<?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 Leela Kumili on Medium]]></title>
        <description><![CDATA[Stories by Leela Kumili on Medium]]></description>
        <link>https://medium.com/@leela.kumili?source=rss-d30d57fc8a21------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*VPB2QX0sv_eQWN-x4jIGVA.png</url>
            <title>Stories by Leela Kumili on Medium</title>
            <link>https://medium.com/@leela.kumili?source=rss-d30d57fc8a21------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 23:01:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@leela.kumili/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[Every Large Migration Becomes a Data Quality Project. Plan for It.]]></title>
            <link>https://medium.com/@leela.kumili/every-large-migration-becomes-a-data-quality-project-plan-for-it-3163b609b1c9?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/3163b609b1c9</guid>
            <category><![CDATA[what-not-to-do]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[what-to-do]]></category>
            <category><![CDATA[migration]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 01:56:01 GMT</pubDate>
            <atom:updated>2026-10-27T01:56:01.444Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a9EFL8B5r2pnHqf5" /></figure><p>A database migration sounds simple when you describe it in one sentence.</p><p>Move the data from the old system to the new one.</p><p>That was the assignment.</p><p>Last month, I wrote about the tooling side of this migration and the lesson that nobody else could run the tool I built.</p><p>This is the other half of that story.</p><p>And it turned out to be the harder one.</p><p>Because then you discover that the source database accepts things the target database does not.</p><p>A field that was supposed to have one shape has several. A value that looked harmless for years suddenly violates a constraint. Records that were technically valid are no longer meaningful in the new model.</p><p>And now you are not just migrating data.</p><p>You are deciding what the data should mean.</p><p>That is the part of large migrations that deserves more attention.</p><p>The first instinct in a migration is usually to measure completeness.</p><p>Did we move every record?</p><p>Does the row count match?</p><p>Can we query the new database?</p><p>Those are necessary checks, but they are not enough.</p><p>A successful migration preserves meaning, not just records.</p><p>If an account had a particular state before the migration, that state should represent the same business reality afterward.</p><p>If a transaction had a relationship to another record, that relationship should still mean the same thing.</p><p>If an attribute was historically optional, making it mandatory in the target system may require more than filling in a default value.</p><p>The destination system forces you to confront assumptions that the source system allowed you to ignore.</p><p>That is where the migration becomes a data quality project.</p><p>You find one record that does not fit the new schema.</p><p>It is tempting to fix it.</p><p>Then you find another.</p><p>Then another hundred.</p><p>Eventually, someone asks the dangerous question:</p><p>Sometimes the answer is yes.</p><p>But every transformation introduces a decision.</p><p>Should an invalid value be mapped to a new value?</p><p>Should the record be rejected?</p><p>Should the field become null?</p><p>Should historical data be preserved even though the new system would never allow that state?</p><p>Should two values that were previously different now become one?</p><p>I remember reviewing a transformation rule I had written and realizing I could not explain, in business terms, why it was correct.</p><p>It satisfied the target schema.</p><p>But nobody had actually decided that it was the right interpretation.</p><p>That was an important distinction.</p><p>These are not technical questions alone. They are business semantics disguised as data transformations.</p><p>The migration code can tell you how to transform a value.</p><p>It cannot tell you whether that transformation is correct.</p><p>One of the biggest mistakes in migration planning is designing the transformation from the schema instead of the data.</p><p>The schema tells you what the system was supposed to contain.</p><p>Production data tells you what it actually contains.</p><p>Those can be very different things.</p><p>Before writing migration logic, profile the source data.</p><p>Look for unexpected nulls, multiple formats for the same field, deprecated values still in use, duplicate records, orphaned relationships, values outside expected ranges, records created by older versions of the application, and fields whose meaning changed over time.</p><p>The goal is not to make every anomaly disappear.</p><p>The goal is to understand the population before deciding what to do with it.</p><p>That changes the conversation from:</p><p>to:</p><p>The second statement is much more useful.</p><p>It gives the team something to reason about instead of something to clean up blindly.</p><p>Another lesson is to avoid hiding business decisions inside migration code.</p><p>A one line mapping in a transformation function can look harmless.</p><p>But if nobody can explain why the source value exists, or whether the target value means the same thing, the migration is quietly making a decision that nobody explicitly made.</p><p>I found it useful to think about the migration as two layers.</p><p>Migration mechanics</p><p>Read the source.</p><p>Transform the representation.</p><p>Write to the target.</p><p>Track failures.</p><p>Retry safely.</p><p>Migration decisions</p><p>What does each source value mean?</p><p>Which values are equivalent?</p><p>Which records require remediation?</p><p>Which historical states must be preserved?</p><p>Which records should not migrate at all?</p><p>Keeping those concerns separate makes the migration easier to reason about and easier to audit.</p><p>It also makes the next migration less painful because the decisions are documented somewhere other than inside the code.</p><p>This is where row counts become insufficient.</p><p>Imagine the source contains one million records and the target contains one million records.</p><p>That sounds successful.</p><p>Now imagine 20,000 relationships were broken during transformation.</p><p>Or 5,000 records were mapped into the wrong state.</p><p>Or a timestamp conversion shifted historical events by several hours.</p><p>The row count still matches.</p><p>The migration still failed.</p><p>Validation needs multiple dimensions.</p><p>Completeness</p><p>Did the expected records arrive?</p><p>Structural integrity</p><p>Do relationships, keys, and constraints remain valid?</p><p>Semantic fidelity</p><p>Does the migrated data still represent the same business meaning?</p><p>Behavioral equivalence</p><p>Does the application behave correctly when reading the migrated data?</p><p>Run real read paths against both systems for representative records and compare what the application would actually see.</p><p>This is the kind of check that can catch a timestamp problem even when every row is present and every field satisfies the schema.</p><p>Exception coverage</p><p>Do we understand every record that could not be migrated normally?</p><p>That last one is particularly important.</p><p>A migration that produces a clean success metric by silently dropping exceptional data is not a successful migration.</p><p>It is a successful concealment of a problem.</p><p>Large migrations benefit from treating validation as a first class output.</p><p>Instead of producing only:</p><p>produce evidence that explains the result.</p><p>For example:</p><p>Now the team has something it can investigate.</p><p>The migration becomes observable.</p><p>And when someone asks six months later, “Why were these 500 records handled differently?” there is an answer.</p><p>That is an underrated property of migration systems.</p><p>They should not only move data. They should explain what happened to the data.</p><p>For large systems, you rarely get confidence from manually inspecting records.</p><p>You get it from reconciliation.</p><p>After the data lands in the target, compare the old and new representations using business-meaningful checks.</p><p>Think of it as:</p><p>The important part is deciding what to reconcile.</p><p>Row counts are one signal.</p><p>Aggregates are another.</p><p>Business states, relationships, totals, event counts, and representative samples each reveal different classes of problems.</p><p>And when mismatches occur, preserve enough information to trace them back to the source record and the transformation that produced them.</p><p>Without that traceability, every mismatch becomes an investigation from scratch.</p><p>At scale, that gets expensive very quickly.</p><p>This may be the most important lesson.</p><p>A value can be valid according to the target schema and still be wrong.</p><p>Suppose the old system has three states:</p><p>The new system supports only:</p><p>A migration engineer might decide that UNKNOWN should become SUSPENDED.</p><p>Technically, the record now satisfies the target schema.</p><p>Semantically, something important may have been lost.</p><p>The right approach might be to preserve the unknown state elsewhere, remediate the source data before migration, or change the target model.</p><p>The schema cannot make that decision for you.</p><p>It only tells you what it will accept.</p><p>That is why migrations often expose deeper problems in the systems around them.</p><p>The database was not necessarily storing bad data.</p><p>It may have been storing data whose meaning was never explicitly modeled.</p><p>There is another temptation during a migration.</p><p>Clean everything.</p><p>If you are already touching millions of records, why not fix the historical inconsistencies too?</p><p>Sometimes that is the right thing to do.</p><p>Sometimes it is how a migration turns into an uncontrolled data transformation project.</p><p>Every additional cleanup rule is another place where the migration can silently change behavior.</p><p>A principle I now keep in mind is:</p><p>Migrate what you understand. Transform what you can justify.</p><p>Not every historical imperfection needs to be corrected during the migration.</p><p>Some should be documented.</p><p>Some should be remediated beforehand.</p><p>Some should be handled afterward.</p><p>And some are simply historical facts that should remain historical facts.</p><p>The goal is not to make the data look perfect.</p><p>The goal is to make the new system trustworthy.</p><p>Before a migration, it is easy to think of data as rows and columns.</p><p>During a migration, you start seeing something different.</p><p>Data carries history.</p><p>It carries assumptions.</p><p>It carries workarounds from previous application versions.</p><p>It carries relationships that may not be obvious from the schema.</p><p>It carries business decisions that may never have been documented.</p><p>Moving it safely means understanding all of that.</p><p>That is why I now think about large database migrations less as copying data and more as preserving meaning across a change in system boundaries.</p><p>The migration code is only one part of that problem.</p><p>The harder part is discovering what the old system actually means, determining what the new system can represent, identifying where those models disagree, and making those differences explicit.</p><p>Because when the source and target systems disagree about what is valid, the database is no longer the only thing being migrated.</p><p>The meaning of the data is being migrated too.</p><p>And that is the part you should plan for.</p><p>What about your last migration?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/every-large-migration-becomes-data-quality-project-plan-leela-kumili-djmpc/?trackingId=60GnM6vtScCr7UG6tYhhYw%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3163b609b1c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I Love When Product Teams Bring AI Prototypes]]></title>
            <link>https://medium.com/@leela.kumili/why-i-love-when-product-teams-bring-ai-prototypes-f237884dfd94?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/f237884dfd94</guid>
            <category><![CDATA[product-team]]></category>
            <category><![CDATA[cross-collaboration]]></category>
            <category><![CDATA[withai]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[new-way]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 22:16:01 GMT</pubDate>
            <atom:updated>2026-10-19T22:16:01.626Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZV4LhquiKZBR3W_H" /></figure><p>Recently, I attended a networking event where most people came from product and design backgrounds.</p><p>A question kept coming up:</p><p>“Aren’t engineers worried that we’re building prototypes with AI now?”</p><p>My answer surprised a few people.</p><p>No.</p><p>I love it.</p><p>The question makes sense though.</p><p>For years, software engineers were often the first people who could turn an idea into something real. A product idea would become a document, a user story, or a set of requirements. Then engineering would take that intent and transform it into working software.</p><p>Now product managers and designers can use AI tools to create functional prototypes in hours.</p><p>A flow that once required multiple meetings, design iterations, and engineering involvement can now exist as something clickable before a single engineer joins the conversation.</p><p>Some people see that as crossing boundaries.</p><p>I see it as removing friction.</p><p>Because a prototype is not software.</p><p>A prototype is understanding.</p><p>For most of my career, intent arrived as words.</p><p>A requirements document.</p><p>A user story.</p><p>A design file.</p><p>A product brief.</p><p>I would read it and build a mental model of what the experience should be. Other engineers would do the same.</p><p>The problem?</p><p>We were all solving slightly different versions of the same problem.</p><p>A product manager had one vision.</p><p>A designer had another interpretation.</p><p>An engineer had a third mental model.</p><p>Nobody was wrong.</p><p>We were just translating ideas through different lenses.</p><p>A prototype changes that.</p><p>Instead of reading a sentence like:</p><p>“Users should be able to quickly discover relevant content based on their preferences.”</p><p>I can interact with the experience.</p><p>I can click through the flow.</p><p>I can see where the user gets confused.</p><p>I can understand what “quickly” actually means.</p><p>I can ask better technical questions earlier.</p><p>That is incredibly valuable.</p><p>Engineering has spent years trying to move important activities left.</p><p>Testing earlier.</p><p>Security earlier.</p><p>Performance discussions earlier.</p><p>AI prototypes move something even more expensive left:</p><p>misalignment.</p><p>A prototype creates a shared conversation.</p><p>Instead of debating what a requirement means, the team can discuss what everyone can see.</p><p>That changes the quality of the conversation.</p><p>A prototype showing live search filtering might look simple from a product perspective.</p><p>But an engineer immediately starts thinking:</p><p>Is this client side filtering over a small dataset?</p><p>Does this require real time queries?</p><p>How frequently will data change?</p><p>What happens when millions of users perform this action simultaneously?</p><p>Does the current architecture support this?</p><p>Those questions are much easier to answer when there is something tangible to react to.</p><p>A year ago, that conversation might have happened halfway through implementation.</p><p>Now it can happen before the first commit.</p><p>That timing difference is more valuable than writing code faster.</p><p>The best teams are not the ones where product, design, and engineering have clearly separated responsibilities.</p><p>They are the ones where everyone understands the problem deeply.</p><p>A prototype becomes a shared artifact.</p><p>Product uses it to validate the workflow.</p><p>Design uses it to validate the experience.</p><p>Engineering uses it to validate feasibility.</p><p>Everyone starts from the same picture instead of interpreting the same paragraph differently.</p><p>I think of prototypes as the first artifact that truly belongs to the entire team.</p><p>Requirements usually belong to product.</p><p>Mockups usually belong to design.</p><p>Code belongs to engineering.</p><p>A prototype belongs to the conversation.</p><p>There is one important distinction.</p><p>A prototype shows what an experience should feel like.</p><p>It does not show what the system needs to survive.</p><p>A demo does not handle:</p><p>* failure scenarios * scale requirements * security concerns * data consistency challenges * operational complexity * unexpected user behavior</p><p>And that is completely fine.</p><p>A prototype was never meant to solve those problems.</p><p>The mistake is not creating prototypes.</p><p>The mistake is treating a prototype as an architecture decision.</p><p>A prototype is a question.</p><p>Engineering’s job is to figure out the best answer.</p><p>AI is changing who can create.</p><p>That is a good thing.</p><p>The more people who can explore ideas, test assumptions, and create shared understanding, the better the final product becomes.</p><p>The role of engineering was never just converting ideas into code.</p><p>The value of engineering is understanding complexity, making tradeoffs, and building systems that survive reality.</p><p>The best prototypes do not replace engineering.</p><p>They make engineering conversations better.</p><p>So no, I am not worried when product teams bring AI prototypes.</p><p>Bring them earlier.</p><p>Bring them often.</p><p>The earlier we can react to something real, the better software we build.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/why-i-love-when-product-teams-bring-ai-prototypes-leela-kumili-wed9c/?trackingId=Gnz%2BSG3BTye18PNV5fC%2Flg%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f237884dfd94" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When to Protect the Team and When to Let Them Feel the Heat]]></title>
            <link>https://medium.com/@leela.kumili/when-to-protect-the-team-and-when-to-let-them-feel-the-heat-cdcc67ea67aa?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/cdcc67ea67aa</guid>
            <category><![CDATA[engineering-team]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[waystolead]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 22:16:01 GMT</pubDate>
            <atom:updated>2026-10-12T22:16:01.695Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IRYCASpnFRvmI-TV" /></figure><p>Recently, I was reminded of a valuable leadership lesson by watching the leaders I admire.</p><p>Not from what they said, but from what they did when things got difficult.</p><p>Someone is blocked.</p><p>A production release is held.</p><p>A cross-team dependency isn’t moving.</p><p>A runbook doesn’t work when you actually need it.</p><p>A mistake happens.</p><p>What I’ve noticed is that the leaders I admire don’t always step in.</p><p>And they don’t always step back either.</p><p>They seem to know something much harder to learn:</p><p>When to protect the team, and when to let the team feel the heat.</p><p>I’ve seen a few examples recently that made this distinction much clearer to me.</p><p>In one situation, a Lead Engineer was driving an important cross team effort.</p><p>The work required collaboration with another team, but progress had started to stall. The Lead had competing priorities and couldn’t give the effort the time it needed.</p><p>An Engineering Manager stepped in.</p><p>Not to take the technical ownership away from the Lead.</p><p>Not to start making technical decisions.</p><p>The EM stepped in to drive the cross team conversation, create alignment, and keep the dependency moving.</p><p>That distinction mattered.</p><p>The problem wasn’t that the Lead Engineer couldn’t solve it.</p><p>The problem was that the situation had crossed a boundary where leadership involvement could remove organizational friction.</p><p>The EM recognized that and stepped in.</p><p>That’s one form of protecting a team.</p><p>You don’t make engineers fight organizational battles that leadership is better positioned to resolve.</p><p>I saw another example that stayed with me.</p><p>An engineer was working on a known issue when the change introduced an unintended side effect, and a production release had to be held.</p><p>The engineer who made the change then had to step away unexpectedly for an extended period.</p><p>The team already had several competing priorities. Waiting for that engineer to return wasn’t a realistic option, and the team still had to figure out the safest path forward.</p><p>The Engineering Manager stepped in and took ownership.</p><p>Not because the engineer had done something wrong.</p><p>Not because the manager needed to take over the technical work.</p><p>Because the team was already stretched, and someone needed to own the situation.</p><p>The release was held while the team worked through the safest path forward.</p><p>The decision wasn’t simple. Reverting carried significant cost, while moving forward safely required careful work.</p><p>The manager didn’t try to hide what happened.</p><p>They didn’t blame the engineer who made the change.</p><p>They didn’t leave an already overwhelmed team to figure everything out on its own.</p><p>They took ownership of the situation, created space for the team to work through it, and made sure the problem had an owner even when the original engineer wasn’t available.</p><p>That is a different kind of protection.</p><p>Protecting the team does not mean protecting the team from accountability.</p><p>It means making sure accountability doesn’t turn into blame.</p><p>The team still needs to understand what happened.</p><p>The engineer still needs to learn from the experience.</p><p>But the leader owns the responsibility of making sure the team can move forward when circumstances change.</p><p>This is the example that changed my thinking the most.</p><p>There was a situation where our team had to respond to an operational problem.</p><p>The instinct might have been for a leader to step in and take over.</p><p>Instead, the leader stayed close, contributed where needed, and let the team work through the problem.</p><p>Not because they didn’t care.</p><p>Because they knew we had never really exercised the process before.</p><p>We had a runbook.</p><p>On paper, it looked reasonable.</p><p>In practice, following it exposed problems we hadn’t anticipated.</p><p>Some steps were unclear.</p><p>Some dependencies weren’t obvious.</p><p>Some actions took longer than expected.</p><p>We discovered that having a runbook is very different from knowing how to execute it under pressure.</p><p>The leader didn’t immediately rescue us.</p><p>They helped us work through it without taking the problem away from us.</p><p>We felt the difficulty.</p><p>And that was the point.</p><p>We came out of the exercise with a much better understanding of what needed to change.</p><p>The runbook improved.</p><p>The team’s confidence improved.</p><p>And the next time we faced a similar situation, we were better prepared.</p><p>The leader could have made the immediate situation easier.</p><p>Instead, they made the team stronger.</p><p>As a Lead Engineer, I understand the temptation to step in.</p><p>You see someone struggling with a design and already know how you would approach it.</p><p>You see a difficult conversation and know you could probably handle it faster.</p><p>You see an incident going sideways and think:</p><p>I can fix this.</p><p>I know the system.</p><p>I know the history.</p><p>I have probably seen a similar problem before.</p><p>And when the team is under pressure, speed feels valuable.</p><p>Sometimes you should step in.</p><p>But sometimes the fastest way to solve today’s problem creates a bigger problem tomorrow.</p><p>If I always answer the question, people stop developing the confidence to answer it themselves.</p><p>If I always take over the difficult conversation, someone else never learns to have it.</p><p>If I always jump into an incident, the team never learns where its process breaks down.</p><p>Leadership isn’t measured only by how many problems I personally solve.</p><p>It is also measured by what the team can solve without me.</p><p>That means sometimes accepting a slower path today because it creates a stronger engineer tomorrow.</p><p>There is a difference between helping someone succeed and making sure they never struggle.</p><p>Those are not the same thing.</p><p>And I’ve confused them before.</p><p>I’ve been on both sides of this call, and I haven’t always read it right.</p><p>Someone I worked with regularly brought me a design problem.</p><p>I could see the answer.</p><p>It would have taken me ten minutes to explain it, and they would have left the conversation unblocked.</p><p>I did that.</p><p>It felt like help.</p><p>What I actually did was take the interesting part of their problem and keep it.</p><p>The next time something similar came up, they brought it to me again.</p><p>That was the signal.</p><p>I hadn’t made them more capable.</p><p>I had made myself necessary.</p><p>That changed how I approach those conversations.</p><p>Now I try to resist the urge to immediately give the answer.</p><p>I ask what they think is happening.</p><p>What options have they considered?</p><p>What would they recommend?</p><p>What tradeoff are they worried about?</p><p>Sometimes I still give the answer.</p><p>But not before giving them the opportunity to work through the problem themselves.</p><p>It is genuinely harder to watch someone take longer than I would have.</p><p>But when they hit that class of problem again and don’t need me, that’s the kind of help that compounds.</p><p>The leaders I’ve learned the most from seem to distinguish between two kinds of pressure.</p><p>Some pressure is unnecessary.</p><p>Organizational dysfunction.</p><p>Unclear expectations.</p><p>Cross team friction.</p><p>Blame.</p><p>Unreasonable demands.</p><p>Those are things leaders should absorb or resolve rather than pushing them down to the team.</p><p>Other pressure is useful.</p><p>Owning a difficult decision.</p><p>Defending a design.</p><p>Handling an incident.</p><p>Having a difficult conversation.</p><p>Working through an unfamiliar process.</p><p>Learning from an experience that didn’t go as expected.</p><p>Those experiences can be uncomfortable without being harmful.</p><p>And sometimes people need the opportunity to experience them.</p><p>When I feel the urge to step in, I try to pause.</p><p>Is this something the team genuinely needs me to own?</p><p>Or am I taking it because I know I can do it faster?</p><p>Is the pressure necessary, or is it noise?</p><p>Does this person have enough context and authority to handle it?</p><p>What happens if I solve this for them?</p><p>And the question that matters most:</p><p>Am I protecting this person, or am I protecting myself from watching them struggle?</p><p>That last one is uncomfortable.</p><p>Sometimes I’m not stepping in because the situation is truly too difficult.</p><p>I’m stepping in because watching someone struggle is uncomfortable for me.</p><p>But struggle is not always failure.</p><p>Sometimes it is learning in progress.</p><p>I used to think strong leadership meant creating an environment where the team could focus without unnecessary distractions.</p><p>I still believe that matters.</p><p>But watching these leaders, and making a few mistakes of my own, has added another dimension for me.</p><p>Good leaders don’t remove all the heat.</p><p>They make sure the heat is worth experiencing.</p><p>They protect the team from things that shouldn’t reach them.</p><p>And they allow the team to experience things that will make them stronger.</p><p>That might mean stepping in to resolve a cross-team dependency.</p><p>It might mean taking accountability when something goes wrong.</p><p>Or it might mean standing beside the team while they work through a process they’ve never practiced before.</p><p>The goal isn’t to become the person who fixes everything.</p><p>The goal is to build a team that can handle more without you.</p><p>That’s the leadership judgment I’m still learning.</p><p>Protect people from unnecessary pressure. Do not protect them from necessary growth.</p><p>When have you stepped in and later realized the team needed to feel that one?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/when-protect-team-let-them-feel-heat-leela-kumili-s790c/?trackingId=ZpbDZO8LS%2BGe4GF%2FUQVifw%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cdcc67ea67aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The New Engineering Bottleneck Isn’t Coding. It’s Deciding.]]></title>
            <link>https://medium.com/@leela.kumili/the-new-engineering-bottleneck-isnt-coding-it-s-deciding-80670dc544a3?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/80670dc544a3</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[neway]]></category>
            <category><![CDATA[withai]]></category>
            <category><![CDATA[choose]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 22:16:01 GMT</pubDate>
            <atom:updated>2026-10-05T22:16:01.318Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yZezZpsfk3xpW_hK" /></figure><p>For most of my career, engineering constraints were familiar.</p><p>There was never enough time.</p><p>Not enough engineers.</p><p>Too many bugs.</p><p>Too many requirements.</p><p>Too many systems that needed to be built, maintained, and improved.</p><p>The bottleneck was execution.</p><p>Could we design it fast enough? Could we implement it safely? Could we ship it before the next priority arrived?</p><p>AI is changing that equation.</p><p>Today, I can prototype an idea in an afternoon. Generate a first version of an architecture diagram. Explore implementation approaches. Create documentation. Build a proof of concept.</p><p>The cost of starting has dropped dramatically.</p><p>But something unexpected happened.</p><p>The bottleneck did not disappear.</p><p>It moved.</p><p>The hardest engineering question is no longer:</p><p>“Can we build this?”</p><p>It is:</p><p>“Should we build this?”</p><p>For years, engineering effort acted as a natural prioritization mechanism.</p><p>An idea needed to be valuable enough to justify the investment.</p><p>A new service meant weeks of design and implementation.</p><p>A new tool meant convincing a team that the return justified the maintenance cost.</p><p>A prototype required enough confidence before anyone spent serious time on it.</p><p>That constraint was frustrating, but it served a purpose.</p><p>It filtered.</p><p>Weak ideas rarely survived the cost of execution.</p><p>Not because someone explicitly rejected them, but because the effort required to build them forced a decision.</p><p>AI changed this.</p><p>The cost of exploring ideas has collapsed.</p><p>A developer can now validate an idea before the first meeting ends. A team can experiment with multiple approaches instead of debating one for weeks.</p><p>The old filter is gone.</p><p>And now engineers have to create a new one.</p><p>The interesting thing about AI is that it does not create an idea shortage.</p><p>It creates an abundance problem.</p><p>Every week there is something new.</p><p>A new AI framework.</p><p>A new workflow.</p><p>A new internal automation opportunity.</p><p>A new way to improve developer productivity.</p><p>A new side project idea that suddenly feels possible.</p><p>The challenge is not identifying opportunities.</p><p>The challenge is choosing between them.</p><p>I experienced this recently after attending a technical event.</p><p>I came home with several ideas that I genuinely believed could create value.</p><p>A year ago, the decision would have been easier.</p><p>I would have looked at the effort required, my available time, and naturally narrowed the list.</p><p>Now the barrier to experimentation is much lower.</p><p>I could build a prototype for any of them.</p><p>The constraint was no longer capability.</p><p>The constraint was attention.</p><p>This is the fundamental asymmetry of the AI era.</p><p>AI multiplied what we can create.</p><p>It did not multiply our time.</p><p>We still have the same twenty four hours.</p><p>The same meetings.</p><p>The same responsibilities.</p><p>The same cognitive limits.</p><p>When possibilities increase but capacity stays fixed, prioritization becomes the most valuable skill.</p><p>This is why prioritization is not just a product management skill anymore.</p><p>It is an engineering skill.</p><p>Senior engineers have always made decisions about:</p><p>Which architecture?</p><p>Which technology?</p><p>Which tradeoff?</p><p>Which risk?</p><p>AI adds new questions:</p><p>Which idea?</p><p>Which experiment?</p><p>Which problem?</p><p>Which opportunity?</p><p>As a Lead Engineer, my calendar tells a different story than the AI productivity conversation.</p><p>It is filled with:</p><p>Design reviews.</p><p>Architecture discussions.</p><p>Coaching engineers.</p><p>Partnering with product teams.</p><p>Resolving technical tradeoffs.</p><p>Helping teams align on priorities.</p><p>None of these disappear because AI can generate code faster.</p><p>If anything, they become more important.</p><p>Because look closely at what these activities require.</p><p>They require judgment.</p><p>A design review is not valuable because someone cannot write code.</p><p>It is valuable because someone needs to decide what risks matter.</p><p>A roadmap discussion is not valuable because teams cannot build features.</p><p>It is valuable because teams need to decide where limited engineering capacity should go.</p><p>Coaching is not about teaching someone how to type code faster.</p><p>It is helping someone develop better judgment about what problems are worth solving.</p><p>I do not have a perfect framework.</p><p>But three questions have helped me think differently about opportunities.</p><p>Will this create knowledge, capability, or leverage that helps me or my team in the future?</p><p>A small investment that improves future decisions is often more valuable than a one time impressive prototype.</p><p>A good idea does not automatically mean I should pursue it.</p><p>Sometimes the highest impact decision is enabling someone else to build it.</p><p>Every yes has a hidden no.</p><p>If I cannot explain what I am choosing not to do, I probably have not made a real decision.</p><p>I have just added another commitment.</p><p>That question has become increasingly important.</p><p>Because when building becomes cheaper, saying no becomes harder.</p><p>And more valuable.</p><p>For a long time, engineering excellence was associated with execution.</p><p>Writing better code.</p><p>Building scalable systems.</p><p>Solving harder technical problems.</p><p>Those skills still matter.</p><p>But AI is changing where the leverage comes from.</p><p>The engineers who create the most impact will not necessarily be the ones who can generate the most code.</p><p>They will be the ones who can identify the right problems.</p><p>The ones who can separate interesting ideas from important ones.</p><p>The ones who can make decisions when every option looks possible.</p><p>AI can help us build almost anything.</p><p>The harder question is deciding what deserves to exist.</p><p>The new engineering bottleneck isn’t coding.</p><p>It’s deciding.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/new-engineering-bottleneck-isnt-coding-its-deciding-leela-kumili-g2nrc/?trackingId=qckeRiLXRS6nottGlKV50Q%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80670dc544a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Built a Tool That Worked. Nobody Else Could Run It.]]></title>
            <link>https://medium.com/@leela.kumili/i-built-a-tool-that-worked-nobody-else-could-run-it-f9fd4295101e?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/f9fd4295101e</guid>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 20:21:01 GMT</pubDate>
            <atom:updated>2026-09-21T20:21:01.588Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0e9rJ5Bx1Ih1ZLIZ" /></figure><p>I built a tool that worked.</p><p>We were doing a data intensive migration of a large platform from one database to another.</p><p>Our regular regression and test pipeline, already part of CI/CD, could validate that the system behaved as expected after the migration.</p><p>But it did not tell us what was happening during the migration.</p><p>What happens when traffic is split between the old and new databases?</p><p>What happens as different services move through different migration phases?</p><p>Are the schemas and data mappings actually aligned?</p><p>Are there unexpected differences in data shape across environments?</p><p>Are APIs behaving differently during the transition?</p><p>And with the volume of data involved, how do you find edge cases that manual testing simply cannot cover?</p><p>So I built a tool to help answer those questions.</p><p>It could inspect a cluster, read live configuration, connect to both the old and new databases, compare schemas and data mappings, search logs and events, validate API and database behavior, and run checks across different phases of the migration.</p><p>It gave us visibility into parts of the migration that our regular regression pipeline could not.</p><p>It also caught edge cases, data mismatches, and unexpected impacts that would have been difficult to find through manual testing alone, especially given the volume of data and differences in data shape across environments.</p><p>All of it worked.</p><p>Then I tried to imagine another engineer using it.</p><p>That is when I realized I had built the wrong thing.</p><p>Not because the tool was technically wrong.</p><p>Because I had optimized for capability instead of adoption.</p><p>Our existing regression pipeline answered an important question:</p><p>Does the system work as expected after the migration?</p><p>But the migration itself introduced another set of questions:</p><p>What is happening while we are getting there?</p><p>That distinction mattered.</p><p>A migration is not a single event where you flip a switch from old to new.</p><p>It is a process.</p><p>Different services can be at different stages. Data can exist in different states. Configuration can change. Traffic patterns can expose problems that are invisible in a smaller test environment.</p><p>That meant we needed a different kind of validation.</p><p>Not just validation of the destination, but visibility into the journey.</p><p>The tool became an assistant around that workflow.</p><p>The goal was modest: make the migration easier to understand and easier to validate.</p><p>At some point, I had something that felt genuinely useful.</p><p>Then came the question I had not asked early enough:</p><p>Could another engineer use this tomorrow?</p><p>The honest answer was no.</p><p>The tool ran on my machine.</p><p>It assumed my environment.</p><p>It depended on how I had configured things.</p><p>The interface was built around my own editor setup, cheatsheets, and workflow rather than the developer experience the rest of the team already knows.</p><p>Someone else would have to work through several pieces of setup before reaching the part that was actually useful.</p><p>They would need to understand assumptions that were obvious to me because I had built them.</p><p>That was the moment the problem became obvious.</p><p>I had built a tool that knew how to solve the problem.</p><p>I had not built a path for anyone else to use the solution.</p><p>That distinction sounds small.</p><p>It is not.</p><p>When we build something for ourselves, we naturally optimize for the shortest path between problem and working solution.</p><p>That is often exactly right.</p><p>You have a problem. You need to test an idea. You build the smallest thing that proves it can work.</p><p>But a second problem shows up later, and it is a different kind of engineering.</p><p>How does this capability reach someone who was not there when it was built?</p><p>Setup has to be predictable.</p><p>Configuration has to be legible.</p><p>Dependencies have to be available.</p><p>Authentication has to work in someone else’s environment.</p><p>The workflow needs documentation.</p><p>Errors need to explain themselves.</p><p>Someone needs to know what to do when the happy path stops being happy.</p><p>And ideally, they should not have to find the person who built it on Slack.</p><p>None of that makes the underlying logic smarter.</p><p>It makes the capability usable.</p><p>“Works on my machine” is usually treated as a developer joke.</p><p>Inside an organization, it hides something more expensive.</p><p>If a tool depends on one person’s environment, that person becomes part of the architecture.</p><p>They become the documentation.</p><p>They become the support channel.</p><p>They become the one who knows which configuration matters and the one everyone asks when something breaks.</p><p>The tool that was supposed to reduce engineering effort has created a new dependency instead.</p><p>The irony is not lost on me.</p><p>I built the thing to remove friction from a migration.</p><p>If nobody else can operate it without me, I have not removed the bottleneck.</p><p>I have relocated it, and given it my name.</p><p>This feels especially relevant now because AI has made prototypes dramatically cheaper.</p><p>An engineer can go from idea to working tool in an afternoon.</p><p>That is exciting.</p><p>It also creates a new temptation.</p><p>We can confuse speed of creation with value delivered.</p><p>A working prototype feels like the finish line because the hard technical problem is solved.</p><p>But inside a team, much of the value lives in what happens after that.</p><p>Can someone else discover it?</p><p>Understand it?</p><p>Run it?</p><p>Trust it?</p><p>Extend it?</p><p>Use it without the person who built it?</p><p>The faster we can create software, the easier it becomes to build things that are technically impressive and organizationally isolated.</p><p>I am not arguing that every prototype should be productionized.</p><p>That is its own trap.</p><p>Not every internal experiment deserves a platform, a pipeline, a polished UI, and months of operational investment.</p><p>Sometimes a local tool is exactly the right answer.</p><p>The important part is knowing which one you built.</p><p>For a prototype, the question is:</p><p>Can I make this work?</p><p>That is enough.</p><p>You are testing an idea and reducing uncertainty.</p><p>For a shared engineering capability, there is a second question:</p><p>Can someone else use this successfully without me?</p><p>Those are not the same question.</p><p>The first proves capability.</p><p>The second creates leverage.</p><p>Knowing which one you are answering is engineering judgment, not project management.</p><p>I still like building fast.</p><p>I still think prototypes are allowed to be messy.</p><p>I still do not believe every useful script needs to become a platform.</p><p>But when something starts solving a problem other engineers also have, I run a different set of questions.</p><p>What would someone need to use this for the first time?</p><p>Which assumptions exist only because I am the one who built it?</p><p>What happens on a different machine?</p><p>Where will they get stuck?</p><p>Can they recover when something fails?</p><p>And the one that has turned out to be most useful:</p><p>What would it take for me to stop being part of the workflow?</p><p>That question changes how I think about the work.</p><p>The goal is no longer just to make the tool capable.</p><p>It is to make the capability independent of its creator.</p><p>The clearest sign that an internal tool has matured is not that it has more features.</p><p>It is that the person who built it is no longer required.</p><p>A tool is not finished when it works.</p><p>It is finished when someone else can use it successfully without you.</p><p>Otherwise, it is not a shared capability at all.</p><p>It is a very sophisticated personal script.</p><p>What is one tool on your team that only one person can actually run?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/i-built-tool-worked-nobody-else-could-run-leela-kumili-tmqrc/?trackingId=%2FQQc%2FYTmS%2FCd%2BEpQeoxGZw%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f9fd4295101e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The New Engineering Bottleneck Isn’t Coding. It’s Deciding.]]></title>
            <link>https://medium.com/@leela.kumili/the-new-engineering-bottleneck-isnt-coding-it-s-deciding-a34666b73fb0?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/a34666b73fb0</guid>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 17:21:01 GMT</pubDate>
            <atom:updated>2026-09-18T17:21:01.385Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QJflLbQA6O8SosEf" /></figure><p>For most of my career, engineering constraints were familiar.</p><p>There was never enough time.</p><p>Not enough engineers.</p><p>Too many bugs.</p><p>Too many requirements.</p><p>Too many systems that needed to be built, maintained, and improved.</p><p>The bottleneck was execution.</p><p>Could we design it fast enough? Could we implement it safely? Could we ship it before the next priority arrived?</p><p>AI is changing that equation.</p><p>Today, I can prototype an idea in an afternoon. Generate a first version of an architecture diagram. Explore implementation approaches. Create documentation. Build a proof of concept.</p><p>The cost of starting has dropped dramatically.</p><p>But something unexpected happened.</p><p>The bottleneck did not disappear.</p><p>It moved.</p><p>The hardest engineering question is no longer:</p><p>“Can we build this?”</p><p>It is:</p><p>“Should we build this?”</p><p>For years, engineering effort acted as a natural prioritization mechanism.</p><p>An idea needed to be valuable enough to justify the investment.</p><p>A new service meant weeks of design and implementation.</p><p>A new tool meant convincing a team that the return justified the maintenance cost.</p><p>A prototype required enough confidence before anyone spent serious time on it.</p><p>That constraint was frustrating, but it served a purpose.</p><p>It filtered.</p><p>Weak ideas rarely survived the cost of execution.</p><p>Not because someone explicitly rejected them, but because the effort required to build them forced a decision.</p><p>AI changed this.</p><p>The cost of exploring ideas has collapsed.</p><p>A developer can now validate an idea before the first meeting ends. A team can experiment with multiple approaches instead of debating one for weeks.</p><p>The old filter is gone.</p><p>And now engineers have to create a new one.</p><p>The interesting thing about AI is that it does not create an idea shortage.</p><p>It creates an abundance problem.</p><p>Every week there is something new.</p><p>A new AI framework.</p><p>A new workflow.</p><p>A new internal automation opportunity.</p><p>A new way to improve developer productivity.</p><p>A new side project idea that suddenly feels possible.</p><p>The challenge is not identifying opportunities.</p><p>The challenge is choosing between them.</p><p>I experienced this recently after attending a technical event.</p><p>I came home with several ideas that I genuinely believed could create value.</p><p>A year ago, the decision would have been easier.</p><p>I would have looked at the effort required, my available time, and naturally narrowed the list.</p><p>Now the barrier to experimentation is much lower.</p><p>I could build a prototype for any of them.</p><p>The constraint was no longer capability.</p><p>The constraint was attention.</p><p>This is the fundamental asymmetry of the AI era.</p><p>AI multiplied what we can create.</p><p>It did not multiply our time.</p><p>We still have the same twenty four hours.</p><p>The same meetings.</p><p>The same responsibilities.</p><p>The same cognitive limits.</p><p>When possibilities increase but capacity stays fixed, prioritization becomes the most valuable skill.</p><p>This is why prioritization is not just a product management skill anymore.</p><p>It is an engineering skill.</p><p>Senior engineers have always made decisions about:</p><p>Which architecture?</p><p>Which technology?</p><p>Which tradeoff?</p><p>Which risk?</p><p>AI adds new questions:</p><p>Which idea?</p><p>Which experiment?</p><p>Which problem?</p><p>Which opportunity?</p><p>As a Lead Engineer, my calendar tells a different story than the AI productivity conversation.</p><p>It is filled with:</p><p>Design reviews.</p><p>Architecture discussions.</p><p>Coaching engineers.</p><p>Partnering with product teams.</p><p>Resolving technical tradeoffs.</p><p>Helping teams align on priorities.</p><p>None of these disappear because AI can generate code faster.</p><p>If anything, they become more important.</p><p>Because look closely at what these activities require.</p><p>They require judgment.</p><p>A design review is not valuable because someone cannot write code.</p><p>It is valuable because someone needs to decide what risks matter.</p><p>A roadmap discussion is not valuable because teams cannot build features.</p><p>It is valuable because teams need to decide where limited engineering capacity should go.</p><p>Coaching is not about teaching someone how to type code faster.</p><p>It is helping someone develop better judgment about what problems are worth solving.</p><p>I do not have a perfect framework.</p><p>But three questions have helped me think differently about opportunities.</p><p>Will this create knowledge, capability, or leverage that helps me or my team in the future?</p><p>A small investment that improves future decisions is often more valuable than a one time impressive prototype.</p><p>A good idea does not automatically mean I should pursue it.</p><p>Sometimes the highest impact decision is enabling someone else to build it.</p><p>Every yes has a hidden no.</p><p>If I cannot explain what I am choosing not to do, I probably have not made a real decision.</p><p>I have just added another commitment.</p><p>That question has become increasingly important.</p><p>Because when building becomes cheaper, saying no becomes harder.</p><p>And more valuable.</p><p>For a long time, engineering excellence was associated with execution.</p><p>Writing better code.</p><p>Building scalable systems.</p><p>Solving harder technical problems.</p><p>Those skills still matter.</p><p>But AI is changing where the leverage comes from.</p><p>The engineers who create the most impact will not necessarily be the ones who can generate the most code.</p><p>They will be the ones who can identify the right problems.</p><p>The ones who can separate interesting ideas from important ones.</p><p>The ones who can make decisions when every option looks possible.</p><p>AI can help us build almost anything.</p><p>The harder question is deciding what deserves to exist.</p><p>The new engineering bottleneck isn’t coding.</p><p>It’s deciding.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/new-engineering-bottleneck-isnt-coding-its-deciding-leela-kumili-g2nrc/?trackingId=epuqPwQ2TqyVwc%2BmSc8Wng%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a34666b73fb0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ownership Isn’t About Writing Every Line Anymore]]></title>
            <link>https://medium.com/@leela.kumili/ownership-isnt-about-writing-every-line-anymore-f003c2c03615?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/f003c2c03615</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ownership]]></category>
            <category><![CDATA[career-development]]></category>
            <category><![CDATA[promotion]]></category>
            <category><![CDATA[growth]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Wed, 09 Sep 2026 21:51:01 GMT</pubDate>
            <atom:updated>2026-09-09T21:51:01.572Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_zMgLmDsB2H6lD-t" /></figure><p>“You are responsible for anything your AI generates.”</p><p>I keep hearing this advice everywhere.</p><p>At conferences. In engineering blogs. During AI adoption conversations inside companies.</p><p>And I agree with it.</p><p>If an AI agent generates code that reaches production, the engineer who approves that change owns the outcome.</p><p>You cannot say:</p><p>“The agent wrote it.”</p><p>You cannot say:</p><p>“I didn’t notice.”</p><p>You cannot say:</p><p>“The model made the mistake.”</p><p>Production systems do not care who typed the code.</p><p>But after spending more time working with AI coding agents, I keep coming back to a different question.</p><p>What does ownership actually look like now?</p><p>Because the scale has changed.</p><p>The expectation has not.</p><p>For years, ownership and implementation were closely connected.</p><p>I would spend days building a feature. I knew every file I touched. I understood every design decision because I had made those decisions while writing the code.</p><p>The code review process naturally reinforced that understanding.</p><p>Writing the code was part of reviewing the code.</p><p>AI changes that relationship.</p><p>Today, an AI agent can generate an entire feature in minutes.</p><p>It can update API contracts.</p><p>Modify database models.</p><p>Refactor multiple modules.</p><p>Add tests.</p><p>Update documentation.</p><p>Create configuration changes.</p><p>Before I finish carefully reviewing the first few files, the agent may have already changed twenty more.</p><p>The amount of generated code has increased dramatically.</p><p>But human understanding has not accelerated at the same pace.</p><p>And that creates a new challenge.</p><p>For decades, software engineering had a simple equation:</p><p>More engineers + more time = more code.</p><p>AI changed that equation.</p><p>Now the problem is rarely “Can we write this?”</p><p>The question becomes:</p><p>Can we build enough confidence that this should exist?</p><p>The bottleneck moved from implementation to validation.</p><p>From producing software to trusting software.</p><p>And that requires a different definition of ownership.</p><p>I don’t believe ownership means manually reading every single line because we feel obligated to.</p><p>That approach does not scale.</p><p>And pretending it does creates a false sense of safety.</p><p>Ownership is not about knowing every character in a diff.</p><p>It is about understanding the decisions that matter.</p><p>Can I explain why this solution exists?</p><p>Does it fit our architecture?</p><p>Does it introduce operational risk?</p><p>What assumptions is the AI making?</p><p>What happens when traffic increases ten times?</p><p>How will another team depend on this?</p><p>Can someone maintain this six months from now?</p><p>Those questions determine whether software succeeds in production.</p><p>Not whether every variable was named perfectly.</p><p>Not whether every implementation detail was personally typed.</p><p>The difference is subtle but important.</p><p>The more I use AI tools, the less time I spend thinking about individual implementation choices.</p><p>Instead, I spend more time validating intent.</p><p>Are we solving the right problem?</p><p>Is this the simplest solution?</p><p>Does this move the platform in the right direction?</p><p>Are we introducing complexity we will regret later?</p><p>These are not coding questions.</p><p>They are engineering judgment questions.</p><p>And that judgment becomes more valuable when implementation becomes cheaper.</p><p>AI can generate patterns.</p><p>It cannot understand your organization’s history.</p><p>It cannot know why a previous team rejected a similar design three years ago.</p><p>It cannot fully understand the tradeoffs hidden inside architecture decisions.</p><p>That context still belongs to engineers.</p><p>If ownership is no longer about reading every line, where does confidence come from?</p><p>Not from reviewing harder.</p><p>From building better systems around development.</p><p>Strong automated tests that capture the behaviors we care about.</p><p>Clear architectural boundaries that limit blast radius.</p><p>Observability that quickly exposes when reality differs from expectations.</p><p>Good engineering practices were never just about quality.</p><p>They were mechanisms for creating confidence.</p><p>The engineers who succeed with AI will not be the ones who can read the most generated code.</p><p>They will be the ones who build environments where generated code can be trusted.</p><p>Imagine an AI agent creates a pull request.</p><p>Forty files changed.</p><p>Thousands of lines added.</p><p>Tests pass.</p><p>Documentation updated.</p><p>What do you do?</p><p>Do you read every line?</p><p>If yes, are you still getting the productivity advantage AI promised?</p><p>If no, what exactly are you approving?</p><p>I don’t think our industry has fully answered this yet.</p><p>And maybe that is the most interesting part of this transition.</p><p>We are not just adopting a new coding tool.</p><p>We are redefining what it means to be responsible for software.</p><p>Maybe AI is forcing us to recognize something that was true all along.</p><p>Ownership was never about who wrote the code.</p><p>It was about who made the decision to ship it.</p><p>Who understood the tradeoffs.</p><p>Who considered the risks.</p><p>Who was willing to stand behind the outcome.</p><p>AI simply removes the illusion that implementation and understanding are the same thing.</p><p>As generating software becomes easier, engineering judgment becomes the scarce resource.</p><p>The engineer of the future will not be the person who creates the most code.</p><p>They will be the person who creates the most confidence.</p><p>When an AI agent hands you a forty file pull request, what do you actually do today: read every line, trust the tests, or build confidence another way?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/ownership-isnt-writing-every-line-anymore-leela-kumili-d6kwc/?trackingId=nOLsdqkRQO6tAJYxoIFt%2Fw%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f003c2c03615" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Find a Mentor Without Asking “Will You Be My Mentor?”]]></title>
            <link>https://medium.com/@leela.kumili/how-to-find-a-mentor-without-asking-will-you-be-my-mentor-5947fc1137d9?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/5947fc1137d9</guid>
            <category><![CDATA[mentorship]]></category>
            <category><![CDATA[mentor]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 20:48:02 GMT</pubDate>
            <atom:updated>2026-09-07T20:48:02.891Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NISEQPw2tpYyV80U" /></figure><p>The hardest way to get a mentor is to ask someone to be your mentor.</p><p>I know how that sounds. Stay with me.</p><p>I’ve been on both sides of this. Early in my career, I wanted a mentor badly and had no idea how to get one. Now I mentor engineers, inside my company and outside it, and I can tell you exactly which requests get a yes and which get a polite deflection.</p><p>The difference is almost never the person asking.</p><p>It’s how they ask.</p><p>“Will you be my mentor?” feels like the honest, brave move. It’s actually the hardest possible request to say yes to.</p><p>Think about what it asks for: an open ended commitment of unknown size, on an undefined schedule, toward an unspecified goal. You’re asking someone busy to design the entire relationship, then also show up for it.</p><p>It’s like proposing on a first date. The intention is flattering. The commitment is terrifying.</p><p>So people say some version of “I’m stretched thin right now,” and both sides walk away feeling awkward about something that should have been energizing.</p><p>Here’s the part nobody tells you: the mentors who shaped my career were never asked. And the engineers I invest the most in never asked me either.</p><p>The relationship formed a different way.</p><p>Strip the label away and mentorship is a loop.</p><p>You ask a specific question. You get a perspective you couldn’t see alone. You act on it. You come back and share what happened.</p><p>That’s it.</p><p>That loop, repeated, is the relationship. Everything else, the label, the cadence, the coffee chats, is decoration.</p><p>Which means you don’t need permission to start it.</p><p>You need a first question.</p><p>Most people aim their mentor search at the most senior, most visible person they can find. A director. A famous engineer. A conference speaker.</p><p>Wrong target, for two reasons.</p><p>First, the person twenty steps ahead often can’t remember what your problem feels like. The person two steps ahead solved it recently. Their advice still has the details attached.</p><p>Second, access.</p><p>The best mentors are usually within reach already: the senior engineer on the next team whose design docs you admire, the lead who asks the questions in reviews you wish you’d asked, or someone in a community you’re both part of.</p><p>My filter is simple: find someone whose thinking you can actually observe.</p><p>Not their title. Their thinking.</p><p>If you’ve seen how they reason through a problem and thought, “I want to think like that,” you’ve found a candidate.</p><p>Here’s the move that actually works, and it costs nothing.</p><p>Ask one specific question.</p><p>Not “Can I pick your brain?” Not “Any advice for someone in my position?”</p><p>One real question about one real situation:</p><p>“You led the migration off the old auth service. I’m facing a similar call between refactoring and rebuilding. What did you weigh?”</p><p>Specific questions are easy to answer. They respect the other person’s time. And they show you’ve done your homework, because you can’t ask a specific question without understanding your own problem first.</p><p>Then comes the step almost everyone skips:</p><p>Do the thing, and report back.</p><p>“I took your advice. Here’s what happened.”</p><p>That might be the most underused sentence in career building.</p><p>It converts advice into investment.</p><p>The moment someone sees evidence that their input changed your outcome, something shifts. You stop being a person who asked a question. You become a person worth investing in.</p><p>Run that loop three or four times and look at what you have.</p><p>A person who knows your situation, gives you real input, and follows your progress.</p><p>That’s a mentor.</p><p>No one ever signed anything.</p><p>If the label ever matters, it arrives on its own.</p><p>“You’ve become a mentor to me, and I want you to know I don’t take it for granted” lands beautifully after six months of the loop.</p><p>As an opening line, it’s a burden.</p><p>As a thank you, it’s a gift.</p><p>Mentors aren’t found by asking.</p><p>They’re earned by showing up with real questions and coming back with real results.</p><p>Which means the thing standing between you and a mentor isn’t the courage to make the big ask.</p><p>It’s clarity about your next specific question.</p><p>This is Part 1 of a four part series on mentoring: getting one, being worth one, becoming one, and scaling it beyond yourself.</p><p>So let me ask you the question that starts the loop:</p><p>Who is the person two steps ahead of you, and what would you ask them this week?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/how-find-mentor-without-asking-you-my-leela-kumili-mixxc/?trackingId=uQ9Vc8l6THyWFV%2FBAdGpGA%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5947fc1137d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Most Dangerous Bug in Engineering Teams Is Silence]]></title>
            <link>https://medium.com/@leela.kumili/the-most-dangerous-bug-in-engineering-teams-is-silence-93da27af76d4?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/93da27af76d4</guid>
            <category><![CDATA[team-building]]></category>
            <category><![CDATA[bugs]]></category>
            <category><![CDATA[silent-treatment]]></category>
            <category><![CDATA[team-collaboration]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 17:46:01 GMT</pubDate>
            <atom:updated>2026-08-25T17:46:01.326Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FoIwnL8A4-KYiLw1" /></figure><p>Google spent years studying 180 of its own teams to answer a simple question:</p><p>What separates great teams from average ones?</p><p>Most people expected the answer to be obvious.</p><p>Talent.</p><p>The strongest engineers. The most experienced technical leads. The smartest people in the room.</p><p>But that wasn’t what Google found.</p><p>The strongest predictor of team effectiveness was psychological safety: whether people believed they could speak up, admit mistakes, ask questions, challenge ideas, and raise concerns without negative consequences.</p><p>The finding came from Project Aristotle, Google’s multi-year research into team performance.</p><p>And yet, years later, psychological safety is still often treated as an HR topic.</p><p>A culture initiative.</p><p>Something discussed during an offsite.</p><p>I think that misses the point.</p><p>Psychological safety is not the soft side of engineering.</p><p>It is an engineering system property.</p><p>Because every engineering team depends on one thing before it can build reliable software:</p><p>Accurate information.</p><p>And silence is what happens when important information gets trapped inside people’s heads.</p><p>The dangerous bugs are often the ones nobody reports.</p><p>Think about how we debug production systems.</p><p>A service is unhealthy. A metric stops moving. An alert fires.</p><p>The first question is:</p><p>“What signal did we miss?”</p><p>Engineering teams work the same way.</p><p>A developer sees a risk but does not mention it.</p><p>A reviewer does not understand a design but approves anyway.</p><p>An engineer notices a pattern causing repeated incidents but stays quiet.</p><p>A teammate is blocked for days but says everything is “on track.”</p><p>The problem is not that the team lacks intelligence.</p><p>The problem is that the system is losing signals.</p><p>And when a system loses signals, failures become inevitable.</p><p>You can usually see psychological safety through the artifacts teams already create.</p><p>No surveys required.</p><p>Start with standups.</p><p>Every team has blockers.</p><p>That is not a sign of failure. Complex engineering work creates uncertainty.</p><p>The question is whether people surface it quickly.</p><p>On a healthy team:</p><p>That sentence starts collaboration.</p><p>On an unhealthy team:</p><p>The words sound fine.</p><p>The timeline quietly slips.</p><p>The blocker gets more expensive.</p><p>The team does not have fewer problems. It has fewer visible problems.</p><p>The absence of “I’m stuck” is not evidence that nobody is stuck.</p><p>It is evidence that saying it has a cost.</p><p>Code review is one of the best places to observe team safety.</p><p>Two unhealthy patterns appear repeatedly.</p><p>The first is silence.</p><p>A large change arrives. The reviewer leaves one comment:</p><p>“LGTM.”</p><p>Not because they fully understand the design.</p><p>Because asking questions feels uncomfortable.</p><p>“What happens if this dependency is slow?”</p><p>“Why did we choose this approach?”</p><p>“Can you walk me through this tradeoff?”</p><p>Those questions are not challenges to the author.</p><p>They are how engineering works.</p><p>The second pattern is the opposite.</p><p>A review filled with dozens of comments about naming, formatting, and minor style preferences.</p><p>But no discussion about architecture, scalability, failure modes, or operational risk.</p><p>Nitpicks are easy.</p><p>They feel safe because they do not challenge decisions.</p><p>The important conversations are often the uncomfortable ones.</p><p>When reviews avoid those conversations, the questions do not disappear.</p><p>They simply move to production.</p><p>There is a moment in incident retrospectives where you can feel the room change.</p><p>Someone asks:</p><p>“Who wrote this code?”</p><p>Or:</p><p>“Who approved this change?”</p><p>The conversation immediately shifts.</p><p>People stop investigating and start defending.</p><p>The team is no longer trying to understand the failure.</p><p>It is trying to avoid becoming the failure.</p><p>But the most valuable information in an incident is not who made the mistake.</p><p>It is which assumption failed.</p><p>What did we believe would happen?</p><p>Why was that belief reasonable at the time?</p><p>What signal did we miss?</p><p>What system change prevents this class of failure?</p><p>The goal of an incident review is not to find the person closest to the problem.</p><p>It is to make sure the organization learns faster than it fails.</p><p>A blameless process only works when the room is actually safe.</p><p>Otherwise, you get a meeting that is technically blameless but emotionally defensive.</p><p>The timeline gets documented.</p><p>The action items get assigned.</p><p>The important sentence never gets spoken:</p><p>“Here is what I misunderstood.”</p><p>That sentence is where learning begins.</p><p>Design reviews are another powerful signal.</p><p>Watch what happens when a senior engineer presents an idea.</p><p>Do people ask difficult questions?</p><p>Do people challenge assumptions?</p><p>Does someone junior feel comfortable saying:</p><p>“What happens when this queue backs up?”</p><p>“What happens during a partial failure?”</p><p>“Are we assuming this dependency is always available?”</p><p>On a strong team, disagreement moves in every direction.</p><p>Experience earns influence, but it does not create immunity from questions.</p><p>On an unsafe team, disagreement only flows downward.</p><p>Senior people challenge junior people.</p><p>Junior people approve senior people’s ideas.</p><p>The result looks like alignment.</p><p>But sometimes it is just silence wearing alignment’s clothes.</p><p>We measure everything in engineering.</p><p>Latency.</p><p>Availability.</p><p>Error rates.</p><p>Deployment frequency.</p><p>Incident recovery time.</p><p>But we rarely measure the signals that create those outcomes.</p><p>Consider these as team health indicators:</p><p>How often does someone say “I’m blocked”?</p><p>How many risks are raised before implementation starts?</p><p>How many design reviews contain meaningful questions?</p><p>How many incident retrospectives identify failed assumptions?</p><p>How often does someone challenge a decision regardless of seniority?</p><p>These are not culture metrics.</p><p>They are system health indicators.</p><p>A team that cannot surface reality cannot improve reality.</p><p>The good news is that improving this does not require becoming a manager.</p><p>Small behaviors change the system.</p><p>Ask:</p><p>“What are we missing?”</p><p>before asking:</p><p>“Are we good?”</p><p>Ask:</p><p>“What could fail?”</p><p>before asking:</p><p>“When will it be done?”</p><p>When someone raises a concern, pay attention to your reaction.</p><p>That moment teaches everyone watching.</p><p>“Good catch. Let’s explore it.”</p><p>Creates a very different environment than:</p><p>“Why didn’t we think of this earlier?”</p><p>Credit people for finding problems, not just solving them.</p><p>The person who finds the issue early saved the team time, even if they did not write the fix.</p><p>And one of the simplest habits:</p><p>Speak last.</p><p>When the most senior person shares an opinion first, many people unconsciously adjust their thinking around it.</p><p>Questions create space.</p><p>Answers close it.</p><p>The best leaders create environments where the best idea wins, regardless of who said it.</p><p>We spend our careers building reliable software systems.</p><p>We add observability.</p><p>We remove single points of failure.</p><p>We improve feedback loops.</p><p>We design for failure.</p><p>Teams need the same thinking.</p><p>Silence is a missing signal.</p><p>A missing signal becomes a hidden risk.</p><p>A hidden risk eventually becomes an incident.</p><p>The most dangerous bugs in software are not always the ones with obvious errors.</p><p>They are the ones nobody notices until it is too late.</p><p>The same is true for engineering teams.</p><p>The question is not whether your team has problems.</p><p>Every team does.</p><p>The question is:</p><p>Does your team know how to surface them before they become expensive?</p><p>What is one signal that told you a team was safe enough to speak honestly?</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/most-dangerous-bug-engineering-teams-silence-leela-kumili-3vqec/?trackingId=eNSK7lnGTlG9tFOKdgFBJg%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=93da27af76d4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Code Review Isn’t About Code Anymore]]></title>
            <link>https://medium.com/@leela.kumili/code-review-isnt-about-code-anymore-e66631397ea2?source=rss-d30d57fc8a21------2</link>
            <guid isPermaLink="false">https://medium.com/p/e66631397ea2</guid>
            <category><![CDATA[code-review]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[better-engineers]]></category>
            <dc:creator><![CDATA[Leela Kumili]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 17:46:01 GMT</pubDate>
            <atom:updated>2026-08-24T17:46:01.187Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pHI-uUnbJrMHXCPS" /></figure><p>Last year, every code review required spending time verifying implementation quality. Were failure paths handled correctly? Was the retry strategy resilient? Were edge cases tested? Did the implementation follow our engineering patterns? Did it introduce unnecessary operational complexity? Only after answering those questions could I focus on whether the design itself made sense.</p><p>It was valuable work.</p><p>It was also repetitive.</p><p>Today, AI catches many of those issues before I even open the pull request.</p><p>Surprisingly, my reviews haven’t become shorter.</p><p>They’ve become completely different.</p><p>Somewhere along the way, code review quietly stopped being about code.</p><p>For years, code review combined two completely different responsibilities.</p><p>Most of us never noticed because they happened in the same pull request and ended with the same Approve button.</p><p>The first responsibility was reviewing implementation quality.</p><p>Were failure paths handled correctly? Were tests sufficient? Was logging consistent? Did the implementation follow our engineering patterns? Was it resilient? Did it introduce unnecessary complexity?</p><p>This work is important.</p><p>But much of it is pattern matching.</p><p>The answers already exist somewhere. In engineering documentation. In architectural guidelines. In team conventions. Or simply in the collective experience of the engineers who built the system.</p><p>The second responsibility is very different.</p><p>Is this solving the right problem?</p><p>Should this logic exist in this service?</p><p>What happens when another team depends on this API?</p><p>Will this retry strategy amplify traffic during an outage?</p><p>Are we introducing technical debt that will quietly become everyone’s problem six months from now?</p><p>Is this the simplest solution that will still serve us a year from now?</p><p>These questions don’t have a checklist.</p><p>They require context.</p><p>They require experience.</p><p>They require judgment.</p><p>AI is becoming remarkably good at the first category.</p><p>That shifts much more of our attention toward the second.</p><p>I recently looked through my own review history.</p><p>A year ago, a larger share of my comments focused on implementation quality.</p><p>“This retry policy should use exponential backoff.”</p><p>“Add a circuit breaker for prolonged downstream failures.”</p><p>“Cover the failure path in the tests.”</p><p>Today, those comments are becoming less common.</p><p>Instead, my reviews sound more like architecture discussions.</p><p>“What happens if this retry policy hits a downstream service during an incident?”</p><p>“Should this belong in the shared platform instead of this application?”</p><p>“Will this API contract prevent another team from evolving independently?”</p><p>“Is this optimization worth the additional operational complexity?”</p><p>The conversation shifted.</p><p>I’m spending less time reviewing implementation.</p><p>I’m spending more time asking whether this change belongs in the system at all.</p><p>That feels like a much better use of engineering time.</p><p>Here’s the interesting part.</p><p>Most companies already possess the knowledge required to automate a large portion of implementation review.</p><p>It’s just scattered everywhere.</p><p>Historical pull request comments.</p><p>Internal engineering guidelines.</p><p>Architecture documents.</p><p>Postmortems.</p><p>The unwritten advice every senior engineer repeats:</p><p>“We don’t do it that way here.”</p><p>Those patterns can become an AI review playbook.</p><p>Not a generic linter.</p><p>Not generic best practices.</p><p>Your team’s architectural patterns.</p><p>Your resilience standards.</p><p>Your observability requirements.</p><p>Your deployment practices.</p><p>Your security expectations.</p><p>Your testing philosophy.</p><p>Your operational lessons learned from real production incidents.</p><p>Imagine every pull request being reviewed against your organization’s engineering knowledge before another engineer even opens it.</p><p>An AI reviewer flags retry strategies that violate your resilience standards. It identifies APIs that don’t follow established architectural patterns. It reminds engineers about observability requirements or deployment practices learned from past incidents.</p><p>That’s not generic code review.</p><p>That’s your organization’s engineering knowledge applied automatically.</p><p>By the time a human opens the pull request, most implementation issues have already been identified, explained, and often fixed.</p><p>The human review starts where AI stops.</p><p>Implementation review and engineering review no longer have to happen in the same step. One can increasingly be automated and continuous. The other becomes a conversation about architecture, tradeoffs, operational risk, and long term ownership.</p><p>As implementation quality becomes increasingly automated, the purpose of review changes.</p><p>The question is no longer:</p><p>“Is this implementation correct?”</p><p>It becomes:</p><p>“Is this the right engineering decision?”</p><p>Those are fundamentally different conversations.</p><p>One evaluates implementation.</p><p>The other evaluates engineering tradeoffs.</p><p>That shift changes what makes someone a strong reviewer.</p><p>What becomes more valuable is understanding the system.</p><p>Knowing another team’s roadmap.</p><p>Remembering the outage caused by a similar design.</p><p>Recognizing when an elegant solution creates unnecessary operational complexity.</p><p>Balancing long term maintainability against immediate delivery.</p><p>Understanding which constraints actually matter.</p><p>None of those decisions live inside the diff.</p><p>They live inside the engineer reviewing it.</p><p>Staff and principal engineers have always created value by carrying context across systems and teams. AI doesn’t reduce the value of that context. It makes it even more important.</p><p>The most experienced engineers were always doing this part of the review.</p><p>The problem was that strategic conversations were buried beneath implementation feedback.</p><p>As AI removes more implementation review, those conversations become the review.</p><p>That’s significant because reviewing engineering decisions is fundamentally leadership work.</p><p>You’re aligning teams.</p><p>Protecting architectural direction.</p><p>Sharing operational experience.</p><p>Teaching judgment instead of syntax.</p><p>Helping engineers understand why, not just how.</p><p>In many ways, code review is becoming one of the clearest places engineers practice technical leadership long before they ever become managers or staff engineers.</p><p>Open your last ten pull request reviews.</p><p>How many comments are about implementation?</p><p>How many are about engineering tradeoffs?</p><p>The second list is where your engineering judgment shows up.</p><p>The biggest change AI made to my code reviews wasn’t speed.</p><p>It changed what deserves human attention.</p><p>The less time I spend reviewing implementation, the more time I spend reviewing systems.</p><p>And I think that’s exactly where engineering leadership is heading.</p><p>Not reviewing code.</p><p>Reviewing decisions.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/code-review-isnt-anymore-leela-kumili-byhyc/?trackingId=h3YQJua6SPadXSrgjYw2KQ%3D%3D"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e66631397ea2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>