<?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 Kavindu Sachinthe on Medium]]></title>
        <description><![CDATA[Stories by Kavindu Sachinthe on Medium]]></description>
        <link>https://medium.com/@kavix?source=rss-a8ffbb5b13df------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7Ckx_m1cWs0jnW0B6g2VAg.jpeg</url>
            <title>Stories by Kavindu Sachinthe on Medium</title>
            <link>https://medium.com/@kavix?source=rss-a8ffbb5b13df------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 22:16:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kavix/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[Jev Explained: Turning AI Intelligence into Typed Decisions]]></title>
            <link>https://kavix.medium.com/jev-explained-turning-ai-intelligence-into-typed-decisions-f55ab7f4c4d6?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/f55ab7f4c4d6</guid>
            <category><![CDATA[jevs]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[typesafe-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[system-1-and-system-2]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 12:44:45 GMT</pubDate>
            <atom:updated>2026-09-21T12:44:45.432Z</atom:updated>
            <content:encoded><![CDATA[<h3>Jev Explained The AI Model That Doesn’t Talk — It Decides</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ti-hXRdmJgl4q5Hd.jpg" /></figure><p>Jev is a new kind of AI model released by TypeSafe AI in September 2026.</p><p>It does not generate text not write emails, code, or explanations and It does not chat with you.</p><p>Instead, you give Jev two things</p><p>1. State— any text or JSON (a support ticket, a conversation, a document, an agent’s plan, etc.)<br>2. Questions— typed questions you define in advance</p><p>Jev returns typed answers with probabilities. Your code then decides what to do with those numbers.</p><p>Think of it as a frontier-intelligence function call<br>unstructured state goes in and typed probabilistic decisions come out.</p><p>This is what makes the idea different from simply asking an LLM to generate JSON. The output is constrained by the decision type you define, rather than asking the model to invent a structure and hoping it follows the schema.</p><p>Response time is typically <strong>70–500 milliseconds</strong>, according to TypeSafe’s description. And because Jev’s decision primitives restrict the possible outputs, your application can work with predictable types such as probabilities, choices, and scores.</p><p>That combination <strong>structured decisions, uncertainty awareness, missing-data handling, and low latency</strong> is the part of Jev that I find most interesting. It starts to look less like a chatbot and more like an intelligence layer that can sit directly inside a software system.</p><h3>System One Models and the Name Jev</h3><p>TypeSafe calls Jev the first System One model.</p><p>The name comes from Daniel Kahneman’s book Thinking, Fast and Slow</p><p>- System 1 = fast, intuitive, automatic decisions<br>- System 2 = slow, deliberate, step-by-step reasoning</p><p>Most large language models (GPT, Claude, Grok…) are trained to be good at System 2 style works like writing, reasoning, explaining.</p><p>Jev is built only for System 1 style works quick, structured judgments that software can use directly.</p><p>The model itself is named after William Stanley Jevons, the 19th-century economist. TypeSafe chose the name because they believe cheaper and faster decisions will create more demand for AI decisions (similar to the Jevons paradox in economics).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SPH5wdfI7VJXEfwb.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fYAaZu7KYXGaVNTOfiUaPQ.png" /></figure><p>LLMs are still better when you need the model to write something or do long reasoning. <br>Jev is better when you already know the possible answers and only need the model to pick or score them fast and reliably.</p><h4>How Much Does Jev Cost?</h4><p>- Input: $0.042 per million tokens <br>- Output: Free (there is almost no text to generate)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PyHERqtCHxa5jC59aHJQHw.png" /></figure><p>TypeSafe claims Jev can be 40–200× faster and 40–400× cheaper than frontier LLMs on the same structured decision tasks. Real numbers depend on your exact workload, but the pricing is deliberately aggressive.</p><h3>How Does Jev Work? (Architectural Breakdown)</h3><p>Normal LLMs work like this:</p><p>1. Read the prompt<br>2. Predict the next token<br>3. Feed that token back in<br>4. Predict the next token<br>5. Repeat until the answer is finished</p><p>This is called autoregressive generation. It is powerful but slow, and the model can produce invalid JSON or wander off topic.</p><p>Jev does something different</p><p>1. You send the state + a list of typed questions<br>2. The model encodes the state once<br>3. A <strong>parallel sampler</strong> scores every possible answer for every question at the same time<br>4. It returns the full probability distributions in one forward pass</p><p>There is no token-by-token writing loop. <br>There is no JSON parsing step. <br>The model never gets the chance to invent a value outside your schema.</p><p>TypeSafe trains this system with a method they call <strong>RLCD (Reinforcement Learning for Calibrated Decisions)</strong>. The goal is not “write text that humans like” (RLHF). The goal is “give probabilities that match real outcomes.” A 0.9 should mean the answer is correct about 90% of the time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9rJm0eNfpHOa4Jzy" /></figure><p>Exact internal architecture (parameter count, exact base model, full training details) is still closed. From the outside we only know:</p><p>- Non-autoregressive / parallel sampling<br>- Schema-bound outputs<br>- RLCD training focused on calibration<br>- Hardware-aware design for low latency</p><h4>Noul, Choice, and Score Explained</h4><p>Jev only understands three kinds of questions.</p><p>1. Noul (Yes / No probability)</p><p>Ask a clear true/false question. <br>Jev returns a number between 0 and 1.</p><p>- Near 1.0 → strong yes <br>- Near 0.0 → strong no <br>- Near 0.5 → the model is unsure</p><p>Example: <br>“Does the customer explicitly request a refund?”</p><p>2. Choice (Pick one option)</p><p>Give a list of options (up to 255). <br>Jev returns</p><p>- the winning option<br>- probability for every option<br>- a confidence score</p><p>Example:<br>Which team should handle this ticket? <br>→ billing / technical / sales / other</p><p>3. Score (Ordered scale)</p><p>Give an ordered list of levels (usually 2–10). <br>Jev returns a numeric position that can sit between two levels, plus the full distribution.</p><p>Example levels: <br>0 = Calm <br>1 = Frustrated but civil <br>2 = Very angry</p><p>A score of 1.4 means the model thinks it is between “frustrated” and “very angry.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9B4SfeYedzYR0KRCzrmN9w.png" /></figure><p>You can mix all three types in the same request.</p><h4>Understanding Jev’s Response</h4><p>A typical response looks like this</p><pre>{<br>  &quot;model&quot;: &quot;jev-1.13.0&quot;,<br>  &quot;answers&quot;: {<br>    &quot;department&quot;: {<br>      &quot;type&quot;: &quot;choice&quot;,<br>      &quot;choice&quot;: &quot;billing&quot;,<br>      &quot;confidence&quot;: 0.92,<br>      &quot;probabilities&quot;: {<br>        &quot;billing&quot;: 0.91,<br>        &quot;technical&quot;: 0.07,<br>        &quot;sales&quot;: 0.02<br>      }<br>    },<br>    &quot;is_urgent&quot;: {<br>      &quot;type&quot;: &quot;noul&quot;,<br>      &quot;noul&quot;: 0.87<br>    },<br>    &quot;frustration&quot;: {<br>      &quot;type&quot;: &quot;score&quot;,<br>      &quot;score&quot;: 1.3,<br>      &quot;confidence&quot;: 0.81,<br>      &quot;legend&quot;: {<br>        &quot;0&quot;: &quot;Calm&quot;,<br>        &quot;1&quot;: &quot;Frustrated&quot;,<br>        &quot;2&quot;: &quot;Very angry&quot;<br>      },<br>      &quot;probabilities&quot;: {<br>        &quot;0&quot;: 0.05,<br>        &quot;1&quot;: 0.7,<br>        &quot;2&quot;: 0.25<br>      }<br>    }<br>  },<br>  &quot;usage&quot;: {<br>    &quot;input_tokens&quot;: 412,<br>    &quot;output_tokens&quot;: 48<br>  }<br>}</pre><p>Your code reads these numbers and decides</p><p>- If confidence is high → act automatically <br>- If confidence is low → send to a human or ask another model</p><h4>Multiple Questions in One Request</h4><p>One of the biggest advantages of Jev is that you can ask many questions at once.</p><p>Because everything runs in parallel, asking 10 questions usually costs almost the same time as asking 1 question. You only pay for the extra tokens in the questions themselves.</p><p>This lets you use a pattern called <strong>speculative fan-out</strong></p><p>- Ask every question your code might need<br>- Look at the answers<br>- Only use the ones that are relevant</p><p>For example, you can ask both “Is this a billing issue?” and “How severe is the outage?” in the same call. If it is not a billing issue, you simply ignore the severity answer.</p><p>To show what this looks like in a real production setting, I built a small autonomous FinTech dispute and incident triage gateway in Go.</p><p>Instead of calling a heavy generative language model, waiting several seconds, paying for every output token, and then parsing JSON that might still be malformed, the service talks directly to Jev.</p><p>For every payment dispute or outage webhook it receives, it asks three focused questions at once as<br>a Choice question to route the case to the right team, <br>a Noul question to decide whether it is urgent, <br>a Score question to measure severity.</p><p>Because Jev evaluates all three in a single parallel pass, the whole decision usually comes back in under 400 milliseconds.</p><p>Output tokens are free, so the cost of automated triage drops to roughly $27 per million events. Most importantly, the model can only return values inside the schema we defined, which means zero schema hallucinations and no extra validation or retry loops.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9E57imaxy14ddu6k0iGg6g.png" /></figure><p><a href="https://github.com/kavix/jev-triage-engine">GitHub - kavix/jev-triage-engine</a></p><p>The result is a backend that feels more like a fast, reliable if-statement than a slow chat model.</p><p><a href="https://console.typesafe.ai/playground">https://console.typesafe.ai/playground</a></p><h4>When to Use Jev and When Not To</h4><p>Use Jev when,</p><p>- You need fast classification, routing, or scoring<br>- The possible answers are already known<br>- You want reliable probabilities and confidence<br>- Latency matters (real-time agents, high-volume workflows)<br>- You want to avoid JSON parsing errors</p><p>Do not use when,</p><p>- You need the model to write text, code, or explanations<br>- The task requires multi-step reasoning or tool use in a long chain<br>- You need open-ended creativity<br>- The answer space is completely unknown in advance</p><p>In practice many systems will use both,Jev for the fast structured decisions, and a normal LLM for the parts that need writing or deep reasoning.</p><p>Jev is interesting because it challenges a simple assumption we often make about AI that intelligence always needs to look like text generation.</p><p>Many real-world applications do not need an AI to write a paragraph. they need it to make a fast, structured decision. By turning unstructured state into typed probabilistic outputs, Jev explores a different way to integrate intelligence directly into software.</p><h3>References</h3><ol><li>TypeSafe AI. <em>Introducing System One Models &amp; Jev</em>.<br><a href="https://typesafe.ai/blog/introducing-system-one-models-and-jev">https://typesafe.ai/blog/introducing-system-one-models-and-jev</a></li><li>TypeSafe AI. <em>Jev Documentation</em>.<br><a href="https://docs.typesafe.ai/">https://docs.typesafe.ai/</a></li><li>DataCamp. <em>System One Models and Jev</em>.<br><a href="https://www.datacamp.com/blog/system-one-models-jev">https://www.datacamp.com/blog/system-one-models-jev</a></li><li>Kahneman, Daniel. <em>Thinking, Fast and Slow</em>. Farrar, Straus and Giroux, 2011.</li><li>SalesRLAgent. <em>Reinforcement Learning for Sales Conversion Prediction</em>. arXiv:2503.23303.<br><a href="https://arxiv.org/abs/2503.23303">https://arxiv.org/abs/2503.23303</a></li><li>arXiv. <em>Paper referenced in the discussion of prior decision-model research</em>. arXiv:2510.01237.<br><a href="https://arxiv.org/abs/2510.01237">https://arxiv.org/abs/2510.01237</a></li><li>DeepMostInnovations. <em>Sales Conversion Model — Reinforcement Learning</em>. Hugging Face.<br><a href="https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning">https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning</a></li><li>DeepMostInnovations. <em>SaaS Sales Conversations Dataset</em>. Hugging Face.<br><a href="https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations">https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations</a></li><li>r/LocalLLaMA. <em>I literally built the Jev architecture one year back and completely open-sourced it with model, dataset and paper</em>.<br><a href="https://www.reddit.com/r/LocalLLaMA/comments/1wihgum/i_literally_built_the_jev_architecture_one_year/">https://www.reddit.com/r/LocalLLaMA/comments/1wihgum/i_literally_built_the_jev_architecture_one_year/</a></li><li>Zefan-cai. <em>Open-Jev</em>.<br><a href="https://zefan-cai.github.io/open-jev/">https://zefan-cai.github.io/open-jev/</a></li><li>TypeSafe AI. <em>Jev / System One Video Tutorial</em>. YouTube.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f55ab7f4c4d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sri Lanka’s First Agentic AI Meetup]]></title>
            <link>https://kavix.medium.com/sri-lankas-first-agentic-ai-meetup-6a3e83be519c?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a3e83be519c</guid>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[google-adk]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[web-mcp]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 21:52:55 GMT</pubDate>
            <atom:updated>2026-09-11T21:52:55.882Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M247Yc2gjcwegxHJ8_9zGA.png" /></figure><p>I attended the inaugural Agentic AI Foundation Sri Lanka Chapter Meetup at Hatch Works, Colombo.</p><p>It was a valuable opportunity to learn from industry experts, explore emerging AI technologies, and connect with developers and AI enthusiasts who are passionate about the future of intelligent systems.</p><p>During the meetup I gained insights into Google ADK, Model Context Protocol (MCP), OpenChoreo, Graph RAG, and modern agentic architectures. The biggest shift I noticed is that AI is moving beyond simply generating responses. We are now building systems that can reason, use tools, interact with external systems, and take meaningful actions.</p><p>Here is a quick walk-through of the five sessions and the main ideas that stayed with me.</p><h4><a href="https://medium.com/u/bceb5a3692cc">Ayesha Dissanayaka</a> — MCP Auth Grows Up</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e4o_vnFxy9i_MV9gJqeHYw.png" /></figure><p>Ayesha walked us through the latest MCP specification and why security must be treated as a first-class concern from day one.</p><p>Key points she covered</p><p>- MCP is becoming more mature with a stateless protocol core, header-based routing, WAF filtering, and a format extensions framework.<br>- Authentication is no longer an afterthought. She used a clear hotel key-card analogy to explain OAuth, PKCE, identity as a URL, random secrets generated on the device, and SHA-256 verifiers.<br>- There is no pre-shared relationship between the client and the server. The agent brings its own identity.<br>- Practical advice for people building MCP servers, act as a real OAuth 2.1 Resource Server, publish the rules clearly, return proper 401 (with metadata) and 403 (missing scope), and always validate token signature, audience, and scope.<br>- For people building agents and MCP clients— use PKCE, handle 401/403 gracefully, and sanitize tool inputs.</p><p>Her closing message was strong, if you fund the MCP, consumers will bring their own agents. MCP becomes your sales front. Security is the bottleneck, not a tax.</p><h4><a href="https://medium.com/u/d5eb35aa981b">Isuru Alagiyawanna</a> — Vectors Guess. Graphs Know.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fpQr67cgJVIZbW35AiiiWg.png" /></figure><p>Isuru explained why traditional vector-based RAG is often not enough, especially for questions that need relationships.</p><p>He used a simple family example</p><p>Kamal’s father is Nimal. Nimal’s brother is Sunil. Sunil lives in Kandy. Who is the uncle of Kamal?</p><p>Vector embeddings struggle with this kind of multi-hop reasoning. Graphs handle it naturally because they store explicit relationships.</p><p>His main advice was do not limit yourself to traditional RAG, especially in domains like healthcare where relationships between entities matter a lot. Graph-based approaches give the model real structure instead of just semantic similarity.</p><h4><a href="https://medium.com/u/c1db2763b76e">Nisal Renuja Palliyaguru</a> — Building Agents with Google ADK</h4><p>Nisal focused on how to actually build production-ready agents using Google’s Agent Development Kit (ADK).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DlQXzfO9n4_cQBi_bNMq2w.png" /></figure><p>Important ideas from his session</p><p>- An agent is more than a model call. It has execution flow, memory, and storage choices.<br>- <a href="https://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system">Agentic design patterns</a> control how agents execute tasks using sequential, parallel, and loop patterns<br>- Sub-agents and picking the right model for the job (balancing latency and cost) are practical decisions.<br>- Human-in-the-loop is essential for irreversible actions — ask for verification before the agent does something permanent.<br>- For production, wrap the agent so it can become a service (A2A). <br>- Publish an Agent Card so other agents can discover and talk to it without needing the source code.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0PFAU8XKgdi1_CMjljXuhA.png" /></figure><p>He also shared a demo repository <a href="https://github.com/nisalrenuja/aaif-adk-demo](https://github.com/nisalrenuja/aaif-adk-demo)">https://github.com/nisalrenuja/aaif-adk-demo</a></p><h4>Chamindu Himantha — Building a <a href="https://webmcp.dev/">WebMCP</a> Native Voice Agent with LiveKit</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dvAmYQtEuTET4HNq-58L4A.png" /></figure><p>This was the one of most interactive session for me.</p><p>Chamindu showed how to build a real voice agent that uses <a href="https://webmcp.dev/">WebMCP</a>. He covered the full voice pipeline and the way it Intract with WebMCP.</p><p><a href="https://webmcp.dev/"><strong>WebMCP</strong></a> is a way for websites to provide their functions directly to AI agents as tools. Instead of an AI agent only reading a webpage and trying to understand how to click buttons or fill forms, the website can expose specific actions such as searching products, booking an appointment, checking an application status, or submitting a form.</p><p>The AI agent can then call these tools in a structured and reliable way. This makes AI agents more useful on the web because they can understand what a website can do and perform actions directly, while the website still controls what actions are available.</p><p>Other practical points</p><p>- How to declare tools with name, description, and metadata<br>- Difference between stdio and streamable HTTP (stateless handshake) for remote MCP servers<br>- Attaching tool sources to the agent and even changing the tool list mid-session<br>- How the client backend talks to the MCP server for tools and resources<br>- Pixel-space control and reading markup when the agent needs to interact with a web page</p><p>He was honest that the ecosystem is still maturing and many parts are in trial stage, but the direction is clear, voice agents that can use tools in real time are becoming practical.</p><h4><a href="https://medium.com/u/72cfcfdd38a8">Yashod Perera</a> — Tools Are Intents, Not Endpoints</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8tch8PhMFZiYikzDs9pK8w.png" /></figure><p>This was the deepest and most practical talk of the day. Yashod (OpenChoreo maintainer) shared a real production story of how they reduced their MCP server by 40%.</p><p>Instead of mapping every API endpoint to a tool, they redesigned the tools around <strong>intents</strong>.</p><p>I wrote a full detailed article on this session here [<a href="https://kavix.medium.com/tools-are-intents-not-endpoints-what-i-learned-about-context-budgets-and-tool-design-from-oc-9bc31fef9a9e">Tools Are Intents, Not Endpoints — What I Learned About Context Budgets and Tool Design from OpenChoreo</a>]</p><p>If you are building MCP servers, that article is worth reading.</p><h3>Overall Takeaway</h3><p>AI is evolving from systems that only generate text to systems that can reason, call tools, maintain state, and take real actions. The building blocks we saw MCP, proper auth, Graph RAG, agent patterns, voice pipelines, and careful tool design are all pieces of that shift.</p><p>A special thank you to <a href="https://medium.com/u/24c180217813">Nimesha Jinarajadasa</a>, <a href="https://medium.com/u/da08f4b5da9">Kaveesha Gimhana</a>, and <a href="https://medium.com/u/52ba621bf3e6">Narthana Jananjaya</a> for organising the meetup and bringing together a growing Agentic AI community in Sri Lanka.</p><p>My sincere appreciation to all the speakers for sharing practical knowledge and real experiences.</p><p>Thanks also to @AAIF — Sri Lanka Chapter and <a href="https://medium.com/u/c76afbed391">kodekloud</a> for supporting the event.</p><p>If you attended the meetup or are working on similar topics, I would love to hear what stood out for you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a3e83be519c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tools Are Intents, Not Endpoints - What I Learned About Context Budgets and Tool Design from OC]]></title>
            <link>https://kavix.medium.com/tools-are-intents-not-endpoints-what-i-learned-about-context-budgets-and-tool-design-from-oc-9bc31fef9a9e?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/9bc31fef9a9e</guid>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[openchoreo]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 21:11:20 GMT</pubDate>
            <atom:updated>2026-09-11T21:12:10.832Z</atom:updated>
            <content:encoded><![CDATA[<h3>Tools Are Intents, Not Endpoints — What I Learned About Context Budgets and Tool Design from OpenChoreo</h3><p>I walked into the Agentic AI Foundation Sri Lanka meetup expecting the usual mix of demos and high-level architecture talks. What I got from Yashod Perera’s session on OpenChoreo was something rarer a clear, battle-tested framework for designing tools that agents can actually use without blowing their context budget or making expensive mistakes.</p><p>This is the session that stuck with me the most</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*By2uUbmHXsseHNMvI-ztjA.jpeg" /></figure><h3>Context Is Not Free</h3><p>Yashod opened with a simple but uncomfortable truth<br>context budget and tools are not free.</p><p>Every tool definition, every description, every previous result, and every system prompt instruction competes for the same limited window. When you load too many near-duplicate tools, the model starts making the wrong choice — not because it is hallucinating, but because you forced it to guess.</p><p>He broke the problem into four levers that teams actually control</p><p>1. Context — measured, budgeted, and actively argued about <br>2. Prompt — versioned, evaluated, and iterated for weeks <br>3. Tools — the surface the agent can touch <br>4. Harness — retries, timeouts, middlewares, and guardrails</p><p>Most teams obsess over the first three and treat the fourth as an afterthought. OpenChoreo’s approach was different, they deliberately reshaped the API <strong>into</strong> tools instead of exposing every endpoint as a tool.</p><h3>The Collision Problem</h3><p>Accuracy fell hardest where names collided.</p><p>Examples that kept causing wrong-scope calls,</p><p>- <em>get_component_type</em> vs <em>get_cluster_component_type</em><br>- <em>get_trait</em> vs <em>get_cluster_trait</em><br>- <em>list_prod_builds</em> vs <em>list_staging_builds</em></p><p>These were not hallucinated tools. They were real tools that the model picked in the wrong half of the pair. The request already contained the scope information; the tool names forced the model to re-infer it.</p><h3>The Golden Rule and the Two Collapse Rules</h3><p>Yashod’s golden rule was simple</p><p>You must have very specific differentiation between tools if you want the LLM to pick the correct one.</p><p>Then he gave two practical rules</p><p>Rule 1 — Differs by a noun → collapse<br>Turn the noun into a parameter. <br><em>get_trait</em> / <em>get_cluster_trait</em> becomes one tool with a <em>scope</em> argument.</p><p>Rule 2 — Differs by a verb → keep both<br><em>get_workflow_run</em> and <em>trigger_workflow_run</em> stay separate because the <strong>intent</strong> is different.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TiWw7nV8q2BYpvb-n_JYtg.png" /></figure><p>He summarised it cleanly</p><p>- Nouns are parameters <br>- Verbs are tools <br>- If you can describe the difference between two tools without using a verb, you are looking at a parameter</p><h3>Collapsing Tools ≠ Collapsing Permissions</h3><p>This was one of the most important distinctions in the talk.</p><p>Just because you collapse two tools into one does <strong>not</strong> mean you collapse their permissions. Some users may have permission for only one of the original actions.</p><p>Their solution</p><p>- <strong>Listing tools</strong> show the tool if the user holds either permission <br>- <strong>Calling the tool </strong>enforce the specific permission based on the scope/argument the model actually passed</p><p>This keeps the tool surface clean for the agent while preserving real security boundaries.</p><h3>What They Actually Got Back</h3><p>After applying these principles they measured</p><p>- 40% fewer tools<br>- 27 scope-duplicate tools deleted outright<br>- Survivors grouped into 7 toolsets so a given session only loads a fraction of what is registered<br>- Wrong-scope calls stopped being a major failure category<br>- Descriptions became shorter naturally, when a tool names a clear intent, you no longer need a long sentence explaining which of two near-identical things it is</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-HFK-rkeMz5lUzPBjePK0Q.png" /></figure><p>OpenChoreo doesn’t just expose MCP tools — it ships real built-in agents that use them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9hXM_VvphZSbpIH2jiGz7Q.png" /></figure><h3>The SRE Agent Story</h3><p><a href="https://openchoreo.dev/docs/next/ai/sre-agent/"><em>https://openchoreo.dev/docs/next/ai/sre-agent/</em></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1FBeUqd7mlXlTkdp-1k9-A.png" /></figure><p>This does root cause analysis. When an incident or critical alert happens, it looks at logs, metrics, and traces from your deployments and generates a report with the most likely root causes. It uses LLMs to give intelligent analysis and actionable insights. You can configure it so that only critical alerts automatically trigger RCA (to control costs).</p><p>This matches exactly what Yashod described, the agent gets triggered, investigates using the selected tools and middlewares, and suggests remediation.</p><p>When an incident occurs,</p><p>1. The system notifies the relevant users <br>2. It triggers the SRE agent <br>3. The agent receives only the tools it needs for that investigation (selected tools + middlewares) <br>4. It investigates, reasons, and proposes remediation</p><p>The control plane (OpenChoreo’s abstracted Kubernetes layer) and the MCP server sat in the middle, giving the agent a clean, observable interface instead of raw cluster access.</p><p>This is what “design for both humans and agents” actually looks like in production.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dezm5LBVfu-IzYEkMpZOJQ.png" /></figure><h4><strong>FinOps Agent</strong></h4><p><a href="https://openchoreo.dev/docs/next/ai/finops-agent/">FinOps Agent | OpenChoreo</a></p><p>The FinOps Agent analyses Kubernetes resource usage and cloud cost data. It generates reports with cost optimisation recommendations, cost attribution, and rightsizing suggestions.</p><p>The goal is to help teams reduce cloud spend. Like the SRE Agent, it runs on the observability plane and uses the MCP tools under proper permissions.</p><h4><strong>Portal Assistant</strong></h4><p><a href="https://openchoreo.dev/docs/next/ai/portal-assistant/">Portal Assistant | OpenChoreo</a></p><p>The Portal Assistant is a chat assistant that lives inside the OpenChoreo developer portal (Backstage). It is read-only by design, it can list, get, and query information, but it cannot create, update, delete, or trigger anything.</p><p>Write-style requests are politely declined. It talks to the Control Plane MCP and Observability Plane MCP, and can even call the SRE Agent for deeper analysis when needed. Users interact with it directly from the Logs tab or Build page in Backstage.</p><p>These agents are real examples of what becomes possible once you stop treating every API endpoint as a tool and start designing the MCP surface around actual intents.</p><p>The cleaner toolset (40% fewer tools, grouped into 7 toolsets) makes it much easier for these agents to pick the right tool and stay within the context budget.</p><h3>Final Mental Model</h3><p>Yashod closed with a clean set of principles,</p><p>- APIs are not tools <br>- Tools are intents <br>- Design the way humans already work <br>- Tools must not compromise access control <br>- Agents need the tools they need — not every tool that exists</p><p>Three levers they tuned obsessively (context, prompt, harness). The fourth lever — the tool surface — they reshaped by changing how the underlying API was exposed.</p><h3>Why This Session Mattered to Me</h3><p>Most talks about agentic systems stay at the level of “add tools and the agent becomes powerful.” This one went deeper, it treated the context window as a scarce resource and tool design as a first-class engineering problem.</p><p>If you are building MCP servers or agent harnesses, the practical takeaway is clear</p><ul><li>Stop exposing every endpoint.</li><li>Start exposing intents.</li><li>Measure the cost of every tool definition.</li><li>Never collapse permissions just because you collapsed tools.</li></ul><p>That single shift from “one tool per endpoint” to “one tool per intent” is the highest-leverage change I took away from the entire meetup.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9bc31fef9a9e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside WSO2 — Insights from an Industry Field Visit on Open Source, Culture, and Career Development]]></title>
            <link>https://kavix.medium.com/inside-wso2-insights-from-an-industry-field-visit-on-open-source-culture-and-career-0f2d137463fb?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f2d137463fb</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[ieee]]></category>
            <category><![CDATA[wso2]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 10:05:06 GMT</pubDate>
            <atom:updated>2026-09-08T10:15:01.194Z</atom:updated>
            <content:encoded><![CDATA[<p>I joined a field visit to WSO2 organized by the IEEE Computer Society chapter. Most of us were Software Engineering undergraduates, and the day turned out to be far more than a typical company tour. It was a clear window into how a successful open-source-driven company actually works, and what it takes to thrive there.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/954/1*6Y8WtqyyxjxXD2laszU2xg.jpeg" /></figure><h3>First Impressions — The Energy of a Real Engineering Culture</h3><p>Walking into the WSO2 office, the atmosphere felt immediately familiar and motivating. Seeing teams working as we rode the lift was striking. It was the kind of environment I had heard about but rarely experienced up close focused, collaborative, and quietly energetic. We spent several hours in the conference room, and the culture matched everything I had previously been told. People were approachable, the energy was professinal without being stiff</p><h3>Session 1 — Open Source as a Foundation, Not a Side Project</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4C4hNaujNCXF4UwUSWA5_w.jpeg" /></figure><p>The technical session, led by Mr. <a href="https://medium.com/u/5a25dc777438">Lasantha Samarakoon</a>, focused on WSO2’s architecture and its deep relationship with open source.</p><p>WSO2 is fundamentally a service-based company built on open-source technologies. Enterprise solutions and open-source projects are not separate tracks, they reinforce each other. The company contributes to and builds upon open source while delivering commercial value on top of it.</p><p>The discussion around why students and early career engineers should contribute to open source was particularly practical</p><p>- It creates a public portfolio that speaks louder than a resume.<br>- You learn directly from experienced developers by reading and contributing to real codebases.<br>- It sharpens Git skills and teaches you how large projects are structured.<br>- It helps you stand out in interviews because recruiters can actually see your work.</p><p>Several technical questions came up during the Q&amp;A, including ones about Choreo (its versions and architecture), the reasons behind discontinuing Siddhi, the choice of Go versus Ballerina, and the healthcare accelerator. These exchanges made the session feel less like a presentation and more like a genuine technical conversation.</p><p>As a undergraduate interested in open-source contribution, this session was particularly valuable. It clarified not only why contributing matters, but also how companies like WSO2 integrate open source into their core business model.</p><h3>Session 2 — Culture, Communication, and What They Actually Look For</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z5pJg_ZFm4glOsD4dNpbsg.jpeg" /></figure><p>The culture session, led by Ms. Nikeshala Samanthilake, shifted the focus to how WSO2 operates day to day and what the company values in people.</p><p>A key principle that stood out was simple yet powerful “Treat others as you want to be treated.” Alongside this, the importance of open communication, mutual respect, and working effectively in teams was emphasized. Soft skills, honesty, and a genuine willingness to learn were also highlighted as essential qualities.</p><p>The visit clarified both technical and non-technical priorities for me.</p><p>The day was an excellent introduction not just to WSO2 as a company, but to a way of working that values contribution, transparency, and continuous learning.</p><p>Field visits like this matter. They move abstract career advice into concrete reality. Seeing an engineering culture in action, hearing directly from people who build and ship products, and understanding what open-source contribution actually looks like in a commercial setting is far more useful than any generic “how to get a job” talk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qMdmv8p0J4I-5r4wTsY-IA.jpeg" /></figure><p>Special thanks to the speakers <a href="https://medium.com/u/cc6dba76d00c">Branavan Kuganesan</a> , <a href="https://medium.com/u/5a25dc777438">Lasantha Samarakoon</a>, and Nikeshala Samanthilake and to the IEEE CS chapter for organizing the visit.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0f2d137463fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Day of Auditing Government Websites with Lanka Data Foundation]]></title>
            <link>https://kavix.medium.com/a-day-of-auditing-government-websites-with-lanka-data-foundation-43ed0a28bbdf?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/43ed0a28bbdf</guid>
            <category><![CDATA[open-data]]></category>
            <category><![CDATA[government]]></category>
            <category><![CDATA[digital-governance]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[sri-lanka]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 08:22:26 GMT</pubDate>
            <atom:updated>2026-09-06T09:37:12.530Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_tNcxEYN00NNwOgM.png" /></figure><p>On Saturday, We gathered at the WSO2 Building for something special. Lanka Data Foundation (LDF) had invited volunteers for its first citizen-led <strong>Website Scorecard: Digital Audit Drive</strong>. We simply showed up ready to look closely at Sri Lankan government websites.</p><p>What followed was a practical, eye-opening day of checking more than 50 government sites against clear criteria. We examined availability, SSL certificates, trilingual support, domain hosting, contact emails, MX records, and developer information. The findings will feed into LDF’s public Website Scorecard a tool to measure the state of government websites and track digital readiness over time.</p><p>I walked away with a clearer understanding of what good digital public services require, a stronger appreciation for LDF’s work, and the chance to contribute to their open-source repositories on GitHub.</p><h3>What Lanka Data Foundation Actually Does</h3><p>Lanka Data Foundation is a citizen driven, non partisan, not-for-profit organisation based in Sri Lanka. Its mission is straightforward yet ambitious “<em>make public data open, connected, and useful</em>” so citizens can better understand and participate in governance.</p><p><a href="https://medium.com/u/3d4d7fd8d62d">Sanjiva Weerawarana</a> has written, LDF sources, curates, connects, and delivers verified public data. The goal is to move from raw data to information, knowledge and wisdom. Most government data today sits in isolation PDFs, scattered pages, unstructured files, or outdated sites. When data is structured and linked over time, patterns become visible.</p><p>You can see how government structures change, how institutions evolve, how committees meet (or fail to meet), and how public services perform.</p><p>LDF’s core technical platform is <strong>OpenGIN</strong> (Open General Information Network), an open-source system designed as a digital public good. It can organise and connect any kind of data over time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e6n16dGidznoC3i0hQmMbQ.png" /></figure><p>Anyone another country, a community, or an organisation can adapt it. You can explore the code and related work at <a href="https://github.com/LDFLK">github.com/LDFLK</a>.</p><p>Key projects and tools include</p><p><a href="https://openginxplore.opendata.lk/"><strong>OpenGINXplore</strong></a> — interactive charts, timelines, and visuals of Sri Lanka’s government structures and data.<br><strong>Archives</strong> — searchable government publications by date and category.<br><a href="https://ldflk.github.io/datasets/"><strong>Datasets repository </strong></a>— cleaned, structured JSON datasets from public sources spanning multiple years and ministries.<br><strong>Tools for </strong><a href="https://github.com/LDFLK/gztarchiver"><strong>gazette extraction</strong></a><strong>, organisation charts, and more.</strong></p><p>The Sri Lanka Information Hub at <a href="https://opendata.lk/">opendata.lk </a>brings these together</p><p>LDF members explained the vision simply</p><p>when government data is properly structured, we can organise it by time, institution, and relationships. Citizens, researchers, journalists, and other organisations can then draw clearer insights. The long-term aim is stronger governance through open data not finger-pointing, but giving people the information they need to hold systems accountable.</p><h3>The Digital Audit Day — What We Actually Did</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h-Dg9hByAF5y1GVz5dQItA.jpeg" /></figure><p>The event ran from 8:30 a.m. to 3:30 p.m. We received clear guidelines and a shared evaluation sheet. The rules were practical use a private/fresh browser window for each site, work through the criteria in order, never guess, and flag anything unclear for review.</p><p>We checked five main areas</p><p>1. <strong>Site Availability </strong>— Does the site load? Is it active, down, or redirecting ?<br>2. <strong>SSL Status</strong> — Is the certificate valid, expired, invalid, or missing?<br>3. <strong>The Trilingual Mandate</strong> — Does a language switcher exist? Are at least five key pages fully available in Sinhala, Tamil, and English? Is the translation static or dynamic ?</p><p>Sri Lanka has a clear government rule on this. ICTA’s official Guidelines for Developing Sri Lanka Government Websites say that all content on government websites must be available in Sinhala, Tamil, and English. If something is first written in one language, the other two versions must be properly written or translated and uploaded within seven days. Important content should appear in all three languages at the same time. This means the translations should be human-prepared and permanently published on the site, not generated on the spot by tools like Google Translate.</p><p>4. <strong>Domain Hosting &amp; Official Contact Email </strong>— Is it on a .gov.lk domain? Do contact emails match the domain? Do MX records exist so official email can actually work?<br>5. <strong>Developer Information </strong>— Who designed or developed the site</p><p>We audited well over 50 government websites. The process highlighted both progress and gaps. Some sites were solid. Many still struggle with basic reliability, proper encryption, full trilingual coverage, or professional email infrastructure.</p><p>Small details, clear navigation, readable content, mobile responsiveness, consistent experience matter enormously when a website is the main way citizens reach essential public services.</p><p>One of the biggest takeaways, shared by several of us that day, is that building a system is not enough. It must be usable, accessible, and reliable for everyone who depends on it. Government websites in Sri Lanka still have a lot of work ahead.</p><h3>Networking, Learning, and Open Source</h3><p>The day was also about people. I networked with the LDF team and felt their genuine commitment to data governance. Conversations were practical and forward-looking. I met Mr. Malith (Dr. Malith Jayasinghe), who shared insights on AI agentic systems in enterprise contexts something I plan to write about in a separate piece.</p><p>I have also contributed to <a href="https://github.com/LDFLK">LDF’s GitHub repositories</a>. Their work is fully open source, and they actively welcome both code and data contributions.</p><p>I was happy to contribute a few pull requests myself while Searching about About Open Data Few Months ago. one to the core <a href="https://github.com/LDFLK/OpenGIN/pull/496">OpenGIN platform (#496)</a> and two to the OpenGINXplore frontend (<a href="https://github.com/LDFLK/openginxplore/pull/207">#207</a> and <a href="https://github.com/LDFLK/openginxplore/pull/208">#208</a>). Seeing the collective effort from core engineers to interns and volunteers really shows how open-source communities can create meaningful public impact.</p><h3>Why This Matters</h3><p>Most government data remains unstructured or locked in silos. When LDF structures it, links it across time and institutions, and makes it accessible, ordinary people gain a clearer picture of how the state works. That is the foundation for better questions, better oversight, and stronger democracy.</p><p>The Website Scorecard is one practical step. It turns individual observations into public, trackable data. Over time it can show improvement (or the lack of it) and guide better digital public services.</p><p>I am glad I spent the day contributing. The work is careful, collaborative, and necessary. If you care about open data, transparent governance, or simply better government websites in Sri Lanka, take a look at <a href="https://opendata.lk/">opendata.lk</a>, explore the GitHub organization, and consider joining future efforts.</p><p>There is still a long road ahead but days like this show that citizens, developers, and organizations working together can start building the systems we need.</p><p>Happy to keep learning and contributing.</p><p>Looking forward to what comes next from Lanka Data Foundation.</p><h3>References</h3><ol><li>Wesbite: <a href="https://opendata.lk/">https://opendata.lk</a></li><li>Application: <a href="https://openginxplore.opendata.lk/">https://openginxplore.opendata.lk</a></li><li>Archive: <a href="https://archives.opendata.lk/">https://archives.opendata.lk</a></li><li>OpenGin: <a href="https://ldflk.github.io/OpenGIN/">https://ldflk.github.io/OpenGIN</a></li><li>Research: <a href="https://ldflk.github.io/research">https://ldflk.github.io/research</a></li><li>Datasets: <a href="https://ldflk.github.io/datasets/">https://ldflk.github.io/datasets/</a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yh_8skCrjfBSsDegX2H9Ew.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43ed0a28bbdf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Became an Official Kubernetes Org Member]]></title>
            <link>https://kavix.medium.com/how-i-became-an-official-kubernetes-org-member-417fa4596d8a?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/417fa4596d8a</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 09:55:33 GMT</pubDate>
            <atom:updated>2026-09-02T09:55:33.354Z</atom:updated>
            <content:encoded><![CDATA[<p>Getting an email with the subject <em>@k8s-github-robot has invited you to join the @kubernetes-sigs organization </em>was one of the happiest moments in my developer journey.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E4u2Ko-i0BK1VfJOvf1Y9Q.png" /></figure><p>Contributing to a project as massive as Kubernetes can feel overwhelming at first. Millions of lines of code, hundreds of repositories, and thousands of top engineers worldwide.</p><p>If you are wondering how someone goes from writing their first pull request to becoming an official Kubernetes GitHub Organization Member, this guide is for you. I will walk you through my exact journey, what requirements you need, and practical tips to get there in simple English.</p><h3>What is Kubernetes Community Membership?</h3><p>In the Kubernetes community, there is a clear ladder for contributors:</p><ol><li>Contributor — Anyone who opens an issue or PR.</li><li>Member — An official member of the @kubernetes or @kubernetes-sigs GitHub organizations. You get access to assign issues, trigger CI tests with /ok-to-test, and become a recognized part of the project.</li><li>Reviewer — Assigned to review PRs in specific areas.</li><li>Approver — Has approval rights to merge code into repositories.</li><li>Subproject Lead / Owner — Leads architectural and community decisions.</li></ol><p>Becoming a Member is the first official milestone.</p><h3>The Requirements to Become a Member</h3><p>To apply for membership in kubernetes or kubernetes-sigs, you must meet these criteria</p><ol><li>Enable 2-Factor Authentication (2FA) on your GitHub account.</li><li>Join the community: Subscribe to the kubernetes-dev Google group mailing list.</li><li>Active contributions: Make multiple meaningful contributions (PRs, issues, documentation, code reviews) across Kubernetes subprojects.</li><li>Find 2 sponsors: Two existing Kubernetes reviewers or approvers from two different companies who have actively reviewed your work and agreed to sponsor you.</li><li>Open a Membership Request Issue: Submit an issue in the kubernetes/org repository following their template.</li></ol><h3>My Journey &amp; Contributions</h3><p>I didn’t start with complex core changes on day one. I started with small cleanups, stayed consistent, and gradually moved into core reconciler logic, concurrency, and bug fixes across multiple repositories.</p><p>Here are the key contributions that shaped my journey,</p><h3>1. Contributions to kubernetes-sigs/kueue (Job Queueing &amp; Batch Scheduling)</h3><p>Kueue is a Kubernetes-native workload queueing system. Contributing here helped me understand Kubernetes controllers, custom resource definitions (CRDs), and scheduling deeply.</p><h4>Reconciler &amp; Controller Improvements</h4><ul><li>PR #14721 (StatefulSet Reconciler): Propagated wlName (Workload Name) to reconcileWorkload in the StatefulSet reconciler to avoid redundant lookups and streamline workload sync.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14721">https://github.com/kubernetes-sigs/kueue/pull/14721</a></p><ul><li>PR #14597 (Pod Defaulting Reconciliation): Consolidated StatefulSet Pod defaulting reconciliation logic for cleaner state transitions.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14597">https://github.com/kubernetes-sigs/kueue/pull/14597</a></p><ul><li>PR #14553 (Configurable Workload Popping): Made the number of workloads popped per ClusterQueue cycle configurable, improving queue dispatch flexibility.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14553">https://github.com/kubernetes-sigs/kueue/pull/14553</a></p><h4>Concurrency &amp; Teardown Stability</h4><ul><li>PR #14908 (Simulator Context Race Fix): Discarded logs in the simulator context to prevent race conditions during test suite teardown.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14908">https://github.com/kubernetes-sigs/kueue/pull/14908</a></p><ul><li>PR #14904 (MultiKueue Shutdown Bound): Properly bounded MultiKueue manager shutdown routines during teardown to avoid hanging processes.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14904">https://github.com/kubernetes-sigs/kueue/pull/14904</a></p><h4>Testing, Validation &amp; Bug Hunting</h4><ul><li>PR #14715 (Webhook Validation Tests): Asserted nil admission warnings across webhook validation test suites.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14715">https://github.com/kubernetes-sigs/kueue/pull/14715</a></p><ul><li>PR #14718 (Unit Test Coverage): Added unit tests for IndexPodGroupName to prevent regressions in pod group indexing.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14718">https://github.com/kubernetes-sigs/kueue/pull/14718</a></p><ul><li>PR #14897 (Elastic Job Ungater Tests): Added end-to-end test coverage for elastic job ungater scale-down behaviors.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14897">https://github.com/kubernetes-sigs/kueue/pull/14897</a></p><ul><li>Issue #14611 (Bug Discovery): Reported an edge-case bug where the Pod importer admitted a ResourceFlavor conflicting with the Pod’s nodeSelector.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/issues/14611">https://github.com/kubernetes-sigs/kueue/issues/14611</a></p><ul><li>PR #14546 &amp; PR #14593 (Code Cleanup &amp; Docs): Cleaned up unused functions (resource.IsZero) and sorted CLI importer flag documentation.</li></ul><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14546">https://github.com/kubernetes-sigs/kueue/pull/14546</a></p><p><a href="https://github.com/kubernetes-sigs/kueue/pull/14593">https://github.com/kubernetes-sigs/kueue/pull/14593</a></p><p>Kueue repository — <a href="https://github.com/kubernetes-sigs/kueue">https://github.com/kubernetes-sigs/kueue</a></p><h3>2. Contributions to kubernetes-sigs/cloud-provider-azure</h3><p>Tag Handling &amp; Event Reporting: Fixed case-insensitive duplicate tag handling and improved Azure resource event reporting, preventing duplicate tag conflicts during resource reconciliations.</p><p>Repository — <a href="https://github.com/kubernetes-sigs/cloud-provider-azure">https://github.com/kubernetes-sigs/cloud-provider-azure</a></p><h3>3. Contributions to Core kubernetes/kubernetes</h3><p>Kubelet Status Visibility: Improved init container startup status visibility in Kubelet, making it easier for cluster operators to debug why pods stall during init container phase.</p><p>Repository — <a href="https://github.com/kubernetes/kubernetes">https://github.com/kubernetes/kubernetes</a></p><h3>How to Apply ?</h3><p>Once you have enough contributions and your sponsors agree to support you:</p><ol><li>Go to kubernetes/org on GitHub.</li></ol><p><a href="https://github.com/kubernetes/org/issues/new/choose">https://github.com/kubernetes/org/issues/new/choose</a></p><ol><li>Select “Request: New membership”.</li><li>Fill out the form with your GitHub username, the organization (kubernetes-sigs or kubernetes), list of closed PRs, ongoing PRs, issues filed, and your sponsors’ GitHub handles.</li><li>Your sponsors will comment with a +1 vouching for your contributions.</li><li>A Kubernetes Org Admin (like @jasonbraganza in my case) will review the issue, create an automated PR to add you to the org team configuration, and merge it.</li><li>Check your inbox! You will receive an official invitation from @k8s-github-robot. Accept it, and you’re officially in!</li></ol><h3>4 Tips for Anyone Starting Today</h3><h3>1. Pick One Subproject to Start</h3><p>Don’t try to understand all of Kubernetes on day one. Pick a SIG (Special Interest Group) or subproject that interests you — like Kueue, Cluster API, CLI (kubectl), or Docs.</p><h3>2. Join the Kubernetes Slack</h3><p>Join <a href="https://slack.k8s.io/">https://slack.k8s.io</a></p><p>Subproject maintainers are very active in channels like #kueue, #sig-node, #sig-cli, etc. If you get stuck, people are happy to guide you.</p><h3>3. Start with Tests and Documentation</h3><p>A great way to learn how the codebase works is by adding missing unit tests or improving test coverage. It helps you understand the architecture without breaking existing features.</p><h3>4. Be Patient and Respectful of Maintainers’ Time</h3><p>Maintainers review PRs in their free time or alongside busy day jobs. Be clear in your PR descriptions, test your changes thoroughly before opening a PR, and communicate politely.</p><h3>Special Thanks</h3><p>A huge thank you to everyone who reviewed my code, gave constructive feedback, and sponsored my membership</p><ul><li>Mykola Bobrovskyi (@mbobrovskyi)</li><li>Tenzen (Yuki Iwai) (@tenzen-y)</li><li>Sohan Kunkerkar (@sohankunkerkar)</li><li>Michal Woźniak (@mimowo)</li><li>Jason Braganza (<a href="https://medium.com/u/ef28a6c96ac0">Jason Braganza</a>)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L9k3r4gl4EIQtS7gu9UvEw.png" /></figure><h3>Final Thoughts</h3><p>Open source is all about showing up, being curious, and solving small problems consistently. If I could do it, you definitely can too!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=417fa4596d8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decoy Font vs ChatGPT 5.5 — Can AI Read Hidden Messages?]]></title>
            <link>https://kavix.medium.com/is-decoy-fontcan-confuse-ai-really-f14e89711d9c?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/f14e89711d9c</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[decoy-effect]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 14:35:09 GMT</pubDate>
            <atom:updated>2026-07-20T14:50:25.124Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Decoy Font vs ChatGPT 5.5 — Can AI Read Hidden Messages?</strong></p><p>Recently, I found an interesting experiment called <strong>Decoy Font</strong> by Mixfont. At first, I thought it was just another creative font. But after reading about it, I realized it is actually a clever combination of typography, image processing, and AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JOuE6orEsDrPIAdfwLS35g.png" /></figure><p>As AI models become better at reading text from images, people are also finding new ways to hide information from AI while keeping it readable for humans. Decoy Font is one of those interesting ideas.</p><h3>What is Decoy Font?</h3><p>Decoy Font is a special TrueType Font (TTF) that can display <strong>two different messages at the same time</strong>.</p><ul><li>When you look at the text closely, you see one message (the decoy).</li><li>When you move farther away or squint your eyes, another hidden message appears.</li></ul><p>It feels like an optical illusion, but it is actually based on image processing techniques.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4w7_dIteSeaS5DYh.png" /><figcaption><a href="https://gigazine.net/gsc_news/en/20260717-decoy-font/#gsc.tab=0">https://gigazine.net/gsc_news/en/20260717-decoy-font/#gsc.tab=0</a></figcaption></figure><p>For example</p><p><strong>Up close</strong></p><p>BINGE SHOWS</p><p><strong>From a distance</strong></p><p>SECRET CODE</p><p>The amazing part is that both messages are inside the same letters.</p><h3>How Does It Work?</h3><p>The font uses something called <strong>spatial frequencies</strong>.</p><p>Instead of drawing a normal letter, each character contains two layers.</p><ul><li>A thin, high-frequency outline that AI usually detects first.</li><li>A blurred, low-frequency shape that humans can recognize from a distance.</li></ul><figure><img alt="https://blog.devgenius.io/seeing-the-invisible-how-spatial-frequency-and-fourier-transforms-reveal-hidden-patterns-cebbab6266e3" src="https://cdn-images-1.medium.com/max/1024/0*oZDL_p8hzR5eKU4R.png" /><figcaption><a href="https://blog.devgenius.io/seeing-the-invisible-how-spatial-frequency-and-fourier-transforms-reveal-hidden-patterns-cebbab6266e3">https://blog.devgenius.io/seeing-the-invisible-how-spatial-frequency-and-fourier-transforms-reveal-hidden-patterns-cebbab6266e3</a></figcaption></figure><p>Think about looking at a newspaper photo.</p><ul><li>If you put your face close to it, you only notice the tiny dots.</li><li>If you step back, the full picture becomes clear.</li></ul><p>Decoy Font uses the same idea, but with letters instead of images.</p><h3>Why Does AI Get Confused?</h3><p>Most AI vision models read images by analyzing pixels and edges.</p><p>Since the foreground outlines are much clearer, AI often reads only the decoy message.</p><p>Humans, however, naturally recognize larger patterns when viewing the image from farther away. That allows us to see the hidden message instead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rKbiEkaUDLT-26MyW4XrMA.png" /><figcaption>Qwen3-VL-235B-A22B</figcaption></figure><p>It is an interesting example showing that <strong>humans and AI don’t always “see” images in the same way.</strong></p><h3>The Technology Behind It</h3><p>Decoy Font is based on a well-known image processing technique called <strong>Hybrid Images</strong>.</p><p>One famous example combines the faces of <strong>Albert Einstein</strong> and <strong>Marilyn Monroe</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/380/0*9WZNnz1-DX5nqB9F" /></figure><ul><li>Look closely and you see Albert Einstein.</li><li>Look from far away and you see Marilyn Monroe.</li></ul><p>Decoy Font applies this same concept to text instead of photographs.</p><h3>Can It Really Protect Information?</h3><p>The answer is <strong>partially</strong>.</p><p>It can confuse many AI systems and OCR tools, making it harder for them to instantly read text from screenshots.</p><p>However, it is <strong>not perfect security</strong>.</p><p>A stronger AI model or a specially designed algorithm could eventually detect both layers and recover the hidden message.</p><p>So Decoy Font should be seen as</p><ul><li>an interesting way to hide text,</li><li>a tool that slows down AI reading,</li><li>but <strong>not</strong> a replacement for encryption.</li></ul><p>If someone needs real privacy, encryption is still the correct solution.</p><h3>My Experiment with ChatGPT 5.5</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4BsdcSEYjc3T8VEcZrk4vQ.png" /><figcaption>ChatGPT 5.5</figcaption></figure><p>While reading about Decoy Font, I wanted to test whether modern AI models could actually understand it.</p><p>The original article mentions that some AI models struggled to read the hidden message. So I created my own Decoy Font image and uploaded it to <strong>ChatGPT 5.5</strong>.</p><p>Surprisingly, ChatGPT 5.5 was able to identify <strong>both messages</strong> correctly.</p><ul><li><strong>Visible (decoy) message:</strong> MEDIUMJ</li><li><strong>Hidden message:</strong> KAVINDU</li></ul><p>This was interesting because it shows how quickly AI vision models are improving. A technique that could confuse earlier models is no longer guaranteed to work against newer ones.</p><p>Of course, this doesn’t mean every Decoy Font image can always be decoded. The result depends on factors such as image quality, font size, blur level, and how strongly the hidden message is embedded. However, my small experiment suggests that newer multimodal models have become much better at recognizing these visual tricks.</p><h3>What This Means</h3><p>Decoy Font is still a clever and creative idea, but it should not be considered a secure way to protect sensitive information.</p><p>Instead, it demonstrates an ongoing challenge between <strong>human-designed visual illusions</strong> and <strong>AI perception</strong>. As AI models continue to improve, techniques like Decoy Font become valuable benchmarks for measuring how well machines understand images that intentionally exploit human visual perception.</p><h3>Why I Think This Is Interesting</h3><p>As a Computer Science student, I find this project exciting because it combines several different fields together</p><ul><li>Typography</li><li>Computer Vision</li><li>Image Processing</li><li>Optical Illusions</li><li>Artificial Intelligence</li></ul><p>Usually, we think AI will always become better at reading text.</p><p>This project shows another perspective, what if we design text that is easier for humans than for AI?</p><p>It reminds us that AI still has limitations, especially when human perception and visual illusions are involved.</p><h3>Possible Future Applications</h3><p>Although Decoy Font is still an experiment, I think it could inspire many ideas.</p><p>For example</p><ul><li>Human-only CAPTCHA systems</li><li>Protecting creative work from AI scraping</li><li>Educational demonstrations about computer vision</li><li>Private messages that are difficult for automated systems to read</li><li>Research for testing OCR and AI vision models</li></ul><p>It also raises an interesting question</p><p>Can we design more content that humans understand naturally but AI struggles to interpret?</p><h3>Final Thoughts</h3><p>Decoy Font is one of the most creative AI-related projects I have seen recently.</p><p>It is not about defeating AI. Instead, it shows how understanding human vision and image processing can create clever ways to challenge machine perception.</p><p>As AI continues to improve, techniques like this may become less effective. But they also encourage researchers and developers to think differently about the relationship between typography, computer vision, and artificial intelligence.</p><p>Sometimes, the most interesting innovations come from combining simple ideas from different fields.</p><p>And Decoy Font is a great example of that.</p><p><a href="https://www.mixfont.com/experiments/decoy-font">Decoy Font: A TTF font that hides what you type</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f14e89711d9c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My First Open Source Release Contribution — What OpenChoreo Taught Me]]></title>
            <link>https://kavix.medium.com/my-first-open-source-release-contribution-what-openchoreo-taught-me-96ec100c42de?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/96ec100c42de</guid>
            <category><![CDATA[open-source-contribution]]></category>
            <category><![CDATA[wso2]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[openchoreo]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Sat, 18 Jul 2026 04:46:53 GMT</pubDate>
            <atom:updated>2026-07-18T04:46:53.809Z</atom:updated>
            <content:encoded><![CDATA[<p>A few years ago, I thought open source was simply about fixing bugs or adding new features.</p><p>After contributing to OpenChoreo, I realized that open source is much more than writing code. It is about learning from experienced engineers, understanding how large software projects are built, and improving yourself through collaboration.</p><p>Recently, I was excited to see my contributions included in the <strong>OpenChoreo v1.2.0-rc.1</strong> release. Seeing my GitHub username in the contributors list may look like a small achievement, but for me, it represents months of learning, debugging, reading code, and improving through code reviews.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LdLzHGVYLxxPL9HgzlPmlw.png" /></figure><h3>The Beginning</h3><p>When I first looked at the OpenChoreo repository, I honestly felt lost.</p><p>The project was much larger than anything I had worked on before. There were controllers, APIs, Helm charts, workflows, tests, CI pipelines, and many directories that I had never seen in a single project.</p><p>At first, I thought I needed to understand the whole codebase before contributing.</p><p>I was wrong.</p><p>Instead, I learned to understand one issue at a time.</p><p>One function. One package. One pull request.</p><p>That made everything much more manageable.</p><h3>Every Pull Request Was a Lesson</h3><p>One thing I really appreciate about OpenChoreo is the maintainers.</p><p>They didn’t just review my code and click “Merge.”</p><p>They explained why something should be changed.</p><p>Sometimes they suggested a cleaner approach.</p><p>Sometimes they pointed out edge cases that I had completely missed.</p><p>Sometimes they shared better ways to structure the code to match the project’s standards.</p><p>Those review comments became one of the best learning resources I have ever had.</p><p>Every pull request taught me something new.</p><p>I wasn’t only fixing issues. I was learning how production software is designed, reviewed, tested, and maintained.</p><p>Thank you to all the maintainers who spent their valuable time reviewing my pull requests. Your feedback helped me improve as a developer, and I learned a lot about industrial software development standards through every review.</p><h3>Learning Beyond Writing Code</h3><p>Before this experience, I thought software engineering was mostly about writing code.</p><p>Now I see it differently.</p><p>A good solution is not just one that works.</p><p>It should also be easy to understand, maintain, test, and fit naturally into the existing project.</p><p>Sometimes I spent more time reading existing code than writing new code.</p><p>Sometimes fixing a bug took only a few lines of code after hours of understanding how that part of the system worked.</p><p>That completely changed how I approach programming today.</p><h3>My Contributions</h3><p>During this journey, I had the opportunity to contribute to different parts of OpenChoreo, including workflow validation, API improvements, Helm configuration, CLI enhancements, controller logic, and several bug fixes.</p><ul><li>feat(cli): add resource to occ config context scope by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/4145">https://github.com/openchoreo/openchoreo/pull/4145</a></li><li>feat(helm): enable WebSocket upgrades by default on KGateway by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/4080">https://github.com/openchoreo/openchoreo/pull/4080</a></li><li>fix(helm): gate KGateway TrafficPolicies on gatewayClassName by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/4079">https://github.com/openchoreo/openchoreo/pull/4079</a></li><li>fix(controller): cascade Resource deletions and surface precise finalize statuses by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/3917">https://github.com/openchoreo/openchoreo/pull/3917</a></li><li>feat(helm): make gatewayClassName configurable in control and observability planes by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/3915">#3915</a></li><li>fix(controller): set workflow failed condition on validation failure by <a href="https://github.com/kavix">@kavix</a> in <a href="https://github.com/openchoreo/openchoreo/pull/3877">#3877</a></li></ul><p>None of these contributions were made alone. Every merged pull request was the result of discussions, review comments, suggestions, and guidance from the maintainers.</p><p>That collaboration is what makes open source so valuable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O6orasRMyLpVRtKbiZNw5g.png" /></figure><h3>More Than Just a Green Checkmark</h3><p>Getting a pull request merged always feels good.</p><p>But what stayed with me the most wasn’t the green “Merged” button.</p><p>It was everything I learned before reaching it.</p><p>How to understand an unfamiliar codebase.</p><p>How to communicate with maintainers.</p><p>How to improve code after reviews.</p><p>How to write software that follows existing project standards instead of personal preferences.</p><p>These are lessons that are difficult to learn from tutorials alone.</p><h3>Looking Forward</h3><p>This is only the beginning of my open source journey.</p><p>There is still so much to learn, and I know I am far from being an expert in Go, Kubernetes, or cloud-native technologies. But every contribution gives me another opportunity to improve.</p><p>I’m grateful to the OpenChoreo community for creating such a welcoming environment for new contributors and for sharing their knowledge through every review and discussion.</p><p>I hope to continue contributing, learning from experienced engineers, and giving back to the open-source community that has already taught me so much.</p><p>Thank you to everyone who reviewed my pull requests, answered my questions, and helped me become a better software engineer.</p><p>Here’s to many more contributions ahead.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96ec100c42de" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[weGuard — Building an AI Future for Sri Lankan Agriculture]]></title>
            <link>https://kavix.medium.com/weguard-building-an-ai-future-for-sri-lankan-agriculture-834c7814419e?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/834c7814419e</guid>
            <category><![CDATA[agriculture]]></category>
            <category><![CDATA[ai-expo]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agriculture-technology]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 07:24:37 GMT</pubDate>
            <atom:updated>2026-07-13T07:24:37.141Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OxUt_o4mbu6c9vdo.png" /></figure><p>Some projects start as ideas.</p><p>Some projects start as solutions to problems we see around us.</p><p>For me, <strong>weGuard</strong> was one of those projects.</p><p>Winning the <strong>National Award for AI Future Vision</strong> at <a href="https://www.aiexpo.lk/news/articles/opening-ceremony-spectacular?utm_source=chatgpt.com"><strong>AI Expo 2025</strong></a> in the University category was not just an achievement for our team. It was a moment that reminded us why we started building this solution in the first place.</p><p>A small idea to use artificial intelligence for agriculture became a platform that was recognized on a national stage.</p><h3>The Beginning of weGuard</h3><p>Agriculture has always been one of the most important sectors in Sri Lanka.</p><p>However, farmers continue to face many challenges, especially when identifying plant diseases at the early stages.</p><p>A farmer noticing unusual changes in a crop often has to depend on experience, external advice, or traditional methods to identify the problem.</p><p>But what if technology could provide immediate assistance?</p><p>What if a farmer could simply capture an image of a plant and receive guidance within seconds?</p><p>That question became the foundation of <strong>weGuard</strong>.</p><h3>Introducing weGuard</h3><p>weGuard is an AI-powered solution designed to support farmers by helping them identify crop diseases using artificial intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MPXyf1qVsktYGAN6v0jMDg.png" /></figure><p>Our vision was to bridge the gap between modern AI technology and real-world agricultural challenges.</p><p>Through weGuard, we focused on creating a solution that is</p><ul><li>Accessible</li><li>Practical</li><li>User-friendly</li><li>Helpful for local farmers</li></ul><p>The goal was not just to build another AI application.</p><p>The goal was to create technology that can create a meaningful impact in people’s lives.</p><h3>Building the Solution</h3><p>Developing weGuard was a journey filled with learning, experimentation, and challenges.</p><p>Building an AI-powered agricultural solution requires much more than training a machine learning model.</p><p>We had to think about the complete user experience.</p><p>From collecting and preparing data to developing the application, every part required careful planning.</p><p>Some of the major areas we focused on included,</p><ul><li>Artificial Intelligence and Machine Learning</li><li>Image-based disease detection</li><li>Mobile application development</li><li>User experience design</li><li>Real-world usability</li></ul><p>The biggest challenge was not only achieving technical accuracy.</p><p>It was making sure the technology could actually be useful for the people who need it.</p><h3>Participating in AI Expo 2025</h3><p>The opportunity to present weGuard at <strong>Sri Lanka’s first National AI Expo &amp; Conference 2025</strong> was a memorable experience for our team.</p><p>The event brought together innovators, researchers, companies, and technology enthusiasts to showcase the future of artificial intelligence in Sri Lanka.</p><p>Standing there and presenting our idea among many innovative AI solutions was both exciting and inspiring.</p><p>It was a moment where we could share our vision</p><p>Using AI not only for advanced technology demonstrations, but for solving real problems faced by communities.</p><h3>A National Recognition</h3><p>We are incredibly honored that <strong>weGuard</strong> was awarded the</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*-38j18ChvaF5C6ug" /></figure><h3>National Award for AI Future Vision — University Category — AI Expo 2025</h3><p>This recognition represents much more than an award.</p><p>It represents</p><ul><li>The dedication of our team</li><li>The countless hours spent learning and building</li><li>The support we received from our mentors</li><li>The encouragement from our university community</li></ul><p>Most importantly, it shows that student-driven innovation can create solutions with national-level impact.</p><h3>Team KODEGAS</h3><p>weGuard was developed by <strong>Team KODEGAS</strong> from the University of Kelaniya.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h_DY2IhYXmmXZgUc" /></figure><p>Every discussion, experiment, failure, improvement, and success was a result of teamwork.</p><p>Working together taught us an important lesson,</p><p>Great ideas become reality when people with different skills and perspectives come together.</p><h3>Beyond the Award</h3><p>This award is not the final destination.</p><p>It is a motivation to continue improving.</p><p>There are still many possibilities ahead</p><ul><li>Improving AI accuracy</li><li>Expanding crop support</li><li>Making technology more accessible</li><li>Helping more farmers through digital solutions</li><li>Hardware Improvement</li></ul><p>The future of agriculture will not only depend on traditional knowledge.</p><p>It will also be shaped by technologies like Artificial Intelligence, Machine Learning, and automation.</p><p>We believe AI can become a powerful tool to support communities and create a smarter, more sustainable future.</p><h3>Thank You</h3><p>We would like to express our sincere gratitude to,</p><p><strong>SLTMobitel</strong> and the organizers of AI Expo 2025 for creating a platform where young innovators could showcase their ideas.</p><p>The respected panel of judges for recognizing and believing in our vision.</p><p>The <strong>University of Kelaniya</strong>, our lecturers, mentors, and everyone who supported us throughout this journey.</p><p>And finally, our teammates who dedicated their time, creativity, and passion to build weGuard.</p><h3>The Journey Continues</h3><p>Winning the National Award for AI Future Vision is a proud milestone for us.</p><p>But more importantly, it reminds us of something bigger</p><p>Technology is not only about writing code or building models.</p><p>Technology is about solving problems.</p><p>It is about creating solutions that improve people’s lives.</p><p>With weGuard and future projects, we hope to continue contributing towards building an AI-powered Sri Lanka.</p><p>This is only the beginning.</p><p><strong>Project Website:</strong><br><a href="https://weguard.kodegas.com/">https://weguard.kodegas.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=834c7814419e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MEDUSA 2.0 — Building Sri Lanka’s First Physical ICS Capture The Flag Competition]]></title>
            <link>https://kavix.medium.com/medusa-2-0-building-sri-lankas-first-physical-ics-capture-the-flag-competition-6adce1cfe798?source=rss-a8ffbb5b13df------2</link>
            <guid isPermaLink="false">https://medium.com/p/6adce1cfe798</guid>
            <category><![CDATA[hackathon-organizing]]></category>
            <category><![CDATA[ctf]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ci]]></category>
            <category><![CDATA[sri-lanka]]></category>
            <dc:creator><![CDATA[Kavindu Sachinthe]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 06:23:16 GMT</pubDate>
            <atom:updated>2026-07-13T06:24:24.962Z</atom:updated>
            <content:encoded><![CDATA[<p>There are some university projects you complete because they are part of the curriculum.</p><p>Then there are projects that completely change you.</p><p>For me, <strong>MEDUSA 2.0</strong> was one of them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4fzTqCrauWJkf_Muj4c0bw.jpeg" /></figure><p>Successfully bringing MEDUSA 2.0 to life has been one of the most challenging, stressful, and rewarding experiences of my university journey. Looking back now, it is difficult to believe how far we came — from an idea, countless meetings, sleepless nights, unexpected obstacles, and months of uncertainty — to finally seeing hundreds of participants competing on a platform we built ourselves.</p><p>This article is not just about a cybersecurity competition.</p><p>It is about vision, teamwork, resilience, and what happens when passionate students decide to build something bigger than themselves.</p><h3>Where Everything Started</h3><p>My MEDUSA journey didn’t begin with MEDUSA 2.0.</p><p>Back in <strong>MEDUSA 1.0</strong>, I had the opportunity to contribute as the <strong>Assistant Technical Coordinator</strong>. That experience introduced me to the world of Capture The Flag (CTF) competitions and cybersecurity event organization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4mJutDncpV99uEBwuNkCTQ.png" /></figure><p>Watching participants solve challenges, learning how competitions were designed, and understanding the technical infrastructure behind everything fascinated me.</p><p>Even during MEDUSA 1.0, I already had ideas.</p><p>I kept thinking,</p><p><em>“What if we make the next version even bigger? How about adding ICS Challenges, ML Challenges etc.”</em></p><p><em>“What if we build our own platform?”</em></p><p><em>“What if we introduce challenges that students have never experienced before in Sri Lanka?”</em></p><p>Those thoughts stayed with me.</p><p>When I later became one of the Project Managers for MEDUSA 2.0, I finally had the opportunity to turn many of those ideas into reality.</p><h3>A Vision Bigger Than a Competition</h3><p>MEDUSA itself was born from the vision of <a href="https://www.linkedin.com/in/dishan-ranasinghe/"><strong>Dishan Ranasinghe</strong></a>.</p><p>His dream was never simply to organize another university event.</p><p>The goal was much bigger.</p><p>He wanted to create a platform that could unite cybersecurity enthusiasts across Sri Lanka, encourage students to learn by doing, and build a stronger cybersecurity community.</p><p>That vision became the foundation of everything we worked toward.</p><p>As organizers, our responsibility wasn’t simply to repeat MEDUSA 1.0.</p><p>We wanted to build something that would raise the standard.</p><h3>Taking MEDUSA to the Next Level</h3><p>Building on the success of MEDUSA 1.0, we asked ourselves one question repeatedly:</p><p><strong>“How can we create an experience participants will never forget?”</strong></p><p>The answer wasn’t easy.</p><p>It required months of planning, research, experimentation, and technical development.</p><p>Eventually, we achieved something we are incredibly proud of.</p><p><strong>MEDUSA 2.0 introduced Sri Lanka’s first Physical Capture The Flag (CTF) experience featuring Industrial Control System (ICS) environments.</strong></p><p>Instead of solving only traditional software-based cybersecurity challenges, participants interacted with realistic industrial systems.</p><p>The competition included attack-and-defense scenarios involving:</p><ul><li>Water Dam Control Systems</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pWWUO3GNstjlos6qg0P0Lw.jpeg" /></figure><ul><li>Traffic Light Infrastructure</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mrf_GWTCeGuh2d7b7pnpDA.jpeg" /></figure><p>Alongside these, participants tackled challenges covering:</p><ul><li>Web Security</li><li>Mobile Security</li><li>Binary Exploitation</li><li>Reverse Engineering</li><li>Digital Forensics</li><li>Cryptography</li><li>Open Source Intelligence (OSINT)</li><li>AI Security</li><li>Machine Learning Security</li></ul><p>Our goal was simple.</p><p>We wanted participants to experience cybersecurity as it exists in the real world.</p><h3>Building Our Own CTF Platform</h3><p>One milestone I’m especially proud of was developing our own CTF platform.</p><p>Instead of relying entirely on existing solutions, we designed and built the infrastructure for the first two online rounds ourselves.</p><p>The authentication portals were developed specifically for MEDUSA</p><ul><li>Round 1 Platform</li><li><a href="https://medusa.ecsc-uok.com/round1-auth">https://medusa.ecsc-uok.com/round1-auth</a></li><li>Round 2 Platform</li><li><a href="https://medusa.ecsc-uok.com/round1-auth">https://medusa.ecsc-uok.com/round2-auth</a></li></ul><p>Building the platform taught us lessons far beyond software development.</p><p>We had to think about:</p><ul><li>scalability</li><li>security</li><li>availability</li><li>challenge deployment</li><li>score calculation</li><li>participant experience</li><li>server management</li><li>monitoring</li><li>handling unexpected failures</li></ul><p>Running a live competition is completely different from developing a normal web application.</p><p>When hundreds of competitors are actively attacking, submitting flags, and interacting with the platform simultaneously, every small mistake becomes visible immediately.</p><p>Thankfully, after many rounds of testing and improvements, everything came together successfully.</p><h3>Creating Challenges Instead of Solving Them</h3><p>One of the most enjoyable parts of MEDUSA 2.0 was creating original CTF challenges.</p><p>Instead of only organizing the event, we also designed problems that would challenge participants to think creatively.</p><p>Some of the challenges I contributed include:</p><ul><li>Registration SQL Injection challenge</li><li>Steganography &amp; Digital Forensics challenge</li><li>Perseus Android Exploitation challenge</li></ul><p>Each challenge required balancing realism, educational value, and difficulty.</p><p>Too easy, and participants finish quickly.</p><p>Too difficult, and nobody learns.</p><p>Finding that balance took many iterations, testing, and feedback.</p><p>For several advanced challenges, including container escaping, we were fortunate to receive valuable support from <strong>HashX</strong>, whose technical expertise helped strengthen the overall competition.</p><h3>Behind Every Successful Event Is a Massive Team</h3><p>Many people see only the competition day.</p><p>Very few see what happens during the months before it.</p><p>MEDUSA 2.0 wasn’t built only by the technical team.</p><p>It was the combined effort of many dedicated teams working together.</p><h3>Fundraising Team</h3><p>If there is one group that deserves enormous appreciation, it is our fundraising team.</p><p>Every rupee used to organize MEDUSA came through their dedication.</p><p>For more than <strong>three months</strong>, they travelled across Colombo Fort and many other parts of Sri Lanka, meeting companies, selling pens, finding sponsors, and raising the funds needed to make MEDUSA possible.</p><p>It was physically demanding, mentally exhausting, and often discouraging.</p><p>Yet they never gave up.</p><p>Without them, MEDUSA simply could not have happened.</p><h3>Design Team</h3><p>The visual identity of MEDUSA 2.0 was another highlight.</p><p>From posters and social media campaigns to banners, merchandise, certificates, and branding, the design team created an identity that participants immediately recognized.</p><p>Every visual element reflected the quality we wanted MEDUSA to represent.</p><h3>Event &amp; Logistics Team</h3><p>Organizing a national competition requires much more than writing code.</p><p>The Event and Logistics team handled venue planning, scheduling, equipment, decorations, registrations, coordination, and countless operational details that many participants never noticed.</p><p>Their work ensured everything happened smoothly on the event day.</p><h3>Many Other Teams</h3><p>Behind MEDUSA were also dedicated teams responsible for:</p><ul><li>Technical Operations</li><li>Public Relations</li><li>Financial Management</li><li>External Relations</li><li>Content Writing</li><li>Media Coverage</li><li>Volunteer Coordination</li></ul><p>Every group played a vital role.</p><p>No single team could have delivered MEDUSA alone.</p><h3>Challenges We Never Expected</h3><p>Every large project encounters obstacles.</p><p>MEDUSA had plenty.</p><p>During our fundraising period, <strong>Cyclone Dituva</strong> affected many of our planned activities.</p><p>Several fundraising sessions had to be postponed or cancelled.</p><p>We also encountered administrative challenges within the university that significantly affected our original plans.</p><p>Initially, MEDUSA 2.0 was scheduled for <strong>22 November 2025</strong>.</p><p>We even planned to host the event at <strong>Dharmaloka Hall, University of Kelaniya</strong>.</p><p>Unfortunately, circumstances beyond our control forced us to postpone the competition.</p><p>Many people would have considered stopping there.</p><p>Instead, we adapted.</p><p>After months of waiting, replanning, and continued preparation, MEDUSA 2.0 was finally held successfully on <strong>21 June 2026</strong>.</p><p>Sometimes success is not about avoiding problems.</p><p>It is about refusing to let problems decide the outcome.</p><h3>Leadership Changed My Perspective</h3><p>Being a Project Manager taught me lessons that no classroom could.</p><p>Leadership isn’t about giving instructions.</p><p>It is about listening.</p><p>Making decisions with incomplete information.</p><p>Helping people stay motivated when progress feels slow.</p><p>Accepting responsibility when things go wrong.</p><p>Celebrating the team’s achievements instead of seeking personal recognition.</p><p>I learned technical skills, but more importantly, I learned patience, communication, planning, conflict resolution, and trust.</p><h3>Thank You</h3><p>A special thank you to my Co-Project Manager <a href="https://www.linkedin.com/in/maleesha-indunil-b19163196/">Maleesha Induni</a>l, whose dedication, countless hours of work, and continuous support played a huge role in making MEDUSA 2.0 successful.</p><p>Thank you to every organizing committee member of the <strong>Electronics and Computer Science Club (ECSC)</strong>.</p><p>Thank you to our lecturers, judges, volunteers, media team, <strong>HashX</strong>, <strong>HACK KAP</strong>, sponsors, participants, and everyone who believed in this vision.</p><p>Every contribution mattered.</p><h3>Looking Forward</h3><p>MEDUSA 2.0 is more than a cybersecurity competition.</p><p>It is becoming a platform where students learn, collaborate, challenge themselves, and discover what cybersecurity really looks like beyond classrooms.</p><p>Seeing hundreds of students participate reminded us why we started this journey in the first place.</p><p>This is only the beginning.</p><p>There is still so much more we want to build.</p><p>In a future article, I’ll dive deeper into how we designed Sri Lanka’s first Physical ICS CTF environment — from concept to implementation — and the technical challenges we faced while bringing industrial systems into a student cybersecurity competition.</p><p>Until then, thank you to everyone who became part of the MEDUSA story.</p><p>We built something truly special together.</p><p><strong>Useful Links</strong></p><p><strong>Competition Platforms</strong></p><ul><li>Round 1: <a href="https://medusa.ecsc-uok.com/round1-auth">https://medusa.ecsc-uok.com/round1-auth</a></li><li>Round 2: <a href="https://medusa.ecsc-uok.com/round2-auth">https://medusa.ecsc-uok.com/round2-auth</a></li></ul><p><strong>Open Source Challenge Repositories</strong></p><ul><li>Registration SQL Injection Challenge<br><a href="https://github.com/kavix/medusa-2-registration-sql-injection">https://github.com/kavix/medusa-2-registration-sql-injection</a></li><li>Steganography &amp; Digital Forensics Challenge<br><a href="https://github.com/kavix/medusa-2-steganography-forensics-challenge">https://github.com/kavix/medusa-2-steganography-forensics-challenge</a></li><li>Perseus Android Exploitation Challenge<br><a href="https://github.com/kavix/medusa-2-perseus-andorid-exploitation-challenge">https://github.com/kavix/medusa-2-perseus-andorid-exploitation-challenge</a></li><li>(Container Escaping Challenge for Round 2 was developed and provided by <strong>HashX</strong>.)</li></ul><p><em>MEDUSA 2.0 was never about proving what one person could achieve.</em></p><p><em>It was about proving what passionate students can build when they believe in a shared vision.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6adce1cfe798" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>