<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Pathum Dilshan</title>
    <description>The latest articles on DEV Community by Pathum Dilshan (@pmpdilshan).</description>
    <link>https://dev.to/pmpdilshan</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4257470%2F7982aaa7-4bc7-4bf4-a518-2cfd529e2aa2.jpg</url>
      <title>DEV Community: Pathum Dilshan</title>
      <link>https://dev.to/pmpdilshan</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/pmpdilshan"/>
    <language>en</language>
    <item>
      <title>I Built a Sales Dashboard. Here's What I'd Change.</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:42:39 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/i-built-a-sales-dashboard-heres-what-id-change-4dhm</link>
      <guid>https://dev.to/pmpdilshan/i-built-a-sales-dashboard-heres-what-id-change-4dhm</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/dashboard-design-lessons.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;A while ago I built a &lt;a href="https://datastudio.google.com/reporting/f2e7038a-fdd7-41a6-80ad-283e66f2b6d4" rel="noopener noreferrer"&gt;Looker Studio dashboard&lt;/a&gt; on 817,741 supermarket transactions: three headline numbers, a daily sales line, sales by district, by day of week and by age group. It works. People can read it. But looking at it again, there's a lot I'd do differently.&lt;/p&gt;

&lt;p&gt;Reviewing your own work in public is uncomfortable, and it's also the fastest way to get better. Here's my honest review.&lt;/p&gt;

&lt;h2&gt;
  
  
  What worked
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Headline numbers first.&lt;/strong&gt; Total sales, gross profit and customers sit at the top, so the big picture takes two seconds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One question per chart.&lt;/strong&gt; Each chart answers exactly one thing: when, where, which day, who.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sorted bars.&lt;/strong&gt; District sales are sorted largest to smallest, so the ranking is obvious without reading a single number.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I'd change
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Replace the pie chart
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fknhoi5buj7rukumr3ueg.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fknhoi5buj7rukumr3ueg.png" alt="Sales by age group as a sorted bar chart" width="800" height="480"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The same age-group data as a sorted bar chart. Compare it with six pie slices.&lt;/p&gt;

&lt;p&gt;The age-group pie has six slices, and people are poor at comparing angles. Is 30–34 (18%) bigger than 40–44 (17.5%)? In a pie you can't tell. In a sorted bar chart it's instant.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Show margin, not just sales
&lt;/h3&gt;

&lt;p&gt;The dashboard shows gross profit as one number but everything else as sales. A district with high sales and thin margins looks like a winner when it might not be. Margin by district and by day would be far more useful to a manager.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Mark missing data on the chart
&lt;/h3&gt;

&lt;p&gt;In late December the daily line drops to almost zero. That's almost certainly missing data, not a business collapse. But nothing on the chart says so. Anyone reading it could draw the wrong conclusion, so I'd shade or annotate those days.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Add comparison
&lt;/h3&gt;

&lt;p&gt;A number on its own means little. "NT\$26M on Sundays" is more useful as "Sunday is 2× a weekday." Comparisons with an average, a target or the previous period turn numbers into decisions.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Write the headline, not just the title
&lt;/h3&gt;

&lt;p&gt;"Sales by Day of Week" describes the chart. "Weekends bring in 43% of sales" tells you what it means. Insight titles save every reader the work of figuring it out.&lt;/p&gt;

&lt;h2&gt;
  
  
  Seven rules I'll follow next time
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt; Start from the decision the dashboard supports, not the data available.&lt;/li&gt;
&lt;li&gt; Headline numbers at the top, detail below.&lt;/li&gt;
&lt;li&gt; One question per chart.&lt;/li&gt;
&lt;li&gt; Bars over pies; sort by value.&lt;/li&gt;
&lt;li&gt; Show comparisons, not just totals.&lt;/li&gt;
&lt;li&gt; Annotate anything unusual, especially gaps.&lt;/li&gt;
&lt;li&gt; Title charts with the insight.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A dashboard isn't finished when it works. It's finished when someone acts on it.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>datascience</category>
      <category>analytics</category>
      <category>dataviz</category>
      <category>beginners</category>
    </item>
    <item>
      <title>What 817,741 Supermarket Transactions Taught Me About Reading Data</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:39:24 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/what-817741-supermarket-transactions-taught-me-about-reading-data-5f2j</link>
      <guid>https://dev.to/pmpdilshan/what-817741-supermarket-transactions-taught-me-about-reading-data-5f2j</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/supermarket-sales-insights.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;For a business intelligence project I built a &lt;a href="https://datastudio.google.com/reporting/f2e7038a-fdd7-41a6-80ad-283e66f2b6d4" rel="noopener noreferrer"&gt;Looker Studio dashboard&lt;/a&gt; on a public supermarket dataset: 817,741 transactions from a Taiwanese supermarket between November 2000 and February 2001. Total sales: NT\$107.8 million. Gross profit: NT\$16.2 million. About 32,300 customers.&lt;/p&gt;

&lt;p&gt;The dataset is old, but the patterns in it are the kind every retailer has. Here's what stood out.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. The weekend is almost half the week
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fsjvsv1omvclno5g767c7.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fsjvsv1omvclno5g767c7.png" alt="Sales by day of week: Sunday and Saturday far ahead" width="800" height="480"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Sales by day of week (NT\$ millions), from the dashboard.&lt;/p&gt;

&lt;p&gt;Sunday alone brought in about NT\$26M, and Saturday about NT\$20M. Together that's roughly &lt;strong&gt;43% of all sales in two days&lt;/strong&gt;. Monday to Friday sit almost flat at NT\$12–14M each.&lt;/p&gt;

&lt;p&gt;For a store manager, that's a staffing schedule, a stock-replenishment plan and a promotion calendar in one chart.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Two districts carry the business
&lt;/h2&gt;

&lt;p&gt;Customers from Nangang (NT\$37.9M) and Xizhi (NT\$31.5M) account for about &lt;strong&gt;64% of sales&lt;/strong&gt;. Every other district is a long tail. If you're deciding where to spend a marketing budget, that's a very short list.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Customers aged 30–44 are the core
&lt;/h2&gt;

&lt;p&gt;The 35–39 age band is the single biggest group at 22.6% of sales, followed by 30–34 (18%) and 40–44 (17.5%). Together, &lt;strong&gt;30–44 year-olds drive almost 60% of revenue&lt;/strong&gt;: likely families doing the weekly shop, which also explains the weekend peak.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Holidays show up as spikes
&lt;/h2&gt;

&lt;p&gt;The daily sales line jumps to its highest point in late January 2001, just before Lunar New Year. Seasonal events like this are why you never compare one month with another without context.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. The finding that wasn't a finding
&lt;/h2&gt;

&lt;p&gt;In late December the daily sales line drops to almost zero for several days. My first reaction was "something happened to the business." But a supermarket doesn't lose almost all its sales for a week and then bounce straight back. Much more likely, those days are &lt;strong&gt;missing from the data&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;This is the most important lesson in the whole project: &lt;strong&gt;before you explain a pattern, check it's real.&lt;/strong&gt; A dashboard makes every number look equally trustworthy. Your job as an analyst is to know which ones aren't.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd do next
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Break the weekend peak down by product category to see what people actually buy.&lt;/li&gt;
&lt;li&gt;Check margin, not just sales, by district: big revenue doesn't always mean big profit.&lt;/li&gt;
&lt;li&gt;Flag missing dates on the chart itself, so nobody misreads them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Good dashboards answer questions. Great ones also tell you when not to trust them.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>datascience</category>
      <category>analytics</category>
      <category>dataviz</category>
      <category>beginners</category>
    </item>
    <item>
      <title>From Drive Tests to Data Pipelines: Why I Went from RF Engineer to Data &amp; AI</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:36:46 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/from-drive-tests-to-data-pipelines-why-i-went-from-rf-engineer-to-data-ai-49om</link>
      <guid>https://dev.to/pmpdilshan/from-drive-tests-to-data-pipelines-why-i-went-from-rf-engineer-to-data-ai-49om</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/rf-engineer-to-data-ai.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Two years ago my working day looked like this: drive-test logs, KPI dashboards, alarm lists, and a lot of Excel. Today I'm an RF and network planning engineer at Mobitel, reading for an MSc in Data Science &amp;amp; AI at the University of Moratuwa, and I build Python and Apache Airflow pipelines that process over 10 million KPI records a day.&lt;/p&gt;

&lt;p&gt;I didn't plan a "career switch." I just kept running into the same wall, and data science was the way through it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The wall: too much data, too few hours
&lt;/h2&gt;

&lt;p&gt;A mobile network produces a staggering amount of data. Every cell reports dozens of counters every hour. When I started, first as a Huawei technical executive deployed to Hutch and then at Mobitel, most of my time went into &lt;em&gt;collecting&lt;/em&gt; and &lt;em&gt;formatting&lt;/em&gt; that data, not understanding it.&lt;/p&gt;

&lt;p&gt;So I started automating. A Python script to merge exports. Then a scheduled job. Then a proper pipeline. Every hour I saved went into the part of the job I actually enjoyed: finding out &lt;em&gt;why&lt;/em&gt; a cell was misbehaving.&lt;/p&gt;

&lt;h2&gt;
  
  
  What carried over from RF engineering
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Thinking in systems.&lt;/strong&gt; A dropped call is never one thing. It's coverage, interference, handover parameters, load and hardware all interacting. That's exactly how you need to think about a machine-learning model and its data.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Distrust of averages.&lt;/strong&gt; Network engineers learn early that a "good" network-wide KPI can hide terrible individual cells. Data scientists call this aggregation bias. Same lesson, different name.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Domain knowledge.&lt;/strong&gt; This is the big one. Knowing that a KPI jumped because two vendors define a counter differently, not because something broke, is worth more than any algorithm.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What didn't carry over
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Statistics, properly.&lt;/strong&gt; I could read a KPI chart, but I had to relearn probability, hypothesis testing and validation from the ground up.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Software habits.&lt;/strong&gt; Version control, testing, reproducible environments. Engineers write scripts; data teams need systems.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Being comfortable with "it depends."&lt;/strong&gt; RF has clear specifications. Machine learning is full of trade-offs where the honest answer is "let's test it."&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I'd tell another engineer
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt; &lt;strong&gt;Start with your own problem.&lt;/strong&gt; Automate one report you hate. You'll learn more than from ten generic tutorials.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Your domain is your advantage.&lt;/strong&gt; There are many people who can train a model. Far fewer can tell you whether its output makes sense in a live network.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Build in public.&lt;/strong&gt; Write about what you learn. My Medium articles on medical AI and reinforcement learning taught me more than the projects themselves, because explaining forces clarity.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Formal study helps, but it's not the starting line.&lt;/strong&gt; I started the MSc after I'd already been automating for a while. The theory landed far better because I had real problems to attach it to.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Where this is going
&lt;/h2&gt;

&lt;p&gt;My MSc research is on drift-aware anomaly detection for mobile network KPIs, a problem I met first as an engineer and now get to study properly. It's the clearest proof I have that the two careers aren't separate. The best work happens where they overlap.&lt;/p&gt;

&lt;p&gt;If you're an engineer thinking about data and AI: you're not starting from zero. You're starting with the hardest part already done.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>career</category>
      <category>datascience</category>
      <category>telecom</category>
      <category>ai</category>
    </item>
    <item>
      <title>How I Automated 10M+ Network KPI Records a Day with Python and Apache Airflow</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:32:40 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/how-i-automated-10m-network-kpi-records-a-day-with-python-and-apache-airflow-2idm</link>
      <guid>https://dev.to/pmpdilshan/how-i-automated-10m-network-kpi-records-a-day-with-python-and-apache-airflow-2idm</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/airflow-kpi-pipeline.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Every hour, a mobile network generates performance counters for every cell: traffic, drops, handovers, throughput, congestion. At Mobitel that adds up to more than 10 million KPI records a day. For a long time, turning that into reports meant exporting files, merging them in Excel and rebuilding the same charts every morning.&lt;/p&gt;

&lt;p&gt;So I built pipelines in Python and Apache Airflow to do it automatically. They cut manual reporting effort by about 40%. Here's the pattern, and what I'd tell anyone building something similar.&lt;/p&gt;

&lt;h2&gt;
  
  
  The pattern
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frxc09jdubgkjh5gnygfx.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frxc09jdubgkjh5gnygfx.png" alt="Pipeline: extract, validate, transform, load, report, alert" width="800" height="288"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The general shape of the pipeline.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt; &lt;strong&gt;Extract:&lt;/strong&gt; pull counter exports from the network management systems on a schedule.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Validate:&lt;/strong&gt; check that files arrived, row counts look sensible and no cells are missing.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Transform:&lt;/strong&gt; clean, standardise names, calculate KPIs from raw counters and aggregate by cell, cluster and region.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Load:&lt;/strong&gt; store the results in a database that reports and analysis can query.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Report and alert:&lt;/strong&gt; refresh dashboards and flag cells that need attention.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Airflow ties it together. Each step is a task, the dependencies are explicit, and if something fails you can see exactly where and rerun just that part.&lt;/p&gt;

&lt;h2&gt;
  
  
  Lessons that mattered most
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Validation is not optional
&lt;/h3&gt;

&lt;p&gt;The most dangerous pipeline isn't the one that fails. It's the one that succeeds with bad data. A late file, a missing region or a changed column name can silently produce a report that looks fine and is wrong. Checking inputs before processing them saved me more than once.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Make every task safe to rerun
&lt;/h3&gt;

&lt;p&gt;Things will fail: a server is down, a file is late. If rerunning a task creates duplicates or overwrites good data, every failure becomes a manual clean-up. Designing each step so it can run twice with the same result (idempotency) makes recovery a click instead of an afternoon.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Calculate KPIs in one place
&lt;/h3&gt;

&lt;p&gt;When three reports each calculate "drop rate" slightly differently, meetings turn into arguments about whose number is right. Defining each KPI once, in the pipeline, means everyone works from the same truth.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Automate the boring part first
&lt;/h3&gt;

&lt;p&gt;I didn't start with machine learning. I started with the report everyone hated building. Once the data was flowing reliably, more interesting work like ML-based anomaly detection became possible, because the foundation was there.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Log as if you'll debug it at 2 a.m.
&lt;/h3&gt;

&lt;p&gt;Because one day you will. Clear logs saying what ran, how many rows and what was skipped turn a mystery into a five-minute fix.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it changed
&lt;/h2&gt;

&lt;p&gt;About 40% less time spent on manual reporting, consistent numbers across teams, and more time for actual optimization work. The lesson I'd pass on: &lt;strong&gt;stop building reports. Build the thing that builds reports.&lt;/strong&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>python</category>
      <category>airflow</category>
      <category>dataengineering</category>
      <category>telecom</category>
    </item>
    <item>
      <title>Swapping 500+ Live Cell Sites Without Customers Noticing: Lessons from a Vendor Swap</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:26:04 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/swapping-500-live-cell-sites-without-customers-noticing-lessons-from-a-vendor-swap-2mg2</link>
      <guid>https://dev.to/pmpdilshan/swapping-500-live-cell-sites-without-customers-noticing-lessons-from-a-vendor-swap-2mg2</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/vendor-swap-lessons.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;At Mobitel I supported a vendor swap of more than 500 cell sites in the Gampaha area, replacing Huawei radio equipment with ZTE. The goal sounds simple: change the equipment, keep the network working. In practice it's one of the most demanding projects a radio network team can take on, because the network never stops. People are making calls, streaming and paying with their phones while you replace the equipment under them.&lt;/p&gt;

&lt;p&gt;Here's what I took away from it.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Your baseline is everything
&lt;/h2&gt;

&lt;p&gt;Before a single site is touched, you need a clear picture of how every cell performs today: drop rate, accessibility, handover success, throughput, traffic. Without a solid pre-swap baseline, you can't tell whether a post-swap change is a problem or just normal variation.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Same network, different numbers
&lt;/h2&gt;

&lt;p&gt;Different vendors don't always define or count KPIs the same way. After a swap, some numbers move even when the customer experience hasn't changed. Teams that don't know this waste days chasing "degradations" that are really definitional differences, or worse, dismiss real problems as "just the new counters."&lt;/p&gt;

&lt;p&gt;(This exact problem later became the starting point for my MSc research on drift-aware anomaly detection.)&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Neighbours and parameters are where problems hide
&lt;/h2&gt;

&lt;p&gt;Most post-swap issues weren't hardware failures. They were in the configuration: neighbour relations, handover parameters, PCI planning, and the border between swapped and not-yet-swapped areas. The cluster boundary is where two vendors' equipment must hand calls back and forth, and it needs extra attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Swap in clusters, not at random
&lt;/h2&gt;

&lt;p&gt;Swapping sites in geographic clusters lets you stabilise one area completely before moving on, and keeps the number of vendor boundaries manageable. Each cluster becomes a small, repeatable project with a clear before-and-after.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. The first 72 hours decide everything
&lt;/h2&gt;

&lt;p&gt;Most issues show up within a few days of a cluster going live. Close KPI monitoring in that window, with a quick way to act on what you find, is what keeps downtime low. Automating that monitoring made a big difference: you can't watch hundreds of cells by hand.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. It's a people project as much as a technical one
&lt;/h2&gt;

&lt;p&gt;Field teams, the vendor, operations, planning and optimization all have to move in step. Clear communication about what's changing tonight, what to watch tomorrow and who to call when something breaks matters as much as any parameter.&lt;/p&gt;

&lt;h2&gt;
  
  
  The result
&lt;/h2&gt;

&lt;p&gt;The swap went ahead with minimum service downtime. Across my optimization work at Mobitel, I've also helped reduce call drop rate by 15% and increase downlink throughput by 20% across 250+ sites. That came from the same habits: measure carefully, change deliberately and watch closely.&lt;/p&gt;

&lt;p&gt;The best vendor swap is the one customers never notice.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>telecom</category>
      <category>networking</category>
      <category>projectmanagement</category>
      <category>career</category>
    </item>
    <item>
      <title>Your Network Alarms Are Lying to You</title>
      <dc:creator>Pathum Dilshan</dc:creator>
      <pubDate>Sat, 21 Nov 2026 14:25:25 +0000</pubDate>
      <link>https://dev.to/pmpdilshan/your-network-alarms-are-lying-to-you-2gdn</link>
      <guid>https://dev.to/pmpdilshan/your-network-alarms-are-lying-to-you-2gdn</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://pmpdilshan.github.io/kpi-anomaly-detection.html" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;At 3 a.m. a cell starts dropping calls. Nobody gets an alert.&lt;/p&gt;

&lt;p&gt;At 9 a.m. the same cell gets busy, as it does every weekday, and the alarm goes off. An engineer checks it, finds nothing wrong, and closes the ticket.&lt;/p&gt;

&lt;p&gt;If you work in network operations, you’ve lived this. I have too, as an RF and network planning engineer on a live 2G/3G/4G/5G network. The cause is almost always the same rule:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;if drop_rate &amp;gt; 2%:
    raise_alarm()
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Static thresholds are simple, explainable and easy to roll out across thousands of cells. They are also wrong far more often than most teams admit. Here’s why, and what to do instead.&lt;/p&gt;




&lt;h2&gt;
  
  
  1. Every cell has its own “normal”
&lt;/h2&gt;

&lt;p&gt;A cell next to a highway, one covering a busy market and one in a quiet village behave completely differently. One threshold can’t fit all three. Set it low and the busy cells flood you with false alarms. Set it high and the quiet cell can degrade for weeks without anyone noticing.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Time changes everything
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5kxfozd5meljfegozu2p.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5kxfozd5meljfegozu2p.png" alt="Chart: a static threshold fires at every peak hour but misses a real fault at night" width="800" height="462"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Illustrative. The static threshold fires at every peak hour but misses a real fault at night, which the learned band catches.&lt;/p&gt;

&lt;p&gt;Traffic at 9 a.m. on a Monday looks nothing like traffic at 2 a.m. on a Sunday. A value that is alarming at night can be perfectly healthy at peak hour. A fixed line through the chart can’t tell the difference.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Networks change on purpose
&lt;/h2&gt;

&lt;p&gt;This is the one that hurt most. I worked on a vendor swap of &lt;strong&gt;more than 500 sites&lt;/strong&gt;, moving from one equipment vendor to another. KPI baselines shifted overnight. Nothing was broken: counters are simply defined differently between vendors.&lt;/p&gt;

&lt;p&gt;Static thresholds lit up everywhere. The real faults were buried in the noise.&lt;/p&gt;




&lt;h2&gt;
  
  
  The first fix: let the model learn “normal”
&lt;/h2&gt;

&lt;p&gt;Instead of telling the system what normal is, let it learn normal &lt;strong&gt;per cell, per hour and per day of the week&lt;/strong&gt;, then flag deviations from that.&lt;/p&gt;

&lt;p&gt;Even simple statistical and machine-learning baselines do this far better than fixed limits. In my work, an ML-based approach improved detection accuracy by &lt;strong&gt;about 18%&lt;/strong&gt; compared with the static thresholds we had been using.&lt;/p&gt;

&lt;p&gt;But learned baselines bring a new problem.&lt;/p&gt;

&lt;h2&gt;
  
  
  The drift trap
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fusndgnmq7tkjr3vxv13x.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fusndgnmq7tkjr3vxv13x.png" alt="Chart: blind retraining learns a real fault as the new normal while a drift-aware baseline does not" width="800" height="462"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Illustrative. Blind retraining follows the vendor swap correctly, then slowly learns a real fault as the new normal. A drift-aware baseline accepts the expected change and still flags the fault.&lt;/p&gt;

&lt;p&gt;A model trained on last month’s data slowly goes stale. New sites go live, parameters are retuned, traffic grows and software is upgraded. This is called &lt;strong&gt;concept drift&lt;/strong&gt;, and in telecom it’s the normal state, not the exception.&lt;/p&gt;

&lt;p&gt;The obvious fix is to keep retraining on recent data. But retrain blindly and the model eventually learns a genuine fault as “normal.” A cell that has been degraded for two weeks stops looking anomalous. It hasn’t recovered; the model has simply got used to it.&lt;/p&gt;

&lt;p&gt;So the real question is:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;How do you adapt to normal change without absorbing genuine faults?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  What I’m researching
&lt;/h2&gt;

&lt;p&gt;That question is the focus of my MSc research in Data Science &amp;amp; AI at the University of Moratuwa: &lt;strong&gt;drift-aware anomaly detection and root-cause ranking for mobile RAN KPI data.&lt;/strong&gt; At a high level, the approach has three parts:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Use what you already know.&lt;/strong&gt; Vendor swaps, new sites and parameter changes are planned in advance. A detector should use that context instead of guessing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Adapt carefully.&lt;/strong&gt; Update the baseline only when a shift is consistent, explainable and not tied to a worse user experience.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Don’t stop at “something is wrong.”&lt;/strong&gt; “Cell 4521 is anomalous” still leaves an engineer digging through dozens of counters. Ranking which KPIs most likely explain each anomaly turns an alert into a starting point for root-cause analysis.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why it matters
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Fewer false alarms&lt;/strong&gt; mean engineers spend their time on problems customers actually feel.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Root-cause ranking&lt;/strong&gt; saves hours on every incident.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A detector that survives network change&lt;/strong&gt; means you don’t rebuild your monitoring every time the network evolves, which in this industry is constantly.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The bigger lesson
&lt;/h2&gt;

&lt;p&gt;The most valuable thing I’ve learned working between RF engineering and data science is this: &lt;strong&gt;domain knowledge decides whether a model is useful.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A data scientist without network context will happily train on data whose definitions changed halfway through. An RF engineer without data skills will keep tuning thresholds by hand. The real value is where the two meet.&lt;/p&gt;

&lt;p&gt;I’ll share results as the research progresses. &lt;strong&gt;How does your team handle KPI drift?&lt;/strong&gt; Tell me in the comments. I read every one.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Pathum Dilshan is an RF &amp;amp; Network Planning Engineer at Mobitel, Sri Lanka, and an MSc Data Science &amp;amp; AI student at the University of Moratuwa. He writes about telecom, data and AI. Connect on &lt;a href="https://www.linkedin.com/in/pathum-dilshan/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt; or visit &lt;a href="https://pmpdilshan.github.io/" rel="noopener noreferrer"&gt;pmpdilshan.github.io&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>telecom</category>
      <category>machinelearning</category>
      <category>datascience</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
