<?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 Lorraine Eyinda on Medium]]></title>
        <description><![CDATA[Stories by Lorraine Eyinda on Medium]]></description>
        <link>https://medium.com/@lorraine.atsulu?source=rss-10de6468dd9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*88IfXlbgWQg_VOthFkbeAA.jpeg</url>
            <title>Stories by Lorraine Eyinda on Medium</title>
            <link>https://medium.com/@lorraine.atsulu?source=rss-10de6468dd9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 12:40:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lorraine.atsulu/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[From Uncertainty to Intelligence: How Probability Powers AI]]></title>
            <link>https://medium.com/@lorraine.atsulu/from-uncertainty-to-intelligence-how-probability-powers-ai-cab1ed5dc552?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/cab1ed5dc552</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[probability]]></category>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Sat, 18 Apr 2026 08:59:34 GMT</pubDate>
            <atom:updated>2026-04-18T12:18:10.696Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Understanding how AI makes decisions in a world full of unknowns</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HlELdMEdAmiTSLxGzSE2rw.png" /></figure><h3><strong>We live in a world full of uncertainty</strong></h3><p>From weather forecasts to financial markets, from medical diagnoses to everyday decisions, we are constantly dealing with situations where outcomes are not guaranteed. We rarely know exactly what will happen, we only know what is likely. But how do we make decisions in such a world?</p><p>And more importantly, how do machines do it?</p><h3><strong>Uncertainty in AI</strong></h3><p>Uncertainty is a fundamental challenge in real-world decision-making.</p><p>In many real-world scenarios, we simply don’t have complete information. And even when we do, the sheer complexity of the systems involved makes exact prediction impossible.</p><p>Modern AI systems often can’t tell when they don’t know something.</p><p>Unlike human experts, who can express uncertainty, AI models can produce outputs that appear highly confident, even when they are wrong.</p><p>In high-stakes domains, this becomes dangerous:</p><ul><li>In healthcare, an AI model might incorrectly diagnose a disease with high confidence, leading to harmful treatment decisions.</li><li>In self-driving cars, failing to recognize an unfamiliar road sign or weather condition can cause accidents.</li><li>In legal applications, an AI recommending bail or sentencing without quantifying its certainty could amplify biases.</li></ul><p>Without explicit uncertainty estimates, users lack critical information about the model’s trustworthiness.</p><h3><strong>Types of Uncertainty in AI</strong></h3><p>AI systems face two main types of uncertainty.</p><ul><li><strong>Data uncertainty (Aleatoric)</strong> arises from noise or ambiguity in the data. This could come from noisy measurements, overlapping classes, or natural variability in the real world. This type of uncertainty is inherent and cannot be eliminated.</li><li><strong>Model uncertainty (Epistemic)</strong> arises from limited knowledge. It occurs when a model has not seen enough data or does not fully understand the underlying patterns.</li></ul><p>In practice, deep neural networks often display both types of uncertainty: they can be overly confident on familiar test data, yet become uncertain when faced with inputs that differ from what they were trained on.</p><p>This highlights the need for methods that can both measure and manage uncertainty.</p><blockquote>Uncertainty does not arise because the system is flawed, it arises because the world is complex and constantly evolving.</blockquote><h3><strong>So, how do we handle uncertainty?</strong></h3><p><strong><em>Probability</em></strong><em> </em>allows us to move from <em>“I don’t know”</em> to <em>“I know how likely this is.” . </em>It makes uncertainty measurable.</p><p>It allows both humans and machines to make informed decisions, even when the full picture is unknown.</p><h3><strong>How AI handles uncertainty in practice</strong></h3><p>Now that we understand what uncertainty is, the next question becomes: <em>How do AI systems actually deal with it?</em></p><p>In practice, AI models don’t try to eliminate uncertainty , they work with it. They do this in a few intuitive ways, among many others used in practice.</p><p><strong>1.Predicting probabilities instead of answers</strong></p><p>In many machine learning systems, predictions are expressed as probabilities.</p><p>Instead of giving a single answer, models assign likelihoods to different possible outcomes.</p><p>For example, in a spam detection task, a model might predict:</p><ul><li>Spam → 0.8</li><li>Not Spam → 0.2</li></ul><p>Rather than simply saying “this is spam,” the model expresses how confident it is.</p><p><strong>2. Bayesian Thinking: updating beliefs</strong></p><p>One way AI handles uncertainty is by updating its beliefs as new information becomes available.</p><p>Instead of making a fixed decision, the model starts with an initial assumption and refines it as it sees more data.</p><p>For example, in medical diagnosis, a model might assign a probability to a disease based on initial symptoms. As more test results come in, it updates that probability.</p><p>This allows AI systems to become more confident as they learn.</p><p><strong>3. Fuzzy logic: handling ambiguity</strong></p><p>Fuzzy logic helps AI deal with situations where things are not clearly defined.</p><p>For example, what does it mean for traffic to be “heavy” or for the weather to be “bad”? These are not yes-or-no conditions.</p><p>Instead of using strict categories, fuzzy logic allows values to exist in degrees — such as low, medium, or high. This makes it especially useful in real-world applications.</p><p>In self-driving cars, for instance, AI must interpret traffic conditions, pedestrian movement, and weather , all of which can be uncertain or imprecise.</p><p>By working with these “partial truths,” fuzzy logic allows systems to make more flexible and realistic decisions.</p><h3>Conclusion</h3><p>Uncertainty is not something AI systems can avoid ,it is something they must learn to handle.</p><p>As we’ve seen, uncertainty arises from both the data and the model, and it plays a central role in how predictions are made. Whether through probabilities, updating beliefs, or handling ambiguity, AI systems are designed to operate in uncertain environments.</p><p>This makes one thing clear: <strong><em>Uncertainty is not a weakness of AI , it is a fundamental part of how it works.</em></strong></p><p>As AI continues to be used in high-stakes domains, understanding and communicating uncertainty becomes essential.</p><p>By building systems that can express doubt, adapt to new information, and operate transparently, we move closer to AI that is not just powerful, but also trustworthy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cab1ed5dc552" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Clear Mental Map of Modern AI: From ML to Agents]]></title>
            <link>https://medium.com/@lorraine.atsulu/a-clear-mental-map-of-modern-ai-from-ml-to-agents-718c1de7db9e?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/718c1de7db9e</guid>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Wed, 08 Oct 2025 18:37:47 GMT</pubDate>
            <atom:updated>2025-10-08T18:37:47.297Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Every time a new AI buzzword trends , from Transformers to Agents. it feels like the future is racing ahead while most of us are still trying to catch up. Even though I’ve been in the data science field for some time, I often pause to ask:</em></p><ul><li><em>“What exactly is AI?”</em></li><li><em>“And how can I explain it simply to someone who’s curious, without drowning them in jargon?”</em></li></ul><p><em>If you’ve ever felt the same, this article is for you. Let’s step back, zoom out, and </em><strong><em>build a clear mental map</em></strong><em> of modern AI , from its origins to its latest breakthroughs.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*YO23WnSUhoH7KB4n7rBA_Q.jpeg" /></figure><h3><strong>What is AI?</strong></h3><p>Artificial Intelligence (AI) is a broad area of computer science focused on building machines that can perform tasks that typically require human intelligence. These include:</p><ul><li>Recognizing patterns (like spotting faces in photos)</li><li>Understanding language (chatbots, translations)</li><li>Supporting decision making (recommendation systems)</li><li>Learning from experience (self-driving cars improving with more data)</li></ul><p>AI is no longer a research experiment. It’s one of the core drivers of the <strong>Fourth Industrial Revolution (Industry 4.0)</strong>. While the earlier revolutions were powered by <strong>steam, electricity, and computing</strong>, today’s is powered by <strong>intelligent automation, data-driven decision-making, and cyber-physical systems, </strong>with<strong> </strong>AI at the center.</p><h3><strong>How Did We Get Here?</strong></h3><p>A quick timeline helps put things in perspective:</p><p><strong>1940s–1950s | The Foundations</strong></p><ul><li><strong>1943: </strong><em>Walter Pitts &amp; Warren McCulloch</em> propose the <strong>artificial neuron model</strong>, laying the groundwork for neural networks.</li><li><strong>1947: </strong>The term <strong>“cybernetics”</strong> is introduced, exploring communication and control in machines and living beings.</li><li><strong>1949: </strong>The <strong>EDVAC</strong> computer is proposed, one of the first to store programs electronically.</li><li><strong>1950</strong> : <em>Isaac Asimov</em> publishes the famous <strong>“Three Laws of Robotics”</strong>, imagining ethical rules for intelligent machines.</li></ul><p><strong>1960s–1970s | Symbolic AI &amp; Early Experiments</strong></p><ul><li><strong>1960s–70s</strong> : AI research focuses on <strong>symbolic reasoning</strong> — teaching machines to manipulate symbols like humans do.</li><li><strong>1964– 66: ELIZA chatbot</strong> (Joseph Weizenbaum, MIT) uses <strong>rule-based pattern matching</strong> to mimic conversation with a psychotherapist.</li><li><strong>1968</strong> :HAL 9000, the fictional AI from <em>“2001: A Space Odyssey,”</em> captures public imagination.</li><li><strong>1971:</strong>The first <strong>autonomous vehicle prototype</strong> is built at Stanford, guided by computer vision.</li><li><strong>1972: MYCIN</strong>, an expert system for diagnosing bacterial infections, shows AI’s medical potential.</li></ul><p><strong>1980s–1990s | AI Winters &amp; Early Robotics</strong></p><ul><li><strong>1980s–90s</strong>: The first <strong>AI Winters</strong> occur as high expectations meet technical limits, leading to funding cuts.</li><li><strong>1980</strong> : <strong>WABOT-2</strong>, a humanoid robot, plays the piano and converses in Japanese.</li><li><strong>1995: ALICE chatbot</strong> demonstrates more advanced rule-based conversation.</li><li><strong>1997: Deep Blue</strong> (IBM) defeats chess champion Garry Kasparov — a landmark AI victory.</li><li><strong>1999</strong> : <strong>Sony AIBO</strong>, a robotic dog, is released, bringing AI robotics to consumers.</li></ul><p><strong>2000s | Data, Robots &amp; the ML Boom</strong></p><ul><li><strong>Early 2000s</strong> : Robots begin to appear in <strong>healthcare</strong>, assisting in surgery and logistics.</li><li><strong>2005: Stanley</strong>, Stanford’s autonomous car, wins the DARPA Grand Challenge.</li><li><strong>Mid-2000s</strong> : The <strong>Big Data era</strong> begins, fueling modern AI.</li><li><strong>2006: </strong><em>Geoffrey Hinton</em> popularizes the term <strong>“deep learning”</strong>, reviving neural networks with new techniques.</li></ul><p><strong>2010s–2020s | AI Everywhere</strong></p><ul><li><strong>2011</strong> — Voice assistants like <strong>Siri</strong> emerge, followed by <strong>Alexa</strong>, making AI mainstream.</li><li><strong>2010s</strong> — AI sees widespread adoption across industries, from military applications to personalized recommendations.</li><li><strong>2020s</strong> — <strong>AI becomes ubiquitous</strong>: chatbots, self-driving cars, smart cities, and intelligent automation define the <strong>Industry 4.0 era</strong>.</li></ul><h3>The Key Components of Modern AI</h3><p>Think of AI as a system made up of <strong>interconnected components</strong>, each building on the other to create more powerful capabilities. Understanding these blocks makes the rapid progress in AI much easier to follow:</p><ol><li><strong>Machine Learning (ML) — Learning from data</strong></li></ol><p>Machine Learning is a subset of AI where systems learn patterns from data instead of following hardcoded rules. How it works: Feed the model large datasets → it finds patterns → uses them to make predictions or decisions. Common ML models also referred to as the traditional ML model include: Linear regression, logistic regression, decision trees, random forests, support vector machines (SVMs), k-nearest neighbors (KNN), clustering. Everyday examples include: YouTube recommending your next video.</p><p><strong>2. Deep learning (DL) — Understanding complexity</strong></p><p>Deep Learning is a subset of ML that uses <strong>neural networks with many layers</strong> to model complex patterns. It’s especially powerful with <strong>unstructured data</strong> like images, speech, and text.</p><p><strong><em>a. Multilayer Perceptron(MLP): </em></strong><em>are simple layered neural networks where data moves only forward. Used in basic classification and machine learning tasks.</em></p><p><strong><em>b. Convolutional Neural Network (CNN) :</em></strong><em> are specifically designed to process image data and are excellent at identifying patterns in images. Moreover they can be adapted to handle sequential data by using 1D convolutional layers.</em></p><p><strong><em>c. Recurrent Neural Network (RNN):</em></strong><em> are specifically designed for sequential data (i.e audio, video, text). Thus can be used for tasks like language modeling, speech recognition &amp; natural language processing.</em></p><p><strong><em>d. Generative Adversarial Networks (GANs): </em></strong><em>GANs are an unsupervised learning method in which two networks , a generator and a discriminator, compete with each other. They are widely used for image, video, and audio generation when labeled data is scarce, and are capable of producing realistic synthetic content for applications like deepfakes or data augmentation.</em></p><p><strong><em>e. Graph Neural Networks(GNNs) : </em></strong><em>Specialized for data represented as graphs, such as social networks, molecules, or transportation systems, capturing relationships between entities.</em></p><p><strong><em>f. Autoencoders :</em></strong><em> Autoencoders are unsupervised neural networks that consist of an encoder and decoder. They learn to compress input data into a compact representation through an encoder and then reconstruct it via a decoder. An autoencoder learns to recognize which aspects of observable data are relevant and limit noise which can be discarded. They are commonly used for anomaly detection, denoising, dimensionality reduction, and generative tasks via Variational Autoencoders (VAEs).</em></p><p><strong><em>g. Transformers — The Breakthrough Deep Learning Architecture</em></strong></p><p><em>Transformers are deep neural networks</em> <em>built around </em><strong><em>self-attention mechanisms</em></strong><em>, which allow models to </em><strong><em>process entire sequences in parallel</em></strong><em> rather than step by step. They typically consist of:</em></p><ul><li><strong><em>Encoder layers</em></strong><em> — process the input and build contextual representations.</em></li><li><strong><em>Decoder layers</em></strong><em> — generate outputs based on those representations.</em></li></ul><p><em>Transformers unlocked a new level of scalability. Building on this, researchers trained enormous models on massive, diverse datasets. These became known as foundation models — the backbone of today’s generative and agentic AI systems.</em></p><p><strong>3. Foundation Models — Scaling Deep Learning</strong></p><p>Foundation models are large, general-purpose AI models pretrained on massive datasets spanning text, code, images, and sometimes even multimodal data. They are almost always built on <strong>Transformer architectures</strong> (sometimes mixed with diffusion or VAEs). Once pretrained, foundation models can be adapted to many downstream tasks without retraining from scratch. Examples include:</p><ul><li><strong>Large Language Models (LLMs):</strong>Specialize in text and code, capable of generating essays, summarizing documents, translating languages, or writing software. GPT (<strong>decoder-only Transformer</strong>), BERT (<strong>encoder-only Transformer</strong>), LLaMA (<strong>decoder-only Transformer</strong>).</li><li><strong>Vision Models</strong>: Specialize in images and video, capable of classification, generation, or understanding visual content. Examples: <strong>DALL·E (text-to-image, encoder-decoder transformer/diffusion)</strong>, <strong>Stable Diffusion (diffusion model)</strong>, <strong>Vision Transformers (ViTs, encoder-transformer)</strong></li><li><strong>Multimodal Models: </strong>Handle multiple types of data (e.g., text + images). Examples: <strong>GPT-4 (decoder-only Transformer)</strong>, <strong>CLIP (image-text, dual-encoder transformer)</strong>.</li></ul><p><strong>4. Generative AI — From understanding to creating</strong><br>Generative AI sits on top of foundation models and uses them to create <strong>new content</strong>: text, code, images, audio, or video.</p><ul><li>LLMs generate text (GPT, BERT, LLaMA).</li><li>Vision and multimodal models generate images and video (DALL·E, Stable Diffusion, CLIP).</li><li>Audio generation tools like <strong>Suno AI</strong> leverage diffusion models to produce music, songs, and lyrics.</li></ul><p>These models are built on deep learning architectures such as <strong>transformers</strong>, <strong>variational autoencoders (VAEs)</strong>, or <strong>diffusion-based networks</strong>. Generative AI powers applications like AI-assisted design, creative writing, music production, interactive assistants, and content summarization.</p><p><strong>5. AI Agents and Agentic AI— From understanding to action</strong><br>Generative AI creates content. <strong>AI Agents</strong> go a step further: they act. Agents are autonomous programs that can:</p><ul><li>Plan tasks</li><li>Use external tools (APIs, apps, databases)</li><li>Adapt to changing situations</li><li>Work with other agents to achieve bigger goals</li></ul><p>At their core, most agents use an <strong>LLM as the “brain.”</strong> They may also integrate vision or multimodal models for richer capabilities.</p><p><strong>Examples:</strong></p><ul><li>A <strong>travel agent</strong> that not only suggests when to climb Mt. Everest but also books your flights and hotel.</li><li>A <strong>coding agent</strong> that remembers past conversations, writes code, and runs it using APIs.</li><li>A <strong>geospatial agent</strong> (e.g., GeoGPT) that analyzes satellite imagery and produces maps or reports.</li></ul><p><strong>Agentic AI</strong> takes this further by coordinating <strong>multiple agents</strong> , like a team where each agent has a specialty. Together, they can tackle complex, multi-step problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/673/1*U4I5C2XidK5g2iPLM_l3xQ.png" /></figure><h3>Where are we headed?</h3><p>Understanding the <strong>levels of AI capability</strong> helps clarify where we are and where we’re going:</p><ol><li><strong><em>Narrow AI — What We Have Today</em></strong></li></ol><p>Narrow AI refers to systems designed to perform specific tasks extremely well, but within a <strong>limited domain</strong>. They don’t possess general reasoning, self-awareness, or the ability to transfer knowledge flexibly between domains. These systems are powerful, but they don’t <em>understand</em> the world like humans do , they excel only within the boundaries they’re trained for.<br><strong>Examples:</strong> Spotify recommending music, Google Translate, facial recognition.</p><p><strong><em>2. Artificial General Intelligence (AGI) — The Next Frontier</em></strong></p><p>AGI refers to machines with <strong>human-level intelligence</strong>: the ability to understand, learn, and apply knowledge flexibly across different tasks — just like we can.</p><p>An AGI would be able to:</p><ul><li>Solve unfamiliar problems without retraining</li><li>Combine knowledge from multiple domains seamlessly</li></ul><p>Think about it this way:<br>Today, we have specialized AI tools for different things: <strong>Chat-GPT for text</strong>, <strong>Suno for music</strong>, <strong>Excel for data analysis</strong>, and so on. Each one is powerful, but <strong>none of them can switch between these tasks on their own. </strong>They need to be retrained or reconfigured for each new domain.</p><p>By contrast, a true AGI would be like having <strong>one AI assistant</strong> that can write an article, compose a song, design the cover art, build the website, and then negotiate ad deals — all without needing a separate tool for each step. That’s why, despite the impressive progress of models like GPT-4, we’re still firmly in the <strong>era of Narrow AI</strong>.</p><p><strong><em>3. Artificial Superintelligence (ASI) — Beyond Human</em></strong></p><p>Super-intelligent AI refers to systems that not only match but <strong>vastly exceed human intellectual capabilities</strong> in every field including creativity, problem-solving, and social intelligence. It goes beyond Narrow or even General AI by possessing adaptability, self-improvement, and decision-making skills that surpass even the most brilliant human minds.</p><p>Futurists like Nick Bostrom popularized the idea of a <em>“superintelligence explosion”</em>, where once an AGI can improve itself, progress could accelerate uncontrollably, potentially transforming society at an unprecedented scale.</p><h3><strong>Conclusion</strong></h3><p>From rule-based expert systems in the 1960s to today’s AI agents, we’ve come a long way. The AI components — ML, DL, Transformers, Foundation Models, Generative AI, and Agents — show how each layer builds on the other, leading us to today’s Industry 4.0 era.</p><p>But here’s the important part: <strong>real-world impact doesn’t always require the most complex tools.</strong> Sometimes, a well-tuned Random Forest can solve a problem more effectively than a giant foundation model. What matters is understanding the landscape clearly enough to pick the right approach for the task.</p><p>The most exciting part? AI isn’t just about smarter machines. It’s about new ways of working, creating, and solving problems.</p><p>The best way to keep up is simple: stay curious, keep learning, and don’t be afraid to pause and ask, <em>“Wait, what’s really happening here?”</em> Because understanding is where the journey begins.</p><p>Happy reading 😊!!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=718c1de7db9e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GeoAI Decoded: Innovations Shaping 2025]]></title>
            <link>https://medium.com/@lorraine.atsulu/geoai-decoded-innovations-shaping-2025-0c11110dd5b5?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c11110dd5b5</guid>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Sun, 05 Oct 2025 21:07:09 GMT</pubDate>
            <atom:updated>2025-10-05T21:07:09.681Z</atom:updated>
            <content:encoded><![CDATA[<p><em>From Google’s foundation models to chat-based GIS assistants, GeoAI is no longer futuristic. It’s here, and it’s transforming how we understand our planet</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*mQYCBY6WHD4umVOgWdXRDQ.jpeg" /></figure><p>Artificial Intelligence (AI) refers to machines that can think, learn, and make decisions , essentially simulating human intelligence. <strong>Geospatial Artificial Intelligence (GeoAI)</strong> takes this further by combining AI with geospatial data, science, and technology.</p><p>This fusion allows us to analyze complex spatial datasets faster than ever, uncovering insights about business opportunities, environmental impacts, and operational risks. From tracking climate change to planning smarter cities, <strong>GeoAI is transforming the way we understand the world</strong>. Simply put, it’s a game changer for tackling some of Earth’s most pressing challenges.</p><p>Some of the most recent developments in the GeoAI field include:</p><p><strong><em>1.AlphaEarth by Google DeepMind</em></strong></p><p>One of the most exciting recent developments in GeoAI is <strong>AlphaEarth</strong>, an advanced AI model by Google DeepMind. Think of it as a <strong>“virtual satellite” with supercharged intelligence</strong>.</p><p>AlphaEarth fuses data from dozens of public sources ,optical imagery, radar, 3D laser mapping, and climate simulations ,all at a 10-meter resolution. It turns this into <em>embeddings</em> that encode spatial and temporal patterns.<br>This enables it to <strong>detect changes over time, compare locations, and provide a comprehensive planetary view</strong> that was previously impossible. AlphaEarth is a strong candidate to become a <strong>foundation model</strong> for monitoring Earth systems.</p><p><strong><em>2.GeoAI agents</em></strong></p><p>While AlphaEarth focuses on understanding the planet, <strong>GeoAI agents bring intelligence closer to the user</strong> through interactive, task-driven geospatial analysis.</p><p>Tools like <strong>GeoGPT</strong>, <strong>EarthGPT</strong>, and <strong>GeoChat</strong> are great examples. They act like smart GIS assistants built on large language models. Instead of writing code or using specialized software, you can simply ask:</p><blockquote><em>“Show me areas affected by deforestation in Country X over the last decade.”</em></blockquote><p>The agent fetches data, runs the analysis, and returns meaningful answers, often with maps or charts. This <strong>lowers the technical barrier</strong> and democratizes access to geospatial intelligence.</p><p><strong><em>3. Tools &amp; Open-Source Infrastructure</em></strong></p><p>Finally, the GeoAI revolution is being powered by an expanding ecosystem of <strong>open-source tools</strong>.</p><p>The <strong>GeoAI Python package</strong> is a great example. It integrates geospatial data into machine learning workflows ,from loading and preprocessing to feature extraction, training, and visualization ,all within Python.</p><p>Combined with libraries like <strong>geopandas</strong>, <strong>rasterio</strong>, and <strong>Google Earth Engine</strong>, these tools <strong>lower the barrier for startups, NGOs, and researchers</strong> to experiment with GeoAI. Open tools and data are becoming the backbone of modern geospatial innovation.</p><p><strong>Conclusion</strong></p><p>Whether it’s foundation models like AlphaEarth, conversational GeoAI agents, or open-source tools, 2025 is shaping up to be the year GeoAI moves from labs to everyday impact.</p><p>The question is, how will <em>you</em> use GeoAI?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c11110dd5b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring PDF Querying with LangChain]]></title>
            <link>https://medium.com/@lorraine.atsulu/exploring-pdf-querying-with-langchain-235594643509?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/235594643509</guid>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Fri, 12 Apr 2024 12:30:10 GMT</pubDate>
            <atom:updated>2024-04-12T12:30:10.706Z</atom:updated>
            <content:encoded><![CDATA[<p><em>In the realm of data, LangChain is the catalyst that transmutes ideas into innovation, unveiling insights from the cryptic language of text.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*emvX2GCPb2Nq-XW89Zxltw.jpeg" /></figure><p>LangChain is an open source framework that allows AI developers to combine Large Language Models (LLMs) like GPT-4 with external sources of data and computation. As a language model integration framework, LangChain’s use-cases largely overlap with those of language models in general, including document analysis and summarization, chatbots, and code analysis. In this article we will dive into the document summarization use case of LangChain specifically PDF documents.</p><p><strong>Flowchart of how LangChain works</strong></p><p>In an instance where you want to know something specific from your own data or document , LangChain allows you to connect LLMs like GPT-4 to your own data by referencing a database that has your data. Once you get the information you need, LangChain can be used to take action i.e retrieving titles of the PDFs with the information you need or send an email to a particular address with the information that you need.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1017/1*q__PvfSZXf2dxMW7jD8GvQ.png" /><figcaption>Document being stored in a vector database</figcaption></figure><p>We first start by identifying the document that you want the LLM to reference, the document is then sliced up to smaller chuncks as illustrated above. The document chunks are then stored as embeddings in a vector database. Embeddings create a vector representation of a piece of text. This is useful because it means we can think about text in the vector space, and do things like semantic search where we look for pieces of text that are most similar in the vector space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2cTtJ_5rTVACAWPwA5OfLA.png" /><figcaption>LangChain General Pipeline</figcaption></figure><p>A user asks an initial question, the question is sent to the LLM and vector representation of the question is used to do a similarity search in the vector database in order to retrieve the relevant information. The relevant information is then fed to the LLM. The LLM contains the initial question and the relevant information from the vector database hence the LLM has the ability to provide an answer or take an action.</p><p>LangChain helps build applications that are data aware( by referencing our own data stored in a vector database) and agentic (can take actions).</p><p><strong>Implementing LangChain using Python</strong></p><p>In this article, I will conduct a mini-project in which I use LangChain to extract the information I need from a PDF. In this tutorial, l’ll be using Hugging Face LLM Provider</p><pre>#loading required libraries<br>from langchain_community.document_loaders import PyPDFLoader<br>from langchain.text_splitter import RecursiveCharacterTextSplitter<br>from langchain.embeddings import HuggingFaceInstructEmbeddings<br>from langchain.vectorstores import FAISS<br>from langchain.llms import HuggingFaceHub<br>from langchain.chains.question_answering import load_qa_chain<br>import warnings<br>warnings.filterwarnings(&quot;ignore&quot;)<br>import os<br>from getpass import getpass<br>import re</pre><p>Next, l load the PDF file that we will be using for this tutorial. Make sure the pdf file is in the same directory as your notebook.</p><pre># Defining a fuction to retrieve the pdf file<br>def get_pdf_text(pdf_docs):<br>    text = &quot;&quot;<br>    #for pdf in pdf_docs:<br>    pdf_reader = PyPDFLoader(pdf_docs)<br>    for page in pdf_reader.load():<br>        text += page.page_content<br>    return text<br># Getting pdf document<br>text= get_pdf_text(&quot;R - into .pdf&quot;)</pre><p>We now split the document into chunks. We use the RecursiveCharacterTextSplitter(), which will recursively split the document using common separators like new lines until each chunk is the appropriate size.</p><pre># Defining the function to perform the splitting<br>text_splitter = RecursiveCharacterTextSplitter(<br>    # Set a really small chunk size, just to show.<br>    chunk_size=500,<br>    chunk_overlap=20,<br>    length_function=len,<br>    is_separator_regex=False,<br>)<br><br># Using the defined function on my document<br>texts = text_splitter.split_text(text)<br>texts[0]</pre><p>Next step is embedding the segmented text. sentence-transformers is a library at Hugging Face that provides easy methods to compute embeddings for sentences, paragraphs and images. For my case l used hkunlp/instructor-large.</p><pre># Embedding the text<br>embeddings = HuggingFaceInstructEmbeddings(model_name=&quot;hkunlp/instructor-large&quot;, model_kwargs={&quot;device&quot;: &quot;cpu&quot;})</pre><p>We now create the vector database. This article uses the FAISS vector database</p><pre># Creating a vector database<br>db = FAISS.from_texts(texts, embeddings)</pre><p>We now need to set up an API token to interact with Hugging Face Hub. You need to create an account with Hugging Face Hub if you don’t have an account in order to get the API token</p><pre># Setting up the API token<br>os.environ[&#39;HUGGINGFACEHUB_API_TOKEN&#39;] = getpass()</pre><p>We now use Mixtral API from Hugging Face Hub for text generation.</p><pre># Using the Mixtral API<br>llm = HuggingFaceHub(repo_id=&quot;mistralai/Mixtral-8x7B-Instruct-v0.1&quot;,task=&quot;text-generation&quot;,model_kwargs={<br>        &quot;max_new_tokens&quot;: 512,<br>        &quot;top_k&quot;: 30,&quot;temperature&quot;:0.1,<br>        &quot;repetition_penalty&quot;: 1.03, &quot;max_length&quot;:64})</pre><p>Lastly we integrate the API with a question-answering (QA) chain for similarity search with user queries. We define a function for this:</p><pre># Defining a function to automate the PDF querying process<br>def get_feedback(query):<br>    # Load the QA chain<br>    chain = load_qa_chain(llm, chain_type=&quot;stuff&quot;)<br>    <br>    # Perform similarity search<br>    docs = db.similarity_search(query)<br>    <br>    # Run the QA chain<br>    chain_response = chain.run(input_documents=docs, question=query)<br>    <br>    # Use regular expressions to find the answer<br>    match = re.search(r&#39;Helpful Answer:(.*)&#39;, chain_response)<br>    if match:<br>        answer = match.group(1).strip()<br>        return answer<br>    else:<br>        return &quot;Answer not found in the response.&quot;<br><br># Get user input<br>user_query = input(&quot;Enter your query: &quot;)<br><br># Get feedback<br>feedback = get_feedback(user_query)<br><br># Print feedback<br>print(&quot;Answer:&quot;, feedback)</pre><p><strong>Conclusion</strong></p><p>LangChain stands as a beacon of innovation, paving the way for transformative applications in natural language processing. Kindly feel free to explore other embedding techniques and LLMs.</p><p>Happy Reading!!!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=235594643509" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Anonymisation]]></title>
            <link>https://medium.com/@lorraine.atsulu/data-anonymisation-1e8be9cb5d94?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e8be9cb5d94</guid>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Thu, 04 May 2023 13:35:18 GMT</pubDate>
            <atom:updated>2023-05-04T13:35:18.596Z</atom:updated>
            <content:encoded><![CDATA[<p><em>One principle of data privacy is confidentiality. Confidentiality involves a promise usually executed through confidentiality agreements that limits the access or places restrictions on certain types of information.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/477/1*lxv_9aF2JtEh-wekaOPM6A.png" /></figure><p>Data anonymization is the process by which personal data is altered in such a way that a data subject can no longer be identified directly or indirectly. Its goal is to safeguard the privacy and confidentiality of data subjects while yet enabling the storage and use of data.</p><h3>Data Anonymisation Techniques</h3><p><strong>Data masking:</strong> is the process of hiding or altering values in a data set in a way that the original values cannot be re-engineered. Masking replaces original information with artificial data that is still highly convincing, yet bears no connection to the true values. For example, you can replace a value character with a symbol such as “#” or “x”.</p><p><strong>Generalization: </strong>involves deliberately removing some of the data to make it less identifiable but also retain some accuracy. Data can be modified into a set of ranges or a broad area with appropriate boundaries. For example instead of using house number in an address you can use the road name.</p><p><strong>Pseudonymization: </strong>a de-identification method that replaces private identifiers with artificial identifiers or pseudonyms. For example, data may be considered pseudonymised if individuals’ email addresses are replaced with numbers — the original, directly identifying information is removed, but each number is specific to a certain email address and can therefore be put back together by anyone with the right knowledge.</p><p><strong>Pertubation</strong>: is a method of altering data by introducing noise or random values. This makes it harder for attackers to determine the original data.A small base may lead to weak anonymisation while a large base can reduce the utility of the dataset. For example, you can add 5000/= to a person’s income and the value may retain its credence. However,adding 100,000/= to a person’s income may make it seem fake.</p><p><strong>Swapping: </strong>is a technique where data is swapped between different individuals or entities such that attribute values don’t correspond with the original records. This can make it difficult to link specific data points to a specific individual.</p><p><strong>Differential privacy</strong>: is a technique that adds noise to the data in a way that preserves privacy while maintaining statistical accuracy. For example, imagine we are measuring the overall trend in searches for influenza across a geographic region. This means we may add noise to the data by adding or subtracting the number of people searching for influenza in a given neighborhood, but doing so would not affect our measurement of the trend across the broader geographic region. It’s also important to note that adding noise to a data set may render it less useful.</p><p><em>NB:Google uses Generalization and Differential privacy techniques to anonymise data.</em></p><h3><strong>Choosing a Data Anonymisation Technique</strong></h3><p>Each organization’s data needs and use cases are different. The best approach to anonymisation is multiple layers of defense. When it comes to prioritizing which methods are right for your data stack, the best approach is to:</p><ol><li><strong>Type of Data:</strong> Consider the type of data that needs to be anonymized. Some data may be more sensitive than others, such as personally identifiable information, financial data, or health records. The level of sensitivity will determine the appropriate technique to use.</li><li><strong>Degree of Anonymization: </strong>The degree of anonymization required will depend on the specific use case. In some cases, it may be necessary to completely anonymize the data, while in others, it may be sufficient to use techniques that only partially anonymize the data.</li><li><strong>Risk of Re-Identification: </strong>Consider the risk of re-identification. Some data may be more prone to re-identification than others. For example, data that is already publicly available may be easier to re-identify than data that is not.</li><li><strong>Utility of the Data:</strong> The level of anonymization should also take into account the utility of the data. The more data is anonymized, the less useful it may be for analysis or research purposes.</li><li><strong>Legal and Regulatory Requirements: </strong>Different legal and regulatory requirements may require different levels of anonymization. For example, certain industries, such as healthcare and finance, have specific regulations governing the handling of sensitive data.</li></ol><h3>Conclusion</h3><p>Data breaches and cyberattacks can cause devastating damages i.e identity theft, fraudelent activities, phishing etc. Organizations need to proactively protect their data and regularly update their protective measures. As the amount of data stored and created increases, so does the importance of data protection.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e8be9cb5d94" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Filtering a Shapefile using Google Earth Engine(GEE)]]></title>
            <link>https://medium.com/@lorraine.atsulu/filtering-a-shapefile-using-google-earth-engine-gee-22b3243ae0cd?source=rss-10de6468dd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/22b3243ae0cd</guid>
            <category><![CDATA[google-earth-engine]]></category>
            <category><![CDATA[gis]]></category>
            <dc:creator><![CDATA[Lorraine Eyinda]]></dc:creator>
            <pubDate>Thu, 06 Apr 2023 17:48:27 GMT</pubDate>
            <atom:updated>2023-04-07T10:04:14.363Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>#GIS #SpatialData #GoogleEarthEngine #JavaScript</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/613/1*FuT1bnNQoj5a5_TmaSbbqQ.png" /></figure><p>Geographic Information Systems(GIS) is the study of fundamental properties of geographic information to aid in making discoveries and solving real world problems.</p><p><strong>TABLE OF CONTENTS</strong></p><blockquote>- What is filtering</blockquote><blockquote>- Procedure of filtering a Shapefile in Google Earth Engine.</blockquote><blockquote>- Conclusion</blockquote><p><strong>What is filtering?</strong></p><p>Filtering refers to removing by means of a filter. When a shapefile is filtered only the areas that we are interested in will be displayed and other areas will be hidden.</p><p>A shapefile is a vector data file format commonly used for geospatial analysis. Shapefiles store the location, geometry, and attribution of point, line, and polygon features.</p><p>For example, using GIS, a single map could include sites that produce pollution, such as factories, and sites that are sensitive to pollution, such as wetlands and rivers. Such a map would help people determine where water supplies are most at risk.</p><p><strong>Procedure of filtering a Shapefile in Google Earth Engine</strong></p><ol><li>Go to: <a href="https://earthengine.google.com/">https://earthengine.google.com/</a> .</li><li>Create an account. Scroll to the bottom of the page you should see a sign up now button. If you already have an account click on the Try Code Editor button.</li><li>On clicking try code editor, the interface below is displayed:</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1GRgtAtk0ll2-NN4DrXCEg.png" /></figure><p>Google Earth Engine has in-built datasets that we can use for this study.</p><p>4. Using the search bar at the top of the page, search for administrative units. Notice that a list of several datasets appears while typing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AM0DxOKlar4M-sbML-FAuA.png" /></figure><p>5. For this case we are interested in the polygons representing the administrative boundaries. Click on the first dataset on the TABLES section. Notice a new interface appears on your screen.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/965/1*OFSWh-rBFyGnfa_10CuXlA.png" /></figure><p>-A brief description about the data and its variables is given.</p><p>6. Click IMPORT. Notice that a line of code appears at the top most section of the script. <strong>NOTE: Any data that is imported into the Google Earth Engine will appear here.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*jvIhYiUwWNiZieSvC0uOKQ.png" /></figure><p>7. To display the imported data on the map we use the code below:</p><pre>// loading the dataset<br>var dataset = admin_boundary;<br><br>// Displaying the map<br>Map.addLayer(dataset, {}, &#39;Second Level Administrative Units&#39;);</pre><p>NB: The initial variable <strong>table</strong> was renamed to<strong> admin_boundary</strong></p><ul><li>The output is as shown below:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_wyuFOtaoEGPHw9uoEod4g.png" /></figure><p><strong>NB: The administrative units dataset contains administrative boundaries in the world.</strong></p><p>8. Filtering the data to our study area i.e Nairobi Region, we use the code below:</p><pre>// loading the dataset<br>var dataset = admin_boundary;<br><br>// filtering the dataset based on our region of interest(Nairobi Region)<br>var Nai = dataset.filter(ee.Filter.eq(&#39;ADM2_NAME&#39;,&#39;Nairobi&#39;));<br><br>// Displaying the map<br>//Map.addLayer(dataset, {}, &#39;Second Level Administrative Units&#39;);<br>Map.addLayer(Nai,{},&#39;Nairobi Region&#39;);</pre><ul><li>The output is as shown below:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gNthx5G5isF4rSdyrmbuPw.png" /><figcaption>Image Displaying Nairobi County</figcaption></figure><p>9. Adding other counties(Nakuru County and Kajiado County )to the map. The code is as shown below:</p><pre>// loading the dataset<br>var dataset = admin_boundary;<br><br>// filtering the dataset based on our region of interest(Nairobi Region)<br>var Nai = dataset.filter(ee.Filter.eq(&#39;ADM2_NAME&#39;,&#39;Nairobi&#39;));<br><br>//Kajiado county<br>var Kajiado = dataset.filter(ee.Filter.eq(&#39;ADM2_NAME&#39;,&#39;Kajiado&#39;));<br><br>//Nakuru county<br>var Nakuru = dataset.filter(ee.Filter.eq(&#39;ADM2_NAME&#39;,&#39;Nakuru&#39;));<br><br><br>// Displaying the map<br>//Map.addLayer(dataset, {}, &#39;Second Level Administrative Units&#39;);<br>Map.addLayer(Nai,{},&#39;Nairobi County&#39;);<br>Map.addLayer(Kajiado,{},&#39;Kajiado County&#39;);<br>Map.addLayer(Nakuru,{},&#39;Nakuru County&#39;);</pre><p>The output is as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1iCYBJ4ICYMl5qOfOcYcsg.png" /></figure><p><strong>Conclusion</strong></p><p>Filtering presents a focused view of a feature layer in a map. By limiting the visibility of features in a layer, you can reveal what’s important depending on your study.</p><p>With GIS technology, people can compare the locations of different things in order to discover how they relate to each other. Streets, buildings, and plants can all be displayed on a single map using GIS. As a result, people can recognize, evaluate, and comprehend patterns and correlations more quickly.</p><p>I sincerely hope that this post will be useful to you as you create your maps, particularly for those who are just beginning their GIS adventure. For feedback and recommendations, please contact me in the comment section or @<a href="https://medium.com/u/10de6468dd9">Lorraine_</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22b3243ae0cd" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>