<?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 Nethmi Tharushika on Medium]]></title>
        <description><![CDATA[Stories by Nethmi Tharushika on Medium]]></description>
        <link>https://medium.com/@gamagetharushy?source=rss-372d1ae4730f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*sP9-46jya7hZIk6y95ho9A.jpeg</url>
            <title>Stories by Nethmi Tharushika on Medium</title>
            <link>https://medium.com/@gamagetharushy?source=rss-372d1ae4730f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 15:04:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gamagetharushy/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[Systems Thinking: The Skill AI Can’t Replace]]></title>
            <link>https://medium.com/@gamagetharushy/systems-thinking-the-skill-ai-cant-replace-5b44276dcdaa?source=rss-372d1ae4730f------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b44276dcdaa</guid>
            <category><![CDATA[system-thinking]]></category>
            <dc:creator><![CDATA[Nethmi Tharushika]]></dc:creator>
            <pubDate>Thu, 23 Jul 2026 17:30:56 GMT</pubDate>
            <atom:updated>2026-07-23T17:31:35.629Z</atom:updated>
            <content:encoded><![CDATA[<p>There’s a question a lot of us in tech keep circling back to , if AI can write code, debug it, and even explain it back to us, what’s actually left for us to master? For a while, that question genuinely unsettled me. But the more I’ve thought about it, the more I’ve landed on an answer and <strong>it’s not “learn to code faster than the AI.” It’s systems thinking.</strong></p><h3>What Changed</h3><p>In the pre-AI era, a huge part of being valuable as an engineer was knowing syntax, memorizing APIs, and being fast at implementation. AI tools have quietly taken over a big chunk of that work. Ask for a function, a script, a boilerplate setup it’s there in seconds. That used to be a differentiator. Now it’s a commodity.</p><p>What hasn’t been commoditized is the ability to see the <em>whole</em> to understand how a piece of code fits into a larger system, how a small change ripples through dependencies, how a decision made in one part of an architecture creates a problem three layers away. That’s systems thinking, and it’s becoming the actual skill that separates people who use AI well from people who just use AI.</p><h3>Why Systems Thinking Matters More Now, Not Less</h3><p>It’s tempting to think that because AI can generate code, understanding the big picture matters less. It’s actually the opposite:</p><ul><li><strong>AI is good at parts, not wholes.</strong> It can generate a component, a function, a query but it doesn’t inherently know how that piece should behave inside your specific system, with your specific constraints and trade-offs. Someone still has to hold that mental model.</li><li><strong>More output means more integration problems.</strong> When code generation gets faster, the bottleneck shifts. It’s no longer “can we write this,” it’s “does this actually fit together correctly.” That’s a systems-level question.</li><li><strong>Debugging becomes a systems problem, not a syntax problem.</strong> AI-assisted code still breaks often in ways that are harder to trace, because the person reading it didn’t write every line themselves. Fixing that requires understanding how pieces interact, not just how one piece works.</li></ul><h3>Reframing the Skill I’m Building</h3><p>This connects to something I’ve been rethinking in how I approach learning as an engineering student: it’s not about memorizing every framework or every tool. It’s about understanding the architecture underneath how components talk to each other, where the failure points are, what trade-offs a design makes and why.</p><p>AI can help me write the pieces faster. But deciding <em>what pieces should exist</em>, <em>how they should connect</em>, and <em>why the system is shaped the way it is</em> that’s still on me. That’s the part that doesn’t get automated away, because it requires judgment, context, and an understanding of the bigger picture that no single prompt can fully capture.</p><h3>The Real Takeaway</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6NXQkVGUFnR--cCGUB1mVA.png" /></figure><p>In the AI era, the engineers who’ll stand out aren’t the ones who can out-code the model. They’re the ones who can look at a sprawling, messy system technical or otherwise and see the structure inside it. They can ask the right questions: What happens if this fails? What depends on what? Where’s the actual bottleneck?</p><p>That’s systems thinking. And if anything, AI has made it more valuable, not less because it’s the one thing you can’t just prompt your way into. You have to build it, piece by piece, by paying attention to how things connect.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b44276dcdaa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I Stopped Trying to Learn Everything: My Journey Into Micro Learning]]></title>
            <link>https://medium.com/@gamagetharushy/why-i-stopped-trying-to-learn-everything-my-journey-into-micro-learning-30cad31e328c?source=rss-372d1ae4730f------2</link>
            <guid isPermaLink="false">https://medium.com/p/30cad31e328c</guid>
            <category><![CDATA[microlearning]]></category>
            <dc:creator><![CDATA[Nethmi Tharushika]]></dc:creator>
            <pubDate>Thu, 23 Jul 2026 16:36:35 GMT</pubDate>
            <atom:updated>2026-07-23T16:36:35.942Z</atom:updated>
            <content:encoded><![CDATA[<p>When I started my computer engineering degree, I had one goal: become a master of everything. Every language, every framework, every tool that showed up on my radar and I wanted to learn it. In my head, that was what it meant to be “industry ready.” A perfect, complete skill set, with no gaps.</p><p>It didn’t take long for that plan to fall apart.</p><h3>The Problem With Trying to Learn Everything</h3><p>The tech world doesn’t sit still. By the time I got reasonably comfortable with one framework, three new ones had already shown up, and the “must-know” list had grown longer, not shorter. I was spreading myself thin, jumping from one language to the next, one framework to the next, always chasing the next shiny thing and always feeling behind.</p><p>That’s when it hit me: this wasn’t sustainable, and more importantly, it wasn’t even the right approach.</p><h3>Solution : Micro Learning</h3><p>The current trend and the one that actually makes sense is micro learning: <strong>breaking knowledge into small, focused pieces and building understanding gradually, rather than trying to swallow an entire domain in one sitting.</strong></p><p>But the real shift for me wasn’t just about learning in smaller chunks. It was realizing <em>what</em> to focus on. Frameworks and tools change constantly. React today, something else tomorrow. But the theory underneath them data structures, algorithms, how systems actually communicate, the core principles of computer science that doesn’t change nearly as fast. If I understood the foundation deeply, picking up any new framework built on top of it would just be a matter of syntax and specifics, not a whole new mountain to climb.</p><p>So I flipped my strategy:</p><ul><li>Instead of learning every framework, I focused on mastering the fundamentals they’re all built on.</li><li>Instead of trying to consume everything at once, I learned in small, focused sessions one concept at a time, deeply, rather than many concepts shallowly.</li><li>Instead of measuring progress by how many tools I’d touched, I started measuring it by how well I understood the “why” behind them.</li></ul><h3>Facing the Fear of Competition</h3><p>As an undergraduate, it’s easy to feel intimidated by the industry. You see people who seem to know everything, companies with impossibly high expectations, and a competitive job market that makes you feel like you’re always one step behind.</p><p>For a long time, that fear held me back. But I’ve come to realize something important: I don’t need to be the finished product right now. I can grow into someone who’s capable of handling an entire IT department’s worth of problems not by knowing everything today, but by training myself slowly, consistently, and being willing to fail along the way.</p><p><em>Failure, it turns out, isn’t the opposite of progress. It’s part of it.</em></p><h3>What I’ve Learned So Far</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/392/0*vYQHVDu861KZpLXE" /></figure><p>Micro learning didn’t just change <em>how</em> I study it changed how I think about competition, growth, and what it actually means to be “skilled.” Instead of chasing a finish line that keeps moving, I’m building a foundation that will hold up no matter what new framework comes along next.</p><p>If you’re a student feeling the same pressure I did to know it all, right now my advice is simple: slow down, focus on the fundamentals, learn in small consistent steps, and let the failures teach you instead of scare you. The industry doesn’t need another person who’s memorized ten frameworks. It needs people who understand <em>why</em> those frameworks work the way they do.</p><p><em>That’s the version of myself I’m building, one small step at a time.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30cad31e328c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Self-Supervised Learning: How AI Learns to Teach Itself]]></title>
            <link>https://medium.com/@gamagetharushy/self-supervised-learning-how-ai-learns-to-teach-itself-c0581261134c?source=rss-372d1ae4730f------2</link>
            <guid isPermaLink="false">https://medium.com/p/c0581261134c</guid>
            <category><![CDATA[pre-training]]></category>
            <category><![CDATA[generative-methods]]></category>
            <category><![CDATA[contrastive-methods]]></category>
            <category><![CDATA[fine-tuned-model]]></category>
            <category><![CDATA[self-supervised-learning]]></category>
            <dc:creator><![CDATA[Nethmi Tharushika]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 02:28:00 GMT</pubDate>
            <atom:updated>2026-07-17T02:28:01.834Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine handing a child a thousand jigsaw puzzles with no picture on the box, no instructions, and no answer key. Just pieces and the challenge of figuring it out. By the time the child finishes, they understand shapes, patterns, and spatial relationships far better than one who only watched someone else solve them. The struggle itself was the lesson .</p><p>This is exactly how self-supervised learning (SSL) works. It’s a type of machine learning where AI models teach themselves by solving puzzles created from their own data without needing humans to provide labels or answers .</p><h3><strong>The Problem With Traditional Learning</strong></h3><p>To understand why SSL is so exciting, let’s look at how AI normally learns.</p><p><strong>Supervised learning</strong> is like a student who only learns with flashcards and answer keys. To train a model to recognize cats versus dogs, you need thousands of images, each carefully labeled “cat” or “dog” by humans. The model studies these examples, learns the patterns, and then tries to apply them to new images .</p><p>This approach has a major problem: labeling data is expensive, time-consuming, and often requires expert knowledge. In medical imaging, for instance, it can take radiologists up to an hour to annotate a single scan which is why countless medical images exist without useful labels .</p><h3><strong>The Self-Supervised Solution</strong></h3><p>Self-supervised learning takes a different approach. Instead of relying on human labels, the model creates its own “homework” from the raw, unlabeled data itself .</p><p>The key insight is that the model learns to predict missing or transformed parts of the input data. The data supervises itself labels emerge from structure, not from human effort .</p><h3><strong>How It Works in Practice</strong></h3><p><strong>Text: Fill in the Blanks</strong></p><p>In language models like BERT, self-supervised learning works through <strong>masked language modeling</strong>. The model is given sentences where 15% of the words are replaced with a [MASK] token, and it must predict the original words .</p><p>For example:<br>- Input: “The [MASK] is in the garden”<br>- The model learns to predict: “cat,” “dog,” or “child”</p><p>This forces the model to deeply understand grammar, context, semantics, and world knowledge. It’s not just becoming a word-guessing machine; it’s building rich internal representations of language .</p><p><strong>Images: The Rotation Game</strong></p><p>For images, one clever self-supervised task is <strong>rotation prediction</strong>. The model is shown an image rotated by 0°, 90°, 180°, or 270° and must identify which rotation was applied .</p><p>To succeed, the model must recognize objects and their orientations. It learns to identify heads, noses, and eyes, and understand their relative positions without ever being told what a “head” or “nose” is .</p><p><strong>Images: The Jigsaw Puzzle</strong></p><p>Another approach treats images like jigsaw puzzles. The model takes shuffled patches from an image and must place them back in the correct positions . To do this, it must understand the spatial context and structure of objects in images.</p><p><strong>Images: Colorization</strong></p><p>The model takes a grayscale image and learns to color it. Since the original color version exists in the training data, the model can verify if it got it right without any human labels .</p><p><strong>Audio-Visual: Matching Sound and Sight</strong></p><p>In video, self-supervised learning can be used for <strong>audio-visual correspondence</strong>. The model sees a video clip and hears audio, then must determine if they match. For example, does this guitar-playing video correspond to this guitar melody? The model learns the relationship between visual and audio events without any labels .</p><h3><strong>Why This Matters</strong></h3><p><strong>1. It Scales With Data</strong></p><p>Supervised learning is limited by how much labeled data you can collect. Self-supervised learning is limited only by how much raw data exists which is essentially unlimited. Every webpage, every book, every image, every video becomes a training signal .</p><p><strong>2. It Learns Better Representations</strong></p><p>The prediction tasks in self-supervised learning aren’t the real goal they’re vehicles. When BERT learns to predict masked words, it develops powerful internal representations that capture grammar, semantics, and context. These representations can then be transferred to other tasks with minimal additional training .</p><p><strong>3. It’s Behind the AI Revolution</strong></p><p>Virtually every high-performing language model of the past several years BERT, GPT, RoBERTa, and their successors is built on self-supervised pre-training . GPT-3, with its 175 billion parameters, was trained entirely using self-supervised causal language modeling on unlabeled internet text .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/917/1*uQMgubftiFr3J6bezoWa-A.png" /></figure><h3><strong>The Two Main Approaches</strong></h3><p><strong>Generative Methods: Predict What’s Missing</strong></p><p>These methods train models to reconstruct hidden or corrupted parts of the input:<br>- Masked Language Modeling (BERT): Predict missing words<br>- Masked Autoencoders (vision): Reconstruct missing image patches</p><p><strong>Contrastive Methods: Learn What’s Similar and Different</strong></p><p>These methods pull similar examples together in the model’s “understanding space” and push dissimilar ones apart:<br>- SimCLR takes a single image, applies two different random transformations (like rotation and cropping), and trains the model to recognize they’re the same image <br>- MoCo uses a queue of negative examples to learn from more comparisons without needing enormous batch sizes</p><h3><strong>From Pre-training to Real-World Use</strong></h3><p>Self-supervised learning typically works in two phases:</p><p><strong>Phase 1: Pre-training.</strong></p><p>The model trains on massive amounts of unlabeled data using pretext tasks (like predicting masked words or rotated images). This builds powerful, general-purpose representations .</p><p><strong>Phase 2: Fine-tuning.</strong></p><p>The pre-trained model is then adapted to specific tasks using a small amount of labeled data. For example, BERT pre-trained on text can be fine-tuned on sentiment classification with just a few hundred labeled examples and outperform models trained from scratch on tens of thousands .</p><h3><strong>Real-World Applications</strong></h3><p>1.Natural Language Processing<br>2.Computer Vision<br>3.Medical Imaging<br>4.Robotics</p><h3><strong>References</strong></h3><ul><li>Zhang, J., Yang, L., &amp; Sajjadi Mohammadabadi, S. M. (2025). A survey on self-supervised learning: Recent advances and open problems. <em>Neurocomputing</em>, *655*, 131409.</li><li>A survey on self-supervised methods for visual representation learning. (2025). <em>Machine Learning</em>, *114*, 111.</li><li>A survey on design choices for self-supervised learning in computer vision. (2026). <em>Artificial Intelligence Review</em>, *59*, 112.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c0581261134c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 2010s: How NLP Learned to Think in Vectors]]></title>
            <link>https://medium.com/@gamagetharushy/the-2010s-how-nlp-learned-to-think-in-vectors-e6f12ea2e41d?source=rss-372d1ae4730f------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6f12ea2e41d</guid>
            <category><![CDATA[neural-embedding]]></category>
            <category><![CDATA[word2vec]]></category>
            <category><![CDATA[gloves]]></category>
            <category><![CDATA[contextual-embedding]]></category>
            <category><![CDATA[nlp]]></category>
            <dc:creator><![CDATA[Nethmi Tharushika]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 17:23:52 GMT</pubDate>
            <atom:updated>2026-07-15T17:23:52.703Z</atom:updated>
            <content:encoded><![CDATA[<p>If you had to pick one idea that defined natural language processing in the 2010s, it wouldn’t be a single model , it would be a shift in <em>representation</em>. At the start of the decade, words were mostly treated as discrete symbols. By the end of it, words, sentences, and even entire documents were points floating in continuous vector space, positioned so that meaning became something you could measure with arithmetic. This is the story of how that happened.</p><h3>Where we started: sparse and symbolic</h3><p>Before the 2010s, most NLP systems represented words with <strong>one-hot vectors, </strong>huge, mostly-zero vectors where each word got its own dimension. A vocabulary of 50,000 words meant 50,000-dimensional vectors, and every word was equally “far” from every other word. The word “dog” had no more in common with “puppy” than it did with “spreadsheet,” at least as far as the representation was concerned.</p><p>They worked reasonably well for tasks like document classification, but they had no real notion of similarity or meaning baked into the representation itself. Any relationships had to be learned downstream, from scratch, for every new task.</p><h3>The groundwork: neural embeddings before word2vec</h3><p>Dense vectors didn’t spring fully formed from word2vec, they came out of over a decade of neural language modeling research that the 2010s inherited and scaled up.</p><p>The key ancestor is Bengio et al.’s <strong>neural probabilistic language model</strong> (2003), which trained a feedforward neural network to predict the next word in a sequence. As a side effect of that training, each word in the vocabulary was assigned a learned, dense vector- an embedding and the network’s hidden layers combined those vectors to make predictions. This was arguably the first time “map words to dense vectors, then let a neural net operate on them” appeared as a coherent recipe, though at the time it was seen mainly as a language modeling technique, not a general-purpose representation tool.</p><p><strong>Collobert and Weston</strong> pushed this further in the late 2000s and into a widely cited 2011 paper, showing that a single neural network, trained with a mostly unsupervised objective over large amounts of raw text, could produce word embeddings useful across many downstream NLP tasks at once part-of-speech tagging, chunking, named entity recognition, semantic role labeling. This “one embedding, many tasks” framing anticipated the transfer-learning mindset that would come to dominate NLP later in the decade.</p><p>What word2vec did, in this light, wasn’t invent the idea of neural embeddings it stripped the architecture down to something radically simpler and faster (a shallow network, sometimes no hidden layer at all) so that training could scale to billions of words on ordinary hardware. That efficiency, more than novelty, is what let dense vectors go from a research curiosity to a standard tool almost everyone could use.</p><h3>The turning point: word2vec (2013)</h3><p>In 2013, a small team at Google released <strong>word2vec</strong>, and the field’s default representation for words changed almost overnight. The idea, distilled from decades-old distributional semantics (“you shall know a word by the company it keeps”), was elegant: train a shallow neural network to predict a word from its context (or vice versa), and use the resulting weight matrix as a dense, low-dimensional vector for each word typically 100 to 300 dimensions instead of 50,000.</p><p>What made word2vec a cultural moment in ML, not just a technical improvement, was that the resulting vector space had <em>structure</em>. The famous example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JZd2W5yZ4iPLoVD0.png" /><figcaption>vector(“king”) — vector(“man”) + vector(“woman”) ≈ vector(“queen”)</figcaption></figure><p>Analogies, clustering by topic, and similarity by cosine distance all just… worked, as an emergent property of training on raw text. No one explicitly told the model that gender or royalty were meaningful axes, they fell out of the statistics of language use.</p><h3>Refinements: GloVe and the embedding ecosystem</h3><p>Shortly after, Stanford’s <strong>GloVe</strong> (Global Vectors, 2014) offered an alternative recipe: instead of a predictive, sliding-window objective like word2vec, GloVe factorized a global word co-occurrence matrix directly. In practice, the two approaches produced embeddings with similar qualitative behavior, and both became standard building blocks.</p><p>This kicked off an ecosystem. Pretrained embeddings became a commodity, you could download a file of vectors trained on billions of words and drop it into the first layer of almost any model, replacing random initialization with something that already encoded useful structure. Papers explored subword embeddings (<strong>fastText</strong>, 2016) to handle rare and out-of-vocabulary words by building word vectors from character n-grams, and researchers began probing embeddings for the biases they inherited from training data , an early and important thread of what would become a much larger conversation about fairness in NLP.</p><h3>The catch: one vector per word</h3><p>For most of the decade, though, these embeddings shared a fundamental limitation: <strong>one vector per word, regardless of context</strong>. The word “bank” got a single representation whether it appeared next to “river” or “loan.” This was a static, dictionary-like view of meaning useful, but clearly not how meaning actually works in language, where context does enormous work.</p><p>Sentence and document-level representations were also an active area approaches like doc2vec extended the same intuition upward but they inherited the same context-blindness at their core.</p><h3>The pivot: contextual embeddings</h3><p>The back half of the decade fixed this. <strong>ELMo</strong> (2018) used bidirectional LSTMs trained as language models to produce word representations that depended on the entire sentence like “bank” near “river” and “bank” near “loan” now got genuinely different vectors. This was a conceptual leap: representation was no longer a fixed lookup table but the <em>output of a model</em>, computed fresh for every context.</p><h3>Why this mattered</h3><p>Looking back, the throughline of the decade is a steady move away from hand-engineered features and symbolic representations toward <strong>learned, dense, continuous representations </strong>first static, then contextual with each step making downstream tasks easier and requiring less task-specific engineering. This is also the foundation the current era of large language models sits on: the vector space, not the symbol, became the fundamental unit of meaning in NLP.</p><p>By 2019, “just use embeddings” had gone from a research idea to an assumption so basic it was barely worth stating which is usually the sign that a genuine paradigm shift has happened.</p><h3>References</h3><ul><li>Bengio, Y., Ducharme, R., Vincent, P., &amp; Jauvin, C. (2003). A Neural Probabilistic Language Model. <em>Journal of Machine Learning Research</em>, 3, 1137–1155.</li><li>Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., &amp; Kuksa, P. (2011). Natural Language Processing (Almost) from Scratch. <em>Journal of Machine Learning Research</em>, 12, 2493–2537.</li><li>Mikolov, T., Chen, K., Corrado, G., &amp; Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. <em>arXiv preprint</em> arXiv:1301.3781.</li><li>Mikolov, T., Sutskever, I., Chen, K., Corrado, G., &amp; Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionality. <em>Advances in Neural Information Processing Systems (NeurIPS) 26</em>, 3111–3119.</li><li>Bojanowski, P., Grave, E., Joulin, A., &amp; Mikolov, T. (2017). Enriching Word Vectors with Subword Information. <em>Transactions of the Association for Computational Linguistics</em>, 5, 135–146.</li><li>Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., &amp; Polosukhin, I. (2017). Attention Is All You Need. <em>Advances in Neural Information Processing Systems (NeurIPS) 30</em>.</li><li>Devlin, J., Chang, M.-W., Lee, K., &amp; Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. <em>arXiv preprint</em> arXiv:1810.04805.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e6f12ea2e41d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Our Journey Began: Building Yuwathi, a Safety App Inspired by Real Voices]]></title>
            <link>https://medium.com/@gamagetharushy/how-our-journey-began-building-yuwathi-a-safety-app-inspired-by-real-voices-dcc8a3ffceec?source=rss-372d1ae4730f------2</link>
            <guid isPermaLink="false">https://medium.com/p/dcc8a3ffceec</guid>
            <dc:creator><![CDATA[Nethmi Tharushika]]></dc:creator>
            <pubDate>Sat, 06 Dec 2025 18:11:06 GMT</pubDate>
            <atom:updated>2025-12-06T18:11:06.146Z</atom:updated>
            <content:encoded><![CDATA[<p>Before developing <strong>Yuwathi</strong>, our women’s safety app, we wanted to understand the real fears, challenges, and expectations that women face every single day. That’s why we conducted a <strong>mini-survey, </strong>not as a formality, but as the <em>foundation</em> of our entire project.</p><p>What we discovered changed everything.</p><p>The survey helped us see women’s safety from a deeply human perspective. It revealed stories of fear, silent struggles, and unspoken concerns that women experience when traveling, commuting, working late, or simply walking alone. These insights shaped every feature and design decision behind Yuwathi.</p><h4>Women’s Safety in 2025: What Our Survey Reveals About Fear, Harassment &amp; the Need for Better Protection</h4><p>Women’s safety has become one of the most urgent social issues of our time. Whether walking home after class, commuting to work, or traveling alone, many women still feel unsafe in public spaces, and our latest <em>Women’s Safety &amp; Awareness Survey</em> proves it.</p><h4>Who Responded to the Women’s Safety Survey?</h4><p>Our survey received responses from both women and men, with a majority falling in the <strong>18–24 age group</strong>. This demographic moves through busy public spaces, daily buses, rideshares, university campuses, and workplaces, making their insights especially critical.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/718/1*OJvQWFTQObtMHKawZSfW8w.png" /><figcaption>Graph of responders’ age distribution</figcaption></figure><p>The participation of male respondents was also encouraging, highlighting that <strong>women’s safety is not only a women’s issue</strong>, but a shared social responsibility.</p><h4>Women Still Feel Unsafe-Even During the Day</h4><p>One of the strongest patterns from the survey was the emotional discomfort women feel while traveling alone. Many shared feelings of:</p><ul><li>Being followed</li><li>Being stared at or sexualized</li><li>Avoiding public transport</li><li>Anxiety about unfamiliar environments</li><li>Fear of harassment during evening or early-morning hours</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*MRHqPimhacx5PV4IaYOgIQ.png" /><figcaption>Responses of males</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/1*IYBIZifkcKWO8teDkUiPug.png" /><figcaption>Responses of females</figcaption></figure><p>These responses reflect a widespread reality: <strong>public spaces often do not feel safe for women</strong>, regardless of time or location.</p><blockquote><em>“I avoid going out alone unless I absolutely have to.”<br> — Survey Respondent</em></blockquote><p>This sentiment appeared repeatedly across answers, showing the psychological burden women carry daily.</p><h4>What Safety Tools Do Women Actually Want?</h4><p>A major portion of the survey focused on what features would genuinely help women feel safer. The top preferences included:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/915/1*9ToirgleUaUKUseh-gwM6A.png" /><figcaption>Responses of the survey</figcaption></figure><p>Most Requested Safety Features:</p><ul><li><strong>Instant SOS button</strong></li><li><strong>Live location sharing</strong> <strong>with trusted contacts</strong></li><li><strong>Voice-activated emergency help</strong></li><li><strong>One-tap access to emergency hotlines</strong></li><li><strong>Educational content about legal rights</strong></li><li><strong>Anonymous reporting options</strong></li></ul><p>Women also expressed <strong>a need for tools that work even when their phones are locked or hidden</strong>, especially during panic situations.</p><h4>Why Women Don’t Report The Silent Crisis</h4><p>Nearly all respondents agreed that women are hesitant to report harassment. The reasons reveal a critical problem:</p><p>Top Reasons Women Don’t Report:</p><ul><li>Fear of not being believed</li><li>Lack of knowledge about reporting procedures</li><li>Emotional trauma or embarrassment</li><li>Social stigma</li><li>Belief that “nothing will happen anyway.”</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*0yOY2O_hCoM5nh0pYj8GbA.png" /><figcaption>Responses of males</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/742/1*oXbo4T538tAS_fDXZ3H_iQ.png" /><figcaption>Responses of females</figcaption></figure><p>One respondent summarized it perfectly:</p><blockquote><em>“I didn’t report it because I didn’t know how.”</em></blockquote><p>This highlights a major knowledge gap: <strong>women need accessible legal awareness</strong>, not just safety tools.</p><h4>Awareness and Education Are Just as Important as Safety Apps</h4><p>Many participants admitted they were not fully aware of:</p><ul><li>What laws protect women</li><li>What constitutes punishable harassment</li><li>How to file a complaint</li><li>What evidence to collect</li><li>Whom to contact in an emergency</li></ul><p>Improving women’s safety means improving <strong>legal literacy</strong>.</p><p><em>Knowledge = empowerment</em></p><p>An app or system that combines both safety features and simplified legal guides could massively increase reporting and confidence.</p><h4>Would Women Pay for Premium Safety Features?</h4><p>When asked if they would pay a small fee for advanced features, many respondents answered:</p><ul><li><strong>“Maybe”</strong></li><li><strong>“If the features are trustworthy.”</strong></li><li><strong>“Only if it’s affordable”.</strong></li></ul><p>This shows that while women value safety, they expect transparency, reliability, and <strong>real value</strong> before paying for upgrades.</p><p>Safety is not a luxury; it is a <strong>basic need</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/735/1*cxfTcP5EK4gb3CEmIQSYIA.png" /><figcaption>Responses of the survey</figcaption></figure><h4><em>What the Survey Really Tells Us</em></h4><p>Even without long paragraphs of feedback, the answers shared a powerful message:</p><p><strong>Women want to feel safe. Women want to feel supported. Women want change.</strong></p><p>This is a reminder that:</p><ul><li>Women’s safety is a daily concern</li><li>Society must listen to women’s experiences</li><li>Technology, education, and community action must work together</li></ul><h4>Moving Forward: Why Surveys Like This Matter</h4><p>Surveys like ours help uncover:</p><ul><li>Real fears women experience</li><li>Gaps in public infrastructure</li><li>Missing safety features</li><li>Emotional burdens women carry</li><li>How technology can bridge safety gaps</li></ul><p>Every insight pushes us one step closer to creating a world where women can move freely, without fear.</p><h4>Final Note: Women’s Safety Should Never Be Optional</h4><blockquote>Every woman deserves the freedom to walk, travel, work, study, and live without fear.<br>Every girl deserves to grow up confident, not cautious.<br>Every family deserves peace of mind knowing their loved ones are protected.</blockquote><p>Our survey is not just data; it is a <strong>call to action</strong>.</p><p><strong>Based on what we learned, we created powerful, practical safety tools inside Yuwathi. Here are the solutions we developed:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/495/1*3CkPqBBUJgDYUdDnBQoc-A.jpeg" /><figcaption>Features of Yuwathi</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dcc8a3ffceec" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>