviewed survey of over 50 research papers covering the full taxonomy of ML privacy attacks: membership inference, model inversion, property inference, and model stealing. For data scientists and engineers building models on sensitive enterprise data, this is the clearest map of what the threat landscape looks like and what defenses are available against each attack type. <br>https://dl.acm.org/doi/full/10.1145/3624010</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=445746504ce0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/octave-john-keells-group/privacy-security-protecting-the-system-445746504ce0">Privacy &amp; Security — Protecting the System</a> was originally published in <a href="https://medium.com/octave-john-keells-group">OCTAVE — John Keells Group</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>