<?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 Keshax on Medium]]></title>
        <description><![CDATA[Stories by Keshax on Medium]]></description>
        <link>https://medium.com/@keshax16?source=rss-485e7025fea3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Cp7_Ava_xJ2LeRC8</url>
            <title>Stories by Keshax on Medium</title>
            <link>https://medium.com/@keshax16?source=rss-485e7025fea3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 17:17:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@keshax16/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[SentinelPay: From Sensor Readings to Insurance Evidence]]></title>
            <link>https://medium.com/@keshax16/sentinelpay-from-sensor-readings-to-insurance-evidence-c14efc57b17c?source=rss-485e7025fea3------2</link>
            <guid isPermaLink="false">https://medium.com/p/c14efc57b17c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agriculture-technology]]></category>
            <category><![CDATA[students]]></category>
            <category><![CDATA[insurance]]></category>
            <category><![CDATA[iot]]></category>
            <dc:creator><![CDATA[Keshax]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 10:08:53 GMT</pubDate>
            <atom:updated>2026-10-30T10:08:53.120Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xSqgEY278irwXn_E-oy-Pg.png" /></figure><h3>What happens when the data collected by a small IoT prototype is connected to a real-world problem?</h3><h3>The problem that caught my attention</h3><p>In November 2025, the Food and Agriculture Organization of the United Nations reported that disasters had caused an estimated <strong>US$3.26 trillion in agricultural losses between 1991 and 2023</strong>. That works out to an average of US$99 billion every year.</p><p>These losses come from events such as floods, droughts and other disasters that affect agricultural production. The scale of the problem made me think about a technical question: how can we collect information about changing environmental conditions and make that information useful?</p><p>My initial interest was disaster detection using the Internet of Things (IoT). I wanted to understand how sensors could collect readings from the environment, how those readings could be transmitted wirelessly, and how a computer could turn them into information that people could interpret.</p><p>I eventually chose agricultural insurance as the application for this exploration.</p><p>When a disaster damages crops, information about the affected conditions can matter when an insurance claim is assessed. I became interested in whether data collected by connected devices could help document field conditions and support that process.</p><p>That idea became <strong>SentinelPay</strong>.</p><h3>Why IoT?</h3><p>The Internet of Things connects physical devices that collect and exchange data. These devices can contain sensors, a microcontroller and a communication system that sends readings to another device.</p><p>The interesting part for me was not just collecting a number from a sensor. It was understanding the journey that number takes.</p><p>A sensor measures something. A microcontroller processes or reads the output. A communication link transfers the information. Software then presents the readings in a form that a person can interpret.</p><p>Each stage introduces its own engineering questions. Is the reading meaningful? Is the data being transmitted correctly? What happens if the connection fails? And does the information actually tell us what we think it tells us?</p><p>I wanted to explore that process by building something, rather than only studying the components separately.</p><h3>Building the hardware</h3><p>My SentinelPay prototype uses an Arduino-compatible microcontroller board with several sensors:</p><ul><li><strong>DHT11:</strong> collects temperature and humidity readings.</li><li><strong>MQ-135:</strong> produces a response to certain gases and air-quality-related conditions.</li><li><strong>GY-521 / MPU6050:</strong> measures acceleration and angular velocity, allowing motion to be observed.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/969/1*c34hhfvAystP1vnTi9apYw.png" /><figcaption><strong>Exploded CAD View of the POD Prototype</strong><br>A simulated breakdown of the POD system, illustrating its component layout and how the physical elements fit together.</figcaption></figure><p>I also incorporated 433 MHz RF transmitter and receiver modules to explore wireless communication between the hardware and the computer-side system.</p><p>These components serve different purposes. The DHT11 provides environmental readings, the MQ-135 responds to gases, and the MPU6050 provides motion measurements. They are not interchangeable, and none of them independently confirms that a flood or agricultural disaster has occurred.</p><p>That distinction matters because collecting sensor data and detecting a disaster are two different tasks.</p><p>A device can produce a reading without understanding its cause. To identify a particular event reliably, a system would need suitable sensors, a defined detection method and testing against real conditions.</p><p>My current prototype is an exploration of the sensing, communication and software pipeline. It is not yet a validated disaster-detection system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/465/1*i9cC07tJNQgptxiqJe4YfA.png" /><figcaption><strong>POD Hardware Prototype and Wiring</strong></figcaption></figure><h3>From hardware to software</h3><p>Collecting readings is only one part of the project. I also wanted to see how the information could be represented on a computer.</p><p>I built the software side using VS Code and developed a simulation to demonstrate the system’s behaviour. The interface visualises changing values through graphs and other on-screen elements. I also explored how an AI-assisted reporting component, AGRO AI, could organise information into a structured report for an insurance-related workflow.</p><p>The simulation and the physical prototype help me explore different parts of the system. However, a simulated response is not automatically evidence that a physical sensor has measured the same event. The distinction between simulated data and hardware-generated data needs to remain clear.</p><p>The longer-term idea is to investigate how collected readings could be organised into useful evidence. For example, a report might record available sensor values, their timestamps and the source of the data. That could help a reviewer understand what the device recorded.</p><p>But a report is only as reliable as the information behind it. A graph showing a change in a sensor value does not, by itself, prove that crops were damaged or establish the cause of the damage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eQLBIYvya0MCDlisPoqqwA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m5vpe4j0dD7bBRAqSRjqTg.png" /></figure><p><strong>Live Sensor Monitoring Dashboard</strong></p><p>The VS Code dashboard visualises sensor readings from the prototype, providing a software interface for monitoring incoming data and exploring how it could be organised into useful evidence.</p><h3>Why agricultural insurance?</h3><p>I could have explored IoT through many applications, including environmental monitoring, industrial systems or smart buildings. Agricultural insurance gave me a specific context in which to think about the usefulness of the data.</p><p>Imagine a farmer reporting crop damage after extreme weather. An insurer needs to assess the claim using appropriate evidence. Could connected devices contribute additional information about conditions recorded in the field?</p><p>That is the question I want SentinelPay to help me investigate.</p><p>The potential users would include farmers, who need a fair way to document relevant conditions, and insurers, who need information to assess claims. Whether this approach would actually improve assessments depends on factors that still need to be tested, including sensor accuracy, device placement, connectivity, data integrity and how the evidence is interpreted.</p><p>A prototype that collects data is a starting point, not proof that the insurance process will improve.</p><h3>My connection to CubeSats and AI</h3><p>SentinelPay also builds on the curiosity I developed during my TKS Learn stage, when I explored CubeSats and AI.</p><p>The two projects operate in different contexts. SentinelPay is not a satellite system, and I am not claiming that its current prototype uses CubeSat technology. The connection is the engineering thinking that interests me: how a physical system gathers information, how that information moves through a system, and how software can help make it useful.</p><p>During Replicate, I wanted to take that curiosity into a new application and learn by building. SentinelPay gave me a way to investigate sensors, wireless communication and data visualisation together.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qsGKHfnAY0y0NkEq3WW8pw.png" /></figure><h3>What I would improve next</h3><p>The next stage is to move from demonstrating individual components towards testing the complete system.</p><p>First, I would verify the data path from the physical sensors through the wireless link to the computer. I would compare received readings with the original sensor outputs and investigate missing or inconsistent data.</p><p>Second, I would improve the evidence-reporting workflow. Each report should distinguish measured values from generated summaries and preserve useful context, such as timestamps and sensor identity.</p><p>Third, I would investigate what additional measurements and validation would be necessary for disaster-related applications. Detecting a meaningful event would require more than simply observing a change in temperature, humidity, gas response or motion.</p><p>Finally, I would need to test the concept under realistic conditions and evaluate whether the resulting information could genuinely help an insurance assessor. Until those steps are completed, SentinelPay remains a prototype exploring a possible approach rather than a deployed insurance solution.</p><h3>My hot take</h3><p><strong>Collecting data is the easy part. Knowing what that data can actually prove is the harder engineering problem.</strong></p><p>A sensor reading can look convincing on a dashboard while still being incomplete, misleading or unrelated to the event someone wants to investigate. If SentinelPay is going to develop beyond a demonstration, I need to think about the quality, context and limitations of the evidence, not just whether the hardware and software work.</p><p>For me, that is the most important lesson of this project so far. Building an IoT system is not only about making devices communicate. It is about understanding the entire path from a physical measurement to a decision someone might make using it.</p><p>SentinelPay is my attempt to explore that path, one component at a time.</p><h3>About Me</h3><p>I’m Keshav, a 16-year-old TKS student based in the UAE, with a strong interest in aerospace, embedded systems, robotics, and AI. I enjoy exploring how technology works by building prototypes and experimenting with hardware and software. From exploring CubeSats during my Learn stage to developing SentinelPay during Replicate, I’ve been interested in understanding how systems collect data, communicate, and apply it to real-world problems. I want to keep turning what I learn into projects that I can build, test, and improve.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c14efc57b17c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond “Yes” or “No”: How Spacecraft Could Make Decisions]]></title>
            <link>https://medium.com/@keshax16/beyond-yes-or-no-how-spacecraft-could-make-decisions-1c7ad5dcd03b?source=rss-485e7025fea3------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c7ad5dcd03b</guid>
            <category><![CDATA[emerging-technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cubesat]]></category>
            <category><![CDATA[satellite-technology]]></category>
            <category><![CDATA[space]]></category>
            <dc:creator><![CDATA[Keshax]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 14:30:43 GMT</pubDate>
            <atom:updated>2026-10-22T14:30:43.786Z</atom:updated>
            <content:encoded><![CDATA[<h3>How AI is changing what small satellites can decide on their own</h3><p>The first time I started looking into AI on satellites, I thought the interesting part would be getting an AI model to actually run in space.</p><p>Then I found out people had already done that.</p><p>Small satellites have already run machine-learning models onboard. They have detected clouds, processed Earth-observation images, and experimented with deciding what information is worth keeping or sending back to Earth.</p><p>So I started wondering about something else.</p><p><strong>What happens when the satellite doesn’t know what it’s looking at?</strong></p><p>That question ended up being much more interesting to me than simply asking whether a satellite can run AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2NGVeuZHdytxp7jPHb8YNg.jpeg" /></figure><h3>The problem starts with a very unglamorous limitation: bandwidth</h3><p>A satellite can collect a huge amount of information, but it can’t constantly send everything it collects back to Earth.</p><p>A small satellite might only have short communication windows with a ground station. During that time, it has to decide what information is actually worth transmitting.</p><p>Imagine a satellite taking pictures of Earth.</p><p>It could collect hundreds of images while flying over an area. But maybe half of them are covered by clouds. Some might be blurry. Some might contain nothing useful for the specific mission.</p><p>Sending all of them back to Earth wastes time and bandwidth.</p><p>And this gets more interesting when the thing you’re looking for is time-sensitive.</p><p>If you’re trying to detect a wildfire, for example, finding something interesting and then waiting hours for the raw data to reach Earth isn’t ideal.</p><p>This is where onboard AI starts becoming useful.</p><p>Instead of sending everything down first and letting a computer on Earth figure it out, the spacecraft can process some of the information <strong>before it ever leaves the satellite.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4cGtofAOfLcH25hBbkgykw.jpeg" /></figure><h3>I found out this was already happening</h3><p>One example that caught my attention was ESA’s <strong>Φ-sat-1</strong> mission.</p><p>The satellite used an AI processor to analyse Earth-observation images onboard and identify clouds.</p><p>The basic idea makes sense.</p><p>If an image is mostly covered by clouds, there may be little reason to spend valuable bandwidth sending the entire image to Earth.</p><p>So why not process it in space first?</p><p>That sounds simple, but there is an important difference between doing this in a data centre and doing it on a satellite.</p><p>On Earth, if your model is wrong, you can usually run another model, download more data, change the software, or simply ask a human to look at the problem.</p><p>A satellite doesn’t have that luxury all the time.</p><p>Its computer has limited processing power.</p><p>Its power budget is limited.</p><p>Its communication with Earth is limited.</p><p>And most importantly, <strong>the spacecraft has to deal with whatever the camera actually sees.</strong></p><p>Not whatever the training dataset looked like.</p><h3>The part that made me rethink onboard AI</h3><p>While going through examples of AI being used on small satellites, I kept seeing the same basic pattern:</p><p><strong>observe → process → decide what to do with the data.</strong></p><p>At first, I thought the “decide” part basically meant classification.</p><p>Is this cloud or not?</p><p>Is this wildfire or not?</p><p>Is this image useful or not?</p><p>But I don’t think that is where the really interesting problem is.</p><p>Real environments aren’t that clean.</p><p>Suppose a satellite is looking at an area and its model thinks there might be a wildfire.</p><p>What should it do?</p><p>It could immediately mark the image as important.</p><p>But what if the image is noisy?</p><p>What if smoke looks similar to clouds?</p><p>What if the model has never seen that particular landscape before?</p><p>What if the prediction is only moderately confident?</p><p>The spacecraft shouldn’t necessarily treat that prediction as a fact.</p><p>It could instead decide to collect another observation.</p><p>It could compare the new image with something it captured earlier.</p><p>It could prioritize the image for transmission but keep the original data.</p><p>It could continue observing the area.</p><p>Or, if the mission allows it, it could wait for instructions from Earth.</p><p>That is a very different idea from simply asking an AI model for <strong>YES or NO</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SgAor3qCfA7MtGa3tX3HkA.png" /></figure><h3>This is where I think spacecraft autonomy gets interesting</h3><p>The more I looked into onboard AI, the more I became interested in the space <strong>between</strong> perception and action.</p><p>A model can tell a spacecraft what it thinks it sees.</p><p>But that doesn’t automatically tell the spacecraft what it should do.</p><p>Imagine a satellite observing the same location multiple times.</p><p>The first observation gives it a weak signal.</p><p>The second observation gives it more evidence.</p><p>The spacecraft now has information from both observations.</p><p>It could change its next action based on that history.</p><p>That starts to look less like a camera running a classifier and more like an autonomous system continuously updating what it believes about its environment.</p><p>And this matters far beyond Earth observation.</p><p>A spacecraft approaching another spacecraft has the same basic problem.</p><p>A planetary rover has it.</p><p>A spacecraft navigating around an asteroid has it.</p><p>A group of satellites working together has it.</p><p>The environment is changing, the sensors are imperfect, and the spacecraft rarely has complete information.</p><h3>Small computers make this even harder</h3><p>There is another part of this that I find pretty interesting.</p><p>We normally associate AI with massive computers and huge data centres.</p><p>A satellite can’t just throw a giant GPU cluster into orbit.</p><p>Every extra piece of hardware has consequences for power, mass, thermal management and reliability.</p><p>Researchers have therefore experimented with much smaller processors and models.</p><p>One study I came across explored cloud detection using a microcontroller-level system. Other work has looked at running wildfire-detection models on edge hardware.</p><p>That made me realise that “AI in space” isn’t necessarily about putting the biggest possible model on a satellite.</p><p>It is almost the opposite.</p><p>The interesting engineering problem is:</p><p><strong>How much intelligence can you fit into a very constrained machine?</strong></p><p>And then there is an even harder question:</p><p><strong>What should that intelligence actually be allowed to do?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bhv_gQ3akNiw9gMyXA5ttQ.jpeg" /></figure><h3>Wildfires are a good example</h3><p>Companies such as <strong>OroraTech</strong> are using satellite data for wildfire monitoring, while projects such as <strong>FireSat</strong> are being developed around more frequent and detailed wildfire detection from space.</p><p>The reason onboard processing becomes interesting here is obvious.</p><p>A satellite constellation could potentially observe huge amounts of Earth.</p><p>But you don’t necessarily want every raw image travelling all the way back to Earth before anything happens with it.</p><p>If the spacecraft can identify interesting observations onboard, it can change what information gets prioritized.</p><p>But again, I don’t think the goal should simply be:</p><p>“AI found fire. Send alert.”</p><p>A real system has to deal with uncertainty.</p><p>Smoke, clouds, heat sources, shadows and unusual terrain can all complicate what the sensors see.</p><p>A model can be extremely fast and still be wrong.</p><p>And in an application like wildfire detection, a wrong decision isn’t just a number in a spreadsheet. It could affect what people investigate next.</p><p>So speed isn’t enough.</p><p>Accuracy isn’t enough either.</p><p>The spacecraft needs some way of understanding <strong>how much it should trust its own prediction.</strong></p><h3>What I think the next step is</h3><p>This is the part I am still trying to understand.</p><p>We already know how to put machine-learning models on small spacecraft.</p><p>We are getting better at processing data without immediately sending everything to Earth.</p><p>But what happens when the spacecraft encounters something that doesn’t match what it learned?</p><p>I don’t think the answer should be to make the spacecraft pretend it knows.</p><p>Instead, I think the system should be able to recognise uncertainty.</p><p>For example, instead of thinking:</p><p>“This is definitely a wildfire.”</p><p>the spacecraft could effectively reason:</p><p>“I think this might be a wildfire, but the evidence isn’t strong enough. I should collect more information before changing what I do.”</p><p>That difference sounds small.</p><p>I don’t think it is.</p><p>Because once a spacecraft can react differently depending on <strong>how confident it is</strong>, you can start building much more interesting autonomous behaviour.</p><p>It can choose what information to collect next.</p><p>It can change its observation strategy.</p><p>It can compare new information with previous observations.</p><p>It can decide which data is important enough to send to Earth.</p><p>And eventually, it could decide that a situation is outside what it understands and hand the problem back to a human.</p><h3>My hot take</h3><p>I don’t think the future of onboard AI is a spacecraft that tries to make every decision by itself.</p><p>I’d rather see spacecraft that are good at knowing <strong>when they shouldn’t make the decision yet.</strong></p><p>That sounds less impressive than saying “fully autonomous spacecraft.”</p><p>But I think it could actually be more important.</p><p>If a spacecraft is millions of kilometres away, there will be situations where its sensors don’t give it enough information to be certain.</p><p>The intelligent response isn’t necessarily to choose an answer anyway.</p><p>It might be to gather another piece of information.</p><p>Or change its observation.</p><p>Or compare what it is seeing with what it saw before.</p><p>Or wait.</p><p>Or ask Earth.</p><p>That makes me think the future of autonomous spacecraft isn’t just about giving them better models.</p><p>It is about giving them better <strong>decision-making loops</strong>.</p><p>And that is where I think things get really interesting.</p><p>A spacecraft doesn’t necessarily need to “think like a human.”</p><p>It needs to become good at operating when humans aren’t immediately there to tell it what to do.</p><h3>What I still want to figure out</h3><p>The question I am left with is not:</p><p><strong>“Can we put AI on a satellite?”</strong></p><p>We already can.</p><p>It is:</p><p><strong>“How much decision-making should we actually give it?”</strong></p><p>If a spacecraft can observe something, estimate what is happening, measure its own uncertainty, collect more evidence and then choose an action, where should we draw the line between onboard autonomy and human control?</p><p>I don’t have a complete answer to that yet.</p><p>But that is probably the part of this topic I want to explore next.</p><p>Because the interesting future of spacecraft might not be machines that always know what to do.</p><p>It might be machines that know when they <strong>don’t</strong> know what to do.</p><h3>About Me</h3><p>Hi, I’m Keshav Chaitan, 16, . I enjoy building hands-on tech and breaking down complex space systems into working prototypes. I’m curious about the intersection of aerospace, autonomous robotics, and edge computing because that’s where deep space hardware learns to think for itself. I love working with sensors, running flight simulations, and sharing real proof-of-work.</p><h3>Sources</h3><ul><li>ESA PhiSat page:<a href="https://www.esa.int/Applications/Observing_the_Earth/FutureEO/Phsat"> https://www.esa.int/Applications/Observing_the_Earth/FutureEO/Phsat</a></li><li>ESA presentation on PhiSat-1:<a href="https://indico.esa.int/event/338/contributions/5692/"> https://indico.esa.int/event/338/contributions/5692/</a></li><li>Intel on the first AI satellite:<a href="https://www.intel.com/content/www/us/en/newsroom/news/first-satellite-ai.html"> https://www.intel.com/content/www/us/en/newsroom/news/first-satellite-ai.html</a></li><li>Salazar et al., cloud detection on a microcontroller:<a href="https://www.mdpi.com/2072-4292/14/21/5597"> https://www.mdpi.com/2072-4292/14/21/5597</a></li><li>Thangavel et al., wildfire detection on edge hardware:<a href="https://www.mdpi.com/2072-4292/15/3/720"> https://www.mdpi.com/2072-4292/15/3/720</a></li><li>Earth Fire Alliance, FireSat:<a href="https://earthfirealliance.org/about-firesat/"> https://earthfirealliance.org/about-firesat/</a></li><li>OroraTech technology:<a href="https://ororatech.com/company/technology"> https://ororatech.com/company/technology</a></li><li>Greek wildfire satellites:<a href="https://techstrong.ai/articles/new-greek-satellites-use-ai-to-spot-emerging-wildfires/"> https://techstrong.ai/articles/new-greek-satellites-use-ai-to-spot-emerging-wildfires/</a></li><li>OroraTech wildfire service:<a href="https://aerialfiremag.com/2022/01/07/ororatech-develops-the-first-global-wildfire-intelligence-service-using-satellites/"> https://aerialfiremag.com/2022/01/07/ororatech-develops-the-first-global-wildfire-intelligence-service-using-satellites/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c7ad5dcd03b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>