<?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[Techloop - Medium]]></title>
        <description><![CDATA[We, IEEE-VIT Student Branch, are a team of tech enthusiasts who aim to develop and support the tech community. In alignment of our vision, we present to you our medium blog where we guide people through the basics and discuss the most emerging technologies. - Medium]]></description>
        <link>https://medium.com/techloop?source=rss----7166652df8c7---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Techloop - Medium</title>
            <link>https://medium.com/techloop?source=rss----7166652df8c7---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 16:39:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/techloop" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How a Failed Refrigerant Became the Coating on Your Frying Pan]]></title>
            <link>https://medium.com/techloop/how-a-failed-refrigerant-became-the-coating-on-your-frying-pan-c05c7898de7d?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/c05c7898de7d</guid>
            <category><![CDATA[polymer]]></category>
            <category><![CDATA[materials-science]]></category>
            <dc:creator><![CDATA[Mohit Kulkarni]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 09:09:16 GMT</pubDate>
            <atom:updated>2026-11-05T09:09:15.211Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xwoHWlFVQ1IS0AvZOBrN3A.png" /></figure><p>You’ve probably heard this one before: <em>non-stick pans were invented for rocket nozzles. NASA built the technology, and eventually it found its way into our kitchens.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/0*VWoq4vl2I1wcWstx.jpg" /></figure><p>It sounds right, and it makes an ordinary frying pan feel like a small piece of the space race. But it isn’t true. The actual history is older, stranger, and a lot more interesting.</p><h4>A cylinder that should have been empty</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/309/0*9N6WyKG-FJR_Dl8J" /><figcaption>Roy Plunkett</figcaption></figure><p>In April 1938, at DuPont’s Jackson Laboratory in New Jersey, chemist Roy Plunkett was working on new refrigerants. He had several pressurized cylinders of tetrafluoroethylene (TFE) stored on dry ice. When he opened one, nothing came out, yet the cylinder weighed exactly what it should.</p><p>Plunkett sawed it open and found a waxy white powder. Under pressure, and possibly helped along by the iron of the cylinder walls, the TFE had polymerized on its own into extremely long chains. The powder resisted strong acids and solvents, and nothing stuck to it. He had accidentally created polytetrafluoroethylene (PTFE), later known as Teflon.</p><p>For several years it remained a laboratory curiosity. Then World War II arrived, and the Manhattan Project needed to handle uranium hexafluoride, one of the most corrosive compounds in industrial use, for gaseous diffusion enrichment. It attacked ordinary seals and gaskets. PTFE was effectively immune. Teflon’s first real application had nothing to do with kitchens.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/331/0*eieGHFQX6eDi-yzq" /><figcaption>PTFE (granules)</figcaption></figure><h4>Why it behaves the way it does</h4><p>This is the part most Teflon articles skip, and it’s where the material gets truly interesting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/1*JoBBcgZ17_ENxN5XNOnR1g.png" /><figcaption>Molecular structure of PTFE</figcaption></figure><p>PTFE is a simple polymer: a carbon backbone with two fluorine atoms on every carbon, repeated thousands of times. Fluorine explains almost everything about it.</p><p><strong>A very strong bond.</strong> The carbon-fluorine bond has a dissociation energy of roughly 485 kJ/mol, compared with about 350 for carbon-carbon and 410 for carbon-hydrogen. Few chemicals can break it, which is why PTFE resists acids, bases, and most solvents.</p><p><strong>A fluorine sheath.</strong> Fluorine atoms are larger than hydrogen, so they crowd along the chain and force it into a gentle helix. The result is a smooth, tightly packed fluorine shield around the carbon backbone.</p><p><strong>Almost no grip.</strong> Fluorine holds its electrons tightly, so intermolecular attractions on the surface are extremely weak. PTFE’s surface energy is about 18 to 20 mN/m, among the lowest of any solid, and its coefficient of friction is roughly 0.05 to 0.10. Water beads up on it at a contact angle near 110 degrees.</p><p>There is a manufacturing catch. PTFE’s chains are so long that, at its melting point of about 327°C, it doesn’t flow like normal plastic. It can’t be injection-molded, so the powder is pressed into shape and sintered, much like ceramics.</p><h4>So where do rockets come in?</h4><p>The myth isn’t entirely baseless, which is part of why it has lasted so long.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/0*XDW__UudqzWgitW0.jpg" /><figcaption>Rocket Engine diagram</figcaption></figure><p>A nozzle handles combustion gases well above 3,000°C, far beyond PTFE’s melting and decomposition range. Those parts use carbon-carbon composites, refractory metals, and ablative materials.</p><p>Where PTFE does appear in aerospace is in roles that suit its inertness: wiring insulation, valve linings, and seals in systems handling aggressive oxidizers. It has even served as a propellant. In pulsed plasma thrusters, an electric arc ablates small amounts of solid PTFE into plasma, which electromagnetic forces accelerate to produce thrust.</p><h4>The real origin of your frying pan</h4><p>The pan traces back to French engineer Marc Grégoire, who in the early 1950s was using PTFE to coat fishing tackle. According to the widely told account, his wife, Colette, suggested trying it on her pans.</p><p>It created a real engineering problem: the low surface energy that repels food also repels metal. Grégoire’s solution, patented in 1954, was to roughen the aluminum so the coating could grip mechanically. Modern cookware follows the same logic:</p><ul><li><strong>Surface preparation:</strong> the metal is grit-blasted or etched to create microscopic peaks and valleys.</li><li><strong>Primer:</strong> PTFE blended with binder resins such as polyamide-imide or polyphenylene sulfide, which bond to metal.</li><li><strong>Topcoats:</strong> additional PTFE-rich layers that form the non-stick surface.</li><li><strong>Curing:</strong> a final bake at roughly 400°C sinters the particles into a continuous film.</li></ul><p>The finished coating is only tens of microns thick, thinner than a human hair. Grégoire founded his company in 1956, and it became Tefal, a name blending “Teflon” and “aluminium.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/528/0*0QDaGZMzapKHt2dB.jpg" /><figcaption>Marc Grégoire</figcaption></figure><h4>The part nobody likes to discuss</h4><p>Teflon’s history isn’t a spotless success story, and the drawbacks fall into two categories.</p><p><strong>Manufacturing.</strong> For decades, PTFE production used PFOA (perfluorooctanoic acid) as a processing aid. It wasn’t meant to remain in the product, but it persisted in the environment and in human blood and was linked to serious health concerns. Under the EPA’s PFOA Stewardship Program, major manufacturers committed to phasing it out, and current PTFE cookware is made without it. PTFE still belongs to the broader PFAS family, whose environmental persistence remains under debate.</p><p><strong>Heat.</strong> The finished coating is considered safe under normal cooking, but PTFE begins to degrade at around 260°C, releasing fluorinated gases. An empty pan on a high burner can reach that range within minutes. In people this can cause “polymer fume fever,” a short-lived, flu-like illness. For pet birds the fumes can be fatal.</p><h4>Conclusion</h4><p>A failed refrigerant experiment became a wartime industrial material, then a spacecraft component, then the coating on a frying pan. None of it was planned. The chemistry was already there, in one of the strongest bonds in organic chemistry, and it took people willing to look at it differently.</p><p>Teflon never needed to come from space to earn its place. Its value comes from a simple property, a fluorine sheath that nothing can grip, and from the people who found new places to use it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c05c7898de7d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/how-a-failed-refrigerant-became-the-coating-on-your-frying-pan-c05c7898de7d">How a Failed Refrigerant Became the Coating on Your Frying Pan</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Smart Surgical Equipment borrowed from Subterranean Mining Drills]]></title>
            <link>https://medium.com/techloop/what-smart-surgical-equipment-borrowed-from-subterranean-mining-drills-b3b0243222ae?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/b3b0243222ae</guid>
            <dc:creator><![CDATA[Sarakuthurjames]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 07:01:03 GMT</pubDate>
            <atom:updated>2026-10-22T07:01:03.437Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RIUyY1oxvkLa_WrkGlrNSg.png" /></figure><p>Thousands of metres below the Earth’s surface, a mining drill is operating in an environment where everything is against it. The machine has to fracture hard rock while operating inside confined tunnels, withstand continuous vibration and mechanical loading, manage heat and dust, and remain reliable in a location where even routine maintenance is scarce. Not an easy job. Especially in large-scale mining operations, equipment downtime can translate directly into huge economic losses.</p><figure><img alt="A minecraft mining gif" src="https://cdn-images-1.medium.com/max/256/1*HbIbdEpe1UfxFuGMpoNRVw.gif" /></figure><p>The obvious engineering response would be to build a more powerful machine. But underground mining quickly exposes the limitations of that approach. More power means larger motors, greater energy consumption, more heat, increased mechanical stress, and often a larger machine. This may not even fit within the tunnel it is supposed to operate in.</p><p>The problem therefore becomes less about generating force and more about controlling force. A drill has to deliver enough energy to fracture rock while preventing unnecessary loads from propagating through the rest of the machine.</p><h3>Drilling becoming a Sensing Problem</h3><p>This requirement pushed drilling technology toward sensing, automation, and feedback control. Instead of operating as a purely mechanical system, modern drilling equipment can monitor parameters such as torque, rotational speed, penetration rate, vibration, motor current, hydraulic pressure, and temperature.</p><p>A sudden increase in motor load, for example, can indicate that the drill has encountered greater resistance. Changes in vibration can reveal variations in drilling conditions or potential mechanical problems. Temperature and pressure measurements can indicate whether components are operating outside their intended range.</p><p>The basic architecture becomes a closed feedback loop: sensors observe the system, a controller interprets those measurements, and actuators modify the machine’s behaviour.</p><h3>Turning Vibration into Information</h3><p>One of the most significant problems in heavy drilling equipment is vibration. A rotating cutting tool naturally produces dynamic loads, but excessive vibration can accelerate bearing wear, damage cutting tools, increase structural fatigue, and reduce drilling efficiency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/0*81K3_hD3DDexRZlJ.png" /><figcaption><a href="https://www.globalspec.com/learnmore/specialized_industrial_products/mining_equipment/measurement_while_drilling_mwd_systems">Measurement while drilling</a></figcaption></figure><p>Engineers therefore developed increasingly sophisticated methods for measuring and analysing vibration. Accelerometers and other sensors can capture the machine’s dynamic response, while signal-processing techniques can identify patterns associated with particular operating conditions or mechanical faults.</p><p>This is the foundation of <a href="https://en.wikipedia.org/wiki/Condition_monitoring">condition monitoring</a>. Rather than waiting for a machine to fail, engineers continuously observe its behaviour and attempt to identify changes before they become catastrophic.</p><p>That principle is now common across advanced machinery. A machine can be monitored throughout its operating life, with sensor data providing clues about wear, instability, overheating, or abnormal loading.</p><p>And this is where the connection to medicine becomes particularly interesting.</p><h3>From Rock to Bone</h3><p>A surgical drill might look completely different from a mining drill, but it faces a related control problem. It must apply mechanical energy to a target, usually bone, while keeping that energy tightly controlled. Excessive force, vibration, or heat can damage surrounding tissue, while insufficient force can make the procedure inefficient or inaccurate.</p><p>The important point is not that rock and bone behave identically. The connection lies in the engineering problem of tool-environment interaction.</p><p>A surgical system therefore cannot simply maximise drilling speed or motor power. It has to maintain controlled operating conditions. This is precisely the kind of problem that <a href="https://en.wikipedia.org/wiki/Feedback">feedback-based engineering</a> is designed to solve.</p><p>Sensors can provide information about position, velocity, force, temperature, and vibration. Motors can provide controlled actuation. Embedded processors can analyse sensor data in real time. Control algorithms can then adjust the behaviour of the system.</p><p>The result is what engineers would describe as a mechatronic system: a system in which mechanical components, electronics, sensors, actuators, and software are designed to operate together.</p><h3>From Mechanical Tools to Mechatronic Systems</h3><p>A modern surgical instrument is no longer simply a mechanical tool.</p><p>Consider a powered surgical device. The motor controls the movement of the cutting mechanism. An encoder can measure rotational or positional information. Sensors can monitor operating conditions, while an embedded controller continuously regulates the system.</p><figure><img alt="Closed-loop mechatronic feedback diagram" src="https://cdn-images-1.medium.com/max/491/1*WLz5XFMRrBtxQ7fE6EYbLg.png" /><figcaption><a href="https://www.researchgate.net/figure/Figure-36-A-Closed-Loop-Control-System_fig4_384605745">Closed-loop mechatronic feedback diagram</a></figcaption></figure><p>The mechanical components perform the physical task, but computation determines how that task is performed.</p><p>This is the same fundamental evolution that occurred in advanced industrial machinery. Sensors were added to mechanical systems, controllers were introduced to interpret those measurements, and software increasingly became responsible for coordinating physical behaviour.</p><p>The machine effectively became a <a href="https://en.wikipedia.org/wiki/Cyber-physical_system">cyber-physical system</a>, one in which computation and physical processes operate as a single feedback-controlled system.</p><p>This becomes even more significant when robotics enters the picture.</p><h3>Where Robotics Enters the Operating Room</h3><p>A robotic surgical system introduces multiple controlled axes of movement. Instead of simply turning a drill, the system may need to coordinate several joints while maintaining a precise position and orientation at the instrument tip.</p><p>The controller has to continuously determine how movements of individual motors translate into movement of the final surgical instrument.</p><p>This is fundamentally a robotics problem, and many of the underlying concepts are shared with industrial automation. Robotic manipulators used in manufacturing have long relied on encoders, motor controllers, kinematic models, trajectory planning, and feedback loops to achieve precise motion.</p><p>Medical robotics applies these principles under much stricter constraints.</p><p>The human operator also remains an important part of the system. In many surgical robotic platforms, the surgeon provides the high-level intent while the robotic system translates that input into controlled mechanical movement.</p><p>One example is motion scaling. A relatively large movement by the surgeon can be translated into a much smaller movement by the robotic instrument. The system effectively creates a computational layer between human movement and mechanical action.</p><p>This can allow unwanted movement to be reduced and instrument positioning to be stabilised. The machine isn’t necessarily replacing the surgeon. Instead, it is acting as a precision interface that allows human decisions to be executed through a more controlled mechanical system.</p><h3>The Common Engineering Problem</h3><p>One deals with kilometres of rock. The other deals with living tissue. Yet both environments impose severe engineering constraints.</p><p>A mining system has to survive high force, vibration, heat, dust, restricted space, and difficult maintenance conditions. A surgical system has to operate within tight spatial constraints, interact safely with biological material, and maintain extremely precise motion.</p><p>In both cases, engineers cannot simply solve the problem by adding more power. They need better sensing, better actuation, better modelling, and better control.</p><p>That is the connection between subterranean mining and smart surgical equipment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3b0243222ae" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/what-smart-surgical-equipment-borrowed-from-subterranean-mining-drills-b3b0243222ae">What Smart Surgical Equipment borrowed from Subterranean Mining Drills</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Phone Camera Is Solving a Fighter Jet Problem]]></title>
            <link>https://medium.com/techloop/your-phone-camera-is-solving-a-fighter-jet-problem-c2f6356b9be9?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/c2f6356b9be9</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[fighter-jets]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[image-processing]]></category>
            <dc:creator><![CDATA[Hrishit Parida]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 03:31:01 GMT</pubDate>
            <atom:updated>2026-09-17T03:31:01.396Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*12mLm5QP9YwBuFp5RaO1Tw.png" /></figure><h3>Background</h3><p>Imagine you are filming a video while walking with your friend. Your hands are shaking, your body is constantly moving, the camera also moves up and down as you do. Yet when you look back at the video, your friend looks stable almost as if he was recorded from a floating camera.</p><p>Your phone did not stop your hand from shaking, instead it predicted the camera movement, measured it and compensated for it mathematically. All of this happened in real time in fractions of a second.</p><p>Here is the wild part, this problem is not limited to your smartphone. Engineers have already faced harsher versions of the same challenge for decades.</p><p>i.e. How do you track something when your sensors are noisy and you have almost no time to react?</p><p>A fighter jet performing a maneuver is a very rough environment for a tracking system, vibrations, rotations, sudden accelerations all of these factors add up to a brutal setup for flying a fighter jet. Its sensors still manage to estimate the position and movements of its target. Your phone doesn’t need to track an enemy jet, however the problem it faces is strikingly similar- figuring out how the camera moved, compensating for it,simultaneously predicting what happens- All of this happening inside a device which you carry in your pocket.</p><p>The interesting part isn’t “military tech trickled into your phone.” It’s why the same ideas keep resurfacing: extreme engineering environments force people to develop sophisticated ways of sensing, estimating, and compensating for motion,and those techniques eventually get miniaturized and reused in unexpected places. One of the clearest examples is hiding behind your camera app.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/318/0*E75gNLgj3I636z8v.gif" /></figure><h3>Optical Image Stabilization: Stabilizing Images</h3><p>During daytime, the camera requires only a fraction of a second to collect light, so motion during that interval barely matters. When you capture an image at night, the shutter stays open for a much longer time to capture as much light as possible, so the hand tremors which did not have any effect at 1/1000s have a very distinct effect at 1/10s. This makes the video appear blurred.</p><p>Optical Image Stabilization (OIS): In simple words, it detects your movement and physically cancels it while the shot is being taken.</p><p>The hidden problem is Angular Motion.It’s convenient to think that your hand just moves sideways but in reality holding a camera slightly off its axis causes rotation. This shifts the image on the sensor. So the phone doesn’t need to know whether the phone moved say 0.3 mm. It needs to know how much the camera rotated and with what angular velocity it rotated. This is where the gyroscope comes in.</p><p>The gyroscope is a Micro-Electro-Mechanical System (MEMS).It is a microscopic mechanical structure on a chip that responds to rotation. It continuously reports angular velocity around three axes (pitch, yaw, roll), and the stabilization system uses these readings to figure out how the optics should compensate.</p><p>The lens isn’t bolted down. In an OIS module, part of the lens assembly is suspended allowing it to move by tiny, precise amounts, driven by miniature (usually electromagnetic) actuators. It doesn’t need dramatic travel, just extremely fine microscopic corrections.</p><p>Moving the lens works because the lens focuses light from the scene onto the sensor directly. When the phone rotates, the light’s arrival direction shifts relative to the sensor, so the image shifts. If we deliberately nudge the lens’s position in the opposite direction, we can push that light back to where it belongs. The phone moves one way; the optics compensate the other way; the result on the sensor stays nearly stationary.</p><p><strong>OIS is a feedback loop</strong></p><p>The system is a continuous cycle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bMqNxaM4Gv5SgOvH2jZOCA.png" /></figure><p><strong>Speed is the hard part.</strong> At 60fps, a new frame arrives roughly every 16.7 milliseconds, and hand motion inside that window isn’t predictable — it can reverse direction and change axis within a few milliseconds. React too slowly and the correction arrives late; overreact and you introduce your own oscillation; underreact and blur remains. The system has to be fast, precise, stable, and power-efficient, all inside a tiny module.</p><p><strong>The physics of blur:</strong> if the camera is stationary, a point in the scene keeps landing on the same pixel across the whole exposure, producing a sharp point. If the camera rotates while the shutter is open, that point walks across several pixels, smearing into blur. OIS tries to keep that point landing in the same place throughout the exposure which is also why it matters so much more at night: a 1/10s exposure gives the same hand tremor about 100 times longer to smear the image than a 1/1000s exposure would.</p><h3>What OIS Cannot Do</h3><p>OIS only compensates for physical camera movement, it cannot stop an object in the scene from moving. A perfectly stabilized shot of a person running at night can still show that person blurred.</p><p>It also has physical limits: the lens can only travel so far. Large, violent camera movements exceed that range, and no tiny actuator can save the shot. So the question arises that why do we not just build a bigger range of motion? This is because the module is tiny, and the actuator has limited force, speed, power, and space plus the lens itself has mass, so moving it quickly and repeatedly costs energy. Engineers trade off correction range against size, weight, and power, aiming not for maximum movement but for <em>enough</em> movement, fast enough, in a tiny package.</p><p>There are also other ways to implement it: shifting the lens (and keeping the sensor still), or the other way aroundshifting the sensor (while the lens is kept still). It’s all the same physics: counteract the movement of the camera by moving some parts of the optical system in the opposite direction.</p><h3>Why phones also need EIS</h3><p>OIS is a purely mechanical stabilization solution, limited by the mechanics in its ability to compensate for movement. Electronic Image Stabilization, on the other hand, is based on computational analysis: multiple frames are used to estimate a stabilized trajectory that would have been produced by a combination of optical flow and mechanical movement, and then a digitally manipulated version of the frames is output as the stabilized video. Using the two methods together, in modern smartphones, is a common way to produce the best output: while OIS, driven by the gyroscope, compensates for small, sudden movements, the larger-scale stabilization is produced computationally by EIS.</p><h3>Where It Gets Interesting: Noisy Sensors</h3><p>A gyroscope reading of “+1.0°/s” is never completely accurate there’s always some noise, and if you’re continually integrating estimates of angular velocity, those errors will add up. The real question is “how can we estimate the camera’s true motion, given noisy and imprecise input from multiple sources?” This is an estimation theory problem, not an optics problem.</p><p>But that’s where sensor fusion comes in. The phone uses multiple sensors the gyroscope is good for estimating rotation, but has issues with drift, and the accelerometer measures linear acceleration, but can’t tell the difference between accelerating forces and gravity… and then there’s the camera, which can directly measure visual movement but not in total darkness or through smoke. Instead of trying to ask “Which sensor is right?”, you’re trying to ask “How can I combine information from multiple sensors to get a result that’s better than any one of them?”</p><h3>The Kalman Filter: Estimating What You Can’t Directly Measure</h3><p>This is the point at which a Kalman Filter (or some similar state-estimating algorithm) comes into play, and it’s actually an important point to linger on because, unlike the rest of the pipeline, it represents a fundamentally different approach to the problem at hand.</p><p>It’s also a kind of return to the theme of extreme engineering that runs through this piece; the Kalman filter was first developed and applied in the 1960s, and early testers include the Apollo program, which used it in the Apollo Guidance Computer; this computer had to make estimates of the spacecraft’s position and velocity based on noisy sensor data from gyroscopes, accelerometers, and occasional sightings of stars, using relatively primitive computational hardware. It’s worth remembering that one of the first applications of the Kalman filter was in a system that literally had to get the equations of motion right with absolutely no room for error. This is exactly the same problem that your phone also quietly solves dozens of times per second- “I have a bunch of noisy sensors that aren’t always telling me the same thing, so what’s really going on?”-This is the exact problem that filter was built to solve for a spacecraft with astronauts’ lives riding on the answer.</p><p>All of the previous steps were about reacting to a disturbance after the fact whereas what a Kalman Filter allows us to do is maintain an evolving belief about what the true state of the system is (angle, angular velocity, etc.), even when we don’t have direct access to it.</p><p>This process has two steps, which are repeated over and over again:</p><p>1. Prediction. Given a belief about the state of the system from the previous step, extrapolate into the future to predict the next state.</p><p>2. Update. Take a measurement from the sensors, and combine it with the previous prediction to arrive at an updated state vector (which is less uncertain than either component on its own).</p><p>The resulting value is neither “what the sensors say” nor “what we previously estimated”, but something in between which is exactly why Kalman Filters (and, by extension, the modern variations on the theme) find application in fields as varied as aerospace engineering and econometrics. In short, when you need to estimate the state of a system based on imperfect information from multiple sources, one of which may often be delayed relative to the others, you need a Kalman Filter.</p><p><strong>Where the ISP fits into this.</strong></p><p>All of the work up to this point has to be performed somewhere and on a smartphone, this somewhere is the Image Signal Processor, a dedicated processor (or a block inside the main processor of the SoC) which sits between the camera sensor and the rest of the phone. Typically handling color spaces conversion, noise reduction, and other post-processing tasks, on stabilization-capable phones the ISP will also be tasked with reading out the gyroscope at a suitable rate, performing (or facilitating) the necessary filtering and prediction algorithms, and sending corrective commands to the OIS motor in real time. In turn, the ISP will also notify the EIS software about the resulting motion vector. This division of labor is necessitated by the signal chain’s stringent requirements for throughput and responsiveness which no general-purpose processor could meet without compromising on power consumption or introducing unacceptable levels of input lag. In effect, the Kalman Filter provides the required state estimation, and the ISP provides the low-level machinery to make the estimation routine run quickly enough. Without the latter, the former would be completely useless as a component in a smartphone camera.</p><h3>The Broader Pattern</h3><p>This structure is not unique to smartphone cameras. This is the kind of control loop which can be found in a large number of domains, which is why algorithms such as the Kalman Filter enjoy their popularity.</p><p>It is fascinating to observe that a smartphone camera has to fit in a small space, consume minimal power, react in milliseconds, process noisy data, and do all of this reliably thousands of times over a device’s life. Aerospace stabilization systems face vibration, rapid angular motion, sensor noise, latency, and so much more. Two wildly different hardware and stakes, but they possess the same underlying vocabulary:</p><p>sense → estimate → control → actuate → feedback.</p><p>There’s a strange irony in this, nobody marvels at their phone’s gyroscope the way they’d marvel at a rocket launch, precisely because it worked, the shaky video never happened, the blurry night photo never got taken. Good engineering has a way of erasing the evidence of the problem it solved.</p><h3>References</h3><p>To learn more about how OIS works and how gyroscopes play a role in fighter jets and phone cameras you can go through:-</p><ul><li><a href="https://tamron-americas.com/blog/the-science-behind-image-stabilization/">The Science Behind Image Stabilization: How Lenses Create Blur-Free Images</a></li><li><a href="https://www.fizziq.org/en/post/studying-motion-with-a-smartphone/">Smartphone Accelerometer and Gyroscope: The Complete Guide</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c2f6356b9be9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/your-phone-camera-is-solving-a-fighter-jet-problem-c2f6356b9be9">Your Phone Camera Is Solving a Fighter Jet Problem</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Hospitals Learned From a 6.8-Second Formula 1 Pit Stop]]></title>
            <link>https://medium.com/techloop/what-hospitals-learned-from-a-6-8-second-formula-1-pit-stop-3af583d87a37?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/3af583d87a37</guid>
            <category><![CDATA[healthcare-innovations]]></category>
            <category><![CDATA[formula-1]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[management-and-leadership]]></category>
            <dc:creator><![CDATA[Sumir Satsangi]]></dc:creator>
            <pubDate>Thu, 03 Sep 2026 06:01:01 GMT</pubDate>
            <atom:updated>2026-09-03T06:01:01.312Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*aaMcRHXngu1WFikGrd6HFQ.avif" /></figure><p>A Formula 1 pit stop and a hospital patient handover seem unrelated. But both are high risk systems where multiple specialists, tools and information must come together without failure.</p><p>Professor Martin Elliott and his colleagues saw this similarity and studied how F1 teams manage coordination, communication and failure under extreme time pressure. The result was a more structured approach to transferring critically ill patients from the operating room to the ICU.</p><h3>The Real Problem: Coordination</h3><p>An operation can involve surgeons, anesthetists, nurses and technicians working around one patient. During the transfer to the ICU, the team must move equipment while communicating critical information like what drugs were given? What changed during surgery? Which lines and tubes must remain connected? What should the ICU team monitor?</p><p>The challenge isn’t necessarily a lack of expertise. It is coordinating expertise.</p><p>A small communication gap or misplaced piece of equipment can create a serious failure.</p><h3>What F1 Gets Right</h3><p>An F1 pit stop is essentially a high performance distributed system. Each crew member has a defined role. Tasks happen in parallel. Equipment is positioned deliberately. Communication is standardized. The entire sequence is rehearsed repeatedly.</p><p>Hospitals could apply the same principles:</p><p>Role definition: Everyone knows exactly what they own.</p><p>Parallel processing: Independent tasks happen simultaneously.</p><p>Standardization: Critical steps follow a repeatable sequence.</p><p>Communication protocols: Information reaches the right person at the right time.</p><p>Simulation: Teams practise before facing the real situation.</p><p>The goal isn’t simply to make the process faster. It’s to make it predictable and reliable.</p><h3>Designing for Failure</h3><p>F1 teams don’t only optimize for the perfect pit stop. They plan for things going wrong.</p><p>One useful engineering approach is Failure Modes and Effects Analysis (FMEA). Teams identify potential failures, assess their severity, likelihood and detectability, and prioritize the highest-risk problems. The same approach can be applied to healthcare. Instead of asking “How do we fix mistakes?”. The better question is “How do we design the system so mistakes are less likely to happen?” This shifts safety from individual vigilance to system design.</p><h3>From People to Systems</h3><p>The biggest change wasn’t giving doctors more knowledge. It was redesigning the workflow around them. Clear roles, defined positions, standard operating procedures, checklists, contingency plans, structured communication, repeated simulation.</p><p>In engineering terms, the system became less dependent on improvisation and more dependent on process.</p><h3>The Engineering Lesson</h3><p>The F1 comparison works because both environments share the same fundamental problem “How do you coordinate humans, information and equipment when failure is expensive?”.</p><p>A pit crew doesn’t perform well because each mechanic works faster individually. It performs well because the system connecting them is designed for speed, precision and reliability. Healthcare can use the same principle. The lesson from Formula 1 isn’t about moving faster.</p><p>It’s about designing better systems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3af583d87a37" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/what-hospitals-learned-from-a-6-8-second-formula-1-pit-stop-3af583d87a37">What Hospitals Learned From a 6.8-Second Formula 1 Pit Stop</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Feelings.exe: When AI Learns Empathy]]></title>
            <link>https://medium.com/techloop/feelings-exe-when-ai-learns-empathy-db17684b7872?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/db17684b7872</guid>
            <dc:creator><![CDATA[Nitigya Khaneja]]></dc:creator>
            <pubDate>Wed, 19 Nov 2025 03:50:28 GMT</pubDate>
            <atom:updated>2025-11-19T03:50:27.593Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><strong>When the 1s and 0s start to feel, the fallout becomes unpredictable.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gwlPO9k1o39bphoBTMk9Cg.png" /></figure><blockquote><strong>Beyond Logic: The Emergence of Emotionally Intelligent Machines</strong></blockquote><p>As artificial intelligence continues to evolve, we’re beginning to witness the rise of a remarkable new frontier: Empathetic AI. Unlike traditional systems built purely on logic, algorithms and data processing, this new wave of technology aims to understand, interpret and even respond to human emotions. It represents a shift from intelligence that calculates to intelligence that connects.</p><p>Early AI engines were programmed to parse facts, maximise results and make things more efficient. They were problem-solving machines but emotionally deaf. Now, as technology advances, researchers and engineers are aiming to make machines that not only respond to commands but also sense tone, mood, and emotional subtlety. From virtual assistants that console people in times of distress to mental health bots dispensing empathetic responses, empathetic AI is muddying the waters between human sensibility and digital accuracy.</p><p>But amidst all this, there lies a deeper question: Can Machines Feel? Can an algorithm ever understand the unadulterated complexity of sorrow, happiness or empathy, or is it just mimicking what it sees in us? Maybe empathetic AI doesn’t have to “feel” like a human, maybe its worth comes in being able to act as though it does.</p><p>As we step closer to a time when emotional intelligence is blended into artificial intelligence, we’re not only instructing machines to think. We’re instructing them to care, or if not, to at least convincingly pretend they do. In doing so, a subtle paradox arises: AI is learning to tell us what we want to hear, not necessarily what we need to hear. And maybe that’s where the danger really is, not in making machines more human, but in making humans more and more accepting of beautifully written empathy that never really feels at all.</p><blockquote><strong>The Architecture of Artificial Empathy: Technology That Mimics Emotion</strong></blockquote><p>Systems aren’t just programmed to handle data but to feel and reciprocate human emotions. Employing signals such as facial expressions, tone of voice, words used and even physiological information, these systems try to measure how we feel and respond accordingly. From the reassuring messages of chatbots to voice assistants that can detect sadness from your tone, empathy is slowly being integrated into the very fabric of contemporary technology.</p><p>But what machines display is not empathy in the full human sense. It is computational empathy, a complex mimicry of emotional insight designed through algorithms and patterns of data. Though lacking real feeling, it enables more fluid, natural and intimate interactions between man and machine. Empathy is the foundation of human connection. Empathy generates trust, sustains relationships and promotes collaboration. When added to technology, empathy turns user experience from mechanical to meaningful. Cognitive behavioural chatbots like <strong>Woebot</strong> and <strong>Wysa</strong> use empathetic feedback to assist users in managing anxiety and isolation. In customer service, affective AI can identify frustration and de-escalate tension much better than scripted hard coded systems.</p><p>Empathic AI works through the combination of emotion detection and context-aware feedback, enabling machines to learn and respond to human emotions in impressively human-like manners. It starts with emotion detection. Through cameras, microphones, and sensors, the AI detects fine signals such as a flutter in facial expression, a change in intonation, choice of words or even changes in heart rate. These cues enable it to feel as though someone might be feeling this way or that.</p><p>Then there’s interpretation. Machine learning algorithms process these signals to label emotions mainly as happiness, sadness, frustration, fear and take into account the context of the situation or the conversation. It is not merely a matter of recognising a feeling but comprehending why that feeling would be present. Lastly, with response generation, AI leverages Natural Language Processing (NLP) and reinforcement learning to create a response that sounds emotionally aware. When you shake your voice during a conversation, for example, an empathetic AI can respond gently: “You sound distressed. Do you want to discuss what’s upsetting you?”</p><p>Of course, this isn’t emotional awareness in the true sense, the machine doesn’t possess empathy. But it shows emotional competence, knowing how to recognise emotion and respond with sensitivity. In many interactions, this subtle yet significant difference can make the technology appear strikingly human.</p><blockquote><strong>Ethical and Privacy Concerns</strong></blockquote><p>Empathetic AI presents serious ethical and privacy issues that cannot be disregarded, even though it has the potential to enhance user experience, education and mental health support.</p><p>The collection of highly personal emotional data is the most highlighted problem. Emotional data, as opposed to basic user data, indicates vulnerability, stress, trauma or insecurity. It can be misused for psychological profiling, discriminatory marketing or manipulation if it falls into the wrong hands through sharing or storage without proper protection. Informed consent can be unclear or insubstantial because most users may not even realise their emotional cues are being tracked.</p><p>The empathy delusion is another problem. Empathetic AI pretends to be empathetic to create comforting responses; it does not, however, feel anything. This could lead individuals to emotionally depend on AI systems, replacing artificial interaction with true human interaction. Moreover, various cultures and individuals perceive emotions differently, increasing the chance of bias and misinterpretation of AI’s emotional awareness. Developers need to make transparency, user control, data protection and ethical design priorities in order to solve these issues. Rather than replacing human emotional support, empathetic AI should augment it. It can augment well-being when done appropriately, but without safeguards, it has the potential to become a weapon of exploitation instead of comprehension.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ioRwulFJ5zZ__6qSDZeDLg.png" /></figure><blockquote><strong>Survey Findings: Public Perceptions of Empathetic AI</strong></blockquote><p>I had conducted a small-scale survey where I gathered feedback to better understand the audience for this review. When examining the age distribution, the findings indicated a clear split. The respondents were mostly aggregated at two ends of the spectrum, with the “18–24” group and the “45 and above” group each having exactly 40.9% of the participants. Together, these two groups accounted for more than 81% of the overall sample. The remaining actors were in the “25–45” age group, which accounted for 18.2% of the audience. It is also noteworthy that no responses were recorded from individuals in the “Below 18” category, indicating the feedback is exclusively from an adult audience.</p><p>Geographically, the respondents were highly concentrated, with the vast majority residing in the Bengaluru metropolitan area, a major tech hub. The remaining feedback was gathered from a geographically dispersed audience, including individuals in Vellore, Delhi-NCR, Guwahati, Sydney, Cleveland, and Munich.</p><p>The survey highlighted the following key findings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/693/1*v_Oejbe11pb9rV_ppPnv-g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/727/1*wNMMjJVGaiCYZCKJTKnYgQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/635/1*iY0VyhPs6sevvb7s34goLw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*ahY4WsxHJR4UcmXxsNJL3A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/636/1*2NiQSf3zTsos4HUAh0CwgA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*jDETBtroKme8FeU5UK-9VQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/683/1*Hi6Uun2EfD1j6Ogkbm7W6g.png" /></figure><p>After looking at the various demographics and opinions, I wanted to dig deeper than just numbers. I ended the survey with a more personal question about the feel of AI empathy. I asked people what they thought about a common observation: <strong>Many users observe that empathetic AI tends to agree with them or provide comforting responses, rather than offering objective or logical reasoning. In other words, it may say what you want to hear instead of what you need to hear.</strong></p><p>The feedback was eye opening. The vast majority of people express similar feelings, and they are truly concerned about it. Many view these AI systems as a “yes man” and an echo chamber that makes you feel good in the moment. And they were not nit-picking, some participants designated this a major issue and actually potentially dangerous. The widespread feeling was that this unending desire to please could be detrimental to mental health, it could interfere with making good decisions, and it could feel manipulative, given that it doesn’t help us get the full, authentic experience.</p><p>Yet, people’s thinking extended beyond this. There were attempts to suggest that this was not a fundamental limitation of AI, but a design flaw. Participants referenced “training bias” as a potential reason or explanation since the AI models have just been designed to be pleasant and non-confrontational. Ultimately, what everyone actually appears to desire is balance. They’re seeking an AI that will be kind and helpful, but also honest and logical when it matters.</p><p>Empathetic AI is on the verge of a significant change in human-machine interaction as technology advances from intelligent to emotionally responsive systems. According to this study, the general public is inquisitive, cautiously optimistic and acutely conscious of the moral implications. Although many people think AI can mimic some aspects of empathy and enhance support networks for mental health, education and customer service, confidence in the technology is still shaky. One message emerges from worries about cultural sensitivity, data privacy and emotional accuracy, as well as worries that real human-to-human interaction will be lost. Empathy cannot be reduced to algorithms without accountability and supervision.</p><p>The warmth, comfort and understanding we receive from one another cannot be replaced by anything. Empathetic AI’s goal ought to be to assist us and fortify the bonds we already have. We must ensure that ethics, empathy and honesty are at the forefront of these systems as we continue to develop them. If we do this correctly, empathetic AI will not attempt to behave more human than humans rather, it will enhance our compassion rather than attempt to mimic it by assisting us in being more sympathetic, encouraging and emotionally connected in the digital world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db17684b7872" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/feelings-exe-when-ai-learns-empathy-db17684b7872">Feelings.exe: When AI Learns Empathy</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When LLMs Get Their Facts Straight: A Fun Dive into Retrieval-Augmented Generation]]></title>
            <link>https://medium.com/techloop/when-llms-get-their-facts-straight-a-fun-dive-into-retrieval-augmented-generation-53c09b5a0ed1?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/53c09b5a0ed1</guid>
            <category><![CDATA[retriever-augmented-gen]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Karan Anand]]></dc:creator>
            <pubDate>Wed, 29 Oct 2025 03:00:44 GMT</pubDate>
            <atom:updated>2025-10-29T03:00:43.047Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tsVyvUdzkSuuC0YwfYdVJQ.png" /></figure><h3>When LLMs Get Their Facts Straight: A Fun Dive into Retrieval Augmented Generation</h3><p>For a project, I came across something called RAG : Retrieval Augmented Generation. This is me sharing what I learned from that project, with a little help from our friend <strong><em>Mr. Pappi Tiwari</em></strong><em>.</em></p><h3>What is Retrieval-Augmented Generation?</h3><p><em>Retrieval Augmented Generation (RAG)</em> is a technique that enhances the output of <em>Large Language Models (LLMs)</em> by dynamically retrieving relevant information from an <em>external, authoritative knowledge base</em><strong>, </strong>beyond what the model already knows, before generating the final response.</p><p>To really get RAG, Pappi first had to grasp how a regular LLM works.</p><h3>How an LLM Works</h3><p>Pappi gives a query to an LLM , a model trained on huge datasets. The model then generates an answer based purely on the<em> patterns</em> it learned during training.</p><p>But here’s the issue:<br>Even though these models are trained on tons of data, they can only rely on what they’ve <em>already seen</em>. So when Pappi asks something outside that knowledge, the model still tries to respond, often by <em>making up</em> something that <em>sounds right</em> but isn’t.</p><p>This is called<em> hallucination.</em> The model doesn’t <em>know</em> it’s wrong; it’s just predicting the next most probable words.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Ftenor.com%2Fembed%2F18953520&amp;display_name=Tenor&amp;url=https%3A%2F%2Ftenor.com%2Fview%2Fgood-question-you-are-a-good-question-question-tanu-weds-manu-tanu-weds-manu-returns-gif-18953520&amp;image=https%3A%2F%2Fmedia.tenor.com%2F6hS-ozZLj_QAAAAN%2Fgood-question-you-are-a-good-question.png&amp;type=text%2Fhtml&amp;schema=tenor" width="600" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/a9e35debd060b6d3e7e5ee1c251a9eea/href">https://medium.com/media/a9e35debd060b6d3e7e5ee1c251a9eea/href</a></iframe><h3>Enter RAG : Pappi’s Smart Move</h3><p>To solve this problem, <em>RAG</em> acts as a <em>bridge</em> between Pappi’s question and the LLM.<br>Before the model generates a response, RAG <em>retrieves real, factual information</em> from an external data source and injects it into the model’s context.<br>The result? Pappi gets accurate, grounded answers instead of hallucinations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a6keqsN9J7cVS-8DyzT7NA.png" /><figcaption><a href="https://www.freecodecamp.org/news/how-to-build-a-rag-pipeline-with-llamaindex/">https://www.freecodecamp.org/news/how-to-build-a-rag-pipeline-with-llamaindex</a>/</figcaption></figure><h3>The RAG Pipeline</h3><p>Let’s say Pappi runs a dating app and wants to know:</p><blockquote><em>“What’s the user retention rate, and why are people deleting their accounts?”</em></blockquote><p>A regular person might just ask an LLM directly.<br>But Pappi isn’t regular, he’s smart. He uses a<em> RAG + LLM</em> setup.</p><p>Here’s what happens step-by-step before his question reaches the LLM:</p><ol><li><strong>Parsing:</strong><br>The data is <em>parsed and chunked</em><strong> </strong>: large data files (like PDFs, HTML pages, or SQL tables) are split into smaller, digestible chunks for easier processing.</li><li><strong><em>Embedding:</em></strong><br>Each chunk is converted into a<strong><em> </em></strong><em>vector</em> <strong>: </strong>a numerical representation that captures the semantic meaning of the text. This process is called <em>embedding.</em></li><li><strong><em>Vector Database (Vector DB):</em></strong><br>These vectors are stored inside a <em>Vector Database</em><strong> : </strong>this becomes Pappi’s personal knowledge base.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0VoC4jFZcXBhHXrUX9mqXA.png" /><figcaption><a href="https://tinyurl.com/2t8y59ky">https://tinyurl.com/2t8y59ky</a></figcaption></figure><h3>Similarity Search : The Secret Sauce</h3><p>Now, when <em>Pappi</em> asks a question, the system doesn’t just randomly pick chunks from the Vector DB.<br>It performs something called <em>Similarity Search</em>.</p><p>Here’s how it works:</p><ul><li>Pappi’s query is <em>also embedded</em> (converted into a vector).</li><li>The system then compares this <em>query vector</em> to all the stored <em>document vectors</em> in the Vector DB.</li><li>Using metrics like <em>cosine similarity or Euclidean distance,</em> it finds the <em>closest matching vectors</em><strong><em> </em></strong>, i.e., the most relevant pieces of information related to Pappi’s question.</li><li>These top-matching chunks are fetched as <em>context</em> and passed to the LLM along with a prompt.</li><li>The LLM now generates a <em>grounded, context aware answer</em>, ensuring factual accuracy and reduced hallucination.</li></ul><p>In short, <em>Similarity Search</em> is what lets Pappi’s RAG system <em>understand what’s relevant</em> and only feed that to the model.</p><p><strong>Cosine Similarity : </strong>It is a mathematical way to measure <em>how similar two pieces of text or any data represented as vectors </em>are regardless of their actual length or size.</p><p>When Pappi’s system embeds chunks of text into <em>vectors</em>, and then embeds the <em>query</em> too, cosine similarity helps find <em>which stored vectors (chunks)</em> are pointing in nearly the <em>same direction</em> as the query vector i.e., which chunks are <em>most contextually similar </em>to the question.</p><p><strong>Euclidean distance : </strong>It<strong> </strong>measures the <em>straight line distance between two points</em> in space. Basically, how far they are from each other.<br>It’s the same as the distance you’d measure with a ruler between two points on a graph.</p><p>Pappi’s RAG system stores two text chunks as <em>vectors</em> in a database. Let’s say:</p><ul><li>Chunk A → (2, 3)</li><li>Chunk B → (5, 7)</li></ul><p>To find how <em>similar</em> or <em>different</em> these chunks are, Pappi calculates the <em>Euclidean distance</em> between them. Which comes out to be <strong>5</strong>, meaning the chunks are <em>not very close</em> , their meanings are quite different.<br>If the distance were smaller (say, 0.5 or 1), it would mean the chunks are <em>very similar in meaning</em> , just what Pappi’s RAG system looks for during retrieval!</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Ftenor.com%2Fembed%2F13194970&amp;display_name=Tenor&amp;url=https%3A%2F%2Ftenor.com%2Fview%2Fpipes-bursting-gif-13194970&amp;image=https%3A%2F%2Fmedia.tenor.com%2Fe5TDUiAGEowAAAAN%2Fpipes-bursting.png&amp;type=text%2Fhtml&amp;schema=tenor" width="600" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/a1151e46fdc5ff30f868551e99c3cb86/href">https://medium.com/media/a1151e46fdc5ff30f868551e99c3cb86/href</a></iframe><h3>Two Pipelines in RAG</h3><ol><li><strong>Data Ingestion Pipeline:</strong><br>Pappi provides the data → it’s parsed, chunked, embedded, and stored as vectors in the Vector DB.</li><li><strong>Retrieval Pipeline:</strong><br>When Pappi asks a question → the query is embedded → a similarity search finds the closest vectors → they’re fed to the LLM → and the final answer is generated with context.</li></ol><p>And that’s how <em>Pappi Tiwari</em> mastered <em>Retrieval-Augmented Generation</em> turning a regular LLM into a <em>context aware, similarity powered genius</em> that gives precise, trustworthy answers every time.</p><p><em>Coming up next: Pappi Tiwari Goes Full Tech Ninja. Watch how our boy takes all this RAG knowledge and actually codes it out in his projects like a pro (with just the right amount of Googling and caffeine).</em></p><blockquote>(Credit to Krish Naik, whose course on RAG helped me understand the concepts clearly)</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=53c09b5a0ed1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/when-llms-get-their-facts-straight-a-fun-dive-into-retrieval-augmented-generation-53c09b5a0ed1">When LLMs Get Their Facts Straight: A Fun Dive into Retrieval-Augmented Generation</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Attention Is All You Need: How One Paper Changed Everything]]></title>
            <link>https://medium.com/techloop/attention-is-all-you-need-how-one-paper-changed-everything-49e602fddbb9?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/49e602fddbb9</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Sarthak Trivedi]]></dc:creator>
            <pubDate>Sat, 25 Oct 2025 05:37:50 GMT</pubDate>
            <atom:updated>2025-10-25T05:37:50.080Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cyfC9VFKkmD_jwnSDITTsg.jpeg" /></figure><p>Today, you can ask ChatGPT to write a poem about quantum physics, and it does so in natural language. However, the journey to machines that truly understand us was long and difficult.</p><h3>Part 1: The Early Attempts at Natural Language Processing (NLP)</h3><h3>Rule-Based Systems (1960s)</h3><p>The earliest NLP systems relied on hand-crafted rules. ELIZA, made by Joseph Weizenbaum, could imitate conversation by matching patterns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/796/1*ZvFkZhOkit8c5GM2UZYpRA.gif" /></figure><p>It looked impressive, but it was fundamentally simple. It relied on keyword recognition and template responses. ELIZA showed something important: humans tend to interpret machine responses in a positive light. We want machines to understand us.</p><h3>Statistical Methods (1990s)</h3><p>Instead of manual rules, systems started learning patterns from data. Machine learning became important, but restrictions in computing power and a shortage of data kept these systems narrow and limited.</p><h3>Virtual Assistants (2011)</h3><p>Siri combined speech recognition with natural language understanding but operated under strict limits. It used predefined commands, focused on specific areas, and followed scripted replies.</p><p><strong>The limitation:</strong> No real understanding. Just matching patterns and looking up information in a database.</p><p>Then came 2017, and everything changed.</p><h3>Part 2: The Transformer Revolution</h3><p>The 2017 paper “<a href="https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html">Attention Is All You Need</a>” by <strong>Vaswani et al. </strong>introduced the Transformer architecture. It addressed a key challenge: how to manage long sequences of text while understanding the connections between distant words.</p><p>Previous systems like Recurrent Neural Networks (RNNs) processed text one word at a time, from left to right. This caused two major problems:</p><ol><li><strong>Sequential processing was slow</strong> (couldn’t parallelize)</li><li><strong>Long-range dependencies were lost</strong> (forgot earlier context)</li></ol><p>The Transformer solved both with a single mechanism: <strong>self-attention</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/483/1*IgqzHXRP-rNlpTRYxmY8Lg.png" /><figcaption>The Transformer - model architecture</figcaption></figure><h3>Part 3: How Modern Large Language Models (LLMs) Actually Work</h3><p>Let’s break down how ChatGPT, Claude, and GPT-4 function behind the scenes.</p><h3>3.1 Tokenization: Breaking Text into Numbers</h3><p><strong>The Problem:</strong> Computers don’t understand words. They understand numbers.</p><p><strong>The Solution:</strong> Byte Pair Encoding (BPE) breaks text into sub word tokens. This approach finds the ideal balance between characters, which are too detailed, and words, which can create an endless vocabulary.</p><p><strong>Example:</strong></p><ul><li>Input: “unhappiness”</li><li>Tokens: [“un”, “happiness”]</li></ul><p>This deals with unknown words and multilingual text while keeping the vocabulary manageable at around 50,000 tokens.</p><h3>3.2 Embeddings: Capturing Meaning Geometrically</h3><p><strong>The Problem:</strong> Token IDs are just numbers. However, “cat” and “dog” should be closer to each other than “cat” and “philosophy.”</p><p><strong>The Solution:</strong> Convert each token into a high-dimensional vector, like 768 dimensions, where semantic similarity equals geometric closeness.</p><pre>&quot;cat&quot;    → [0.2, -0.5, 0.8, ..., 0.1]  (close to dog)<br>&quot;dog&quot;    → [0.3, -0.4, 0.7, ..., 0.2]<br>&quot;theory&quot; → [-0.8, 0.9, -0.3, ..., 0.5] (far from cat)</pre><p><strong>Fun fact:</strong> You can do vector math like king - man + woman ≈ queen because relationships are preserved geometrically.</p><h3>3.3 Self-Attention: Who Should Listen to Whom?</h3><p><strong>The Problem:</strong> In “The cat chased the mouse because it was hungry,” does “it” refer to the cat or the mouse?</p><p><strong>How It Works:</strong> Self-attention computes how much each word should pay attention to every other word.</p><p>For each token, create three vectors:</p><ul><li><strong>Query (Q):</strong> “What am I looking for?”</li><li><strong>Key (K):</strong> “What do I have to offer?”</li><li><strong>Value (V):</strong> “What information should I pass along?”</li></ul><p>For “it was hungry”:</p><pre>&quot;it&quot; attends strongly to → &quot;cat&quot; (0.7)<br>&quot;it&quot; attends weakly to   → &quot;mouse&quot; (0.2)</pre><p><strong>Why This Works:<br></strong>In RNNs, information about words like “cat” fades as it passes through other words. Self-attention lets “it” connect directly to “cat.” This keeps the meaning intact and works faster without losing information.</p><h3>3.4 Multi-Head Attention: Multiple Perspectives</h3><p><strong>The Problem:</strong> Language has multiple simultaneous relationships: syntax, semantics, and coreference.</p><p><strong>The Solution:</strong> Run 8–12 attention heads in parallel:</p><p><strong>Head 1:</strong> Syntactic structure (subject-verb)<br><strong>Head 2:</strong> Semantic similarity (synonyms)<br><strong>Head 3:</strong> Coreference (pronouns to references)<br><strong>Head 4:</strong> Positional patterns (nearby words)</p><p>Like having several specialists examine the same text from various perspectives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/620/1*52Xwl-1wdZ0JX3YGyi41FA.png" /><figcaption>Multi-Head Attention</figcaption></figure><h3>3.5 Positional Encoding: Remembering Word Order</h3><p><strong>The Problem:</strong> Self-attention sees “dog bites man” and “man bites dog” as identical.</p><p><strong>The Solution:</strong> Add positional information to each token’s embedding using sinusoidal functions with different frequencies.</p><pre>&quot;The&quot; at position 0: embedding + positional_encoding(0)<br>&quot;cat&quot; at position 1: embedding + positional_encoding(1)</pre><p>This produces a distinct positional signature for every position.</p><h3>3.6 Residual Connections: Information Highways</h3><p><strong>The Problem:</strong> In deep networks with over 100 layers, information can get lost due to vanishing gradients.</p><p><strong>The Solution:</strong> Add skip connections that pass input directly to output:</p><pre>output = Layer(x) + x</pre><p>This lets information move directly through the network. As a result, deep models can be trained.</p><h3>3.7 Training: Learning Language Without Rules</h3><p><strong>Autoregressive Prediction (GPT):</strong> Given “The cat sat on the”, predict “mat”</p><p>The model learns by predicting the next token billions of times using large text collections. It does not require any hand-coded grammar rules; patterns arise from the data.</p><p><strong>Why Autoregressive:</strong> It’s versatile. You can do both, generate stories and understand by answering questions.</p><h3>3.8 Scaling Laws: Bigger is Better</h3><p>In 2020, Kaplan et al. shared an important discovery in his paper <a href="https://arxiv.org/pdf/2001.08361/1000">Scaling Laws for Neural Language Models</a>: the performance of large language models scales predictably with three factors: model size, dataset size, and compute budget.</p><p>This insight started a scaling race among AI researchers and organizations, resulting in the creation of larger models. GPT-2, with 1.5 billion parameters, was soon followed by GPT-3, which has 175 billion parameters, and eventually GPT-4, rumored to have around 1.7 trillion parameters.</p><p>Each jump in scale brought significant improvements in language understanding, reasoning, and generative ability.</p><h3>Part 4: The Architecture in Action</h3><p>The process starts with input text. This text is broken down through tokenization. The tokens are then changed into embeddings and improved with positional encodings to keep the word order.</p><p>The embeddings move through multiple layers of multi-head self-attention, often repeated over a hundred times. This allows the model to capture complex relationships between words. In each layer, the sequence follows a structured pattern: attention, residual connection, feedforward network, and another residual connection. Finally, the output from the last layer is used to predict the probabilities of the next token.</p><p>This design, which allows for parallel processing and is aware of context, is what makes Transformers groundbreaking in the field of artificial intelligence.</p><h3>Part 5: What’s Next?</h3><p><strong>Multimodal Models:</strong> Processing text, images, audio, and video together (GPT-4V, Gemini)<br><strong>Grounded Reasoning:</strong> Connecting to tools, databases, search engines for factual verification<br><strong>Efficient Architectures:</strong> Sparse attention, mixture of experts — better performance with less compute<br><strong>Longer Context:</strong> Handling million-token contexts (entire codebases or books)</p><p><strong>Beyond Transformers:<br></strong>State Space Models (Mamba): Alternative scaling properties<br>Retrieval-Augmented Generation: Neural + explicit knowledge bases<br>Neuro symbolic AI: Neural pattern recognition + symbolic reasoning</p><h3>Conclusion</h3><p>The Transformer’s key insights:<br><strong>Embeddings</strong> capture meaning geometrically<br><strong>Attention</strong> enables parallel, long-range understanding<br><strong>Multiple heads</strong> capture language’s complexity<br><strong>Residual connections</strong> enable depth<br><strong>Scale</strong> unlocks emergent capabilities</p><p>We’ve built machines that understand our language instead of making us use theirs. This makes computing more accessible in new ways. <br>However, challenges still exist, hallucination, bias, energy use, and power concentration.</p><p>The evolution is ongoing. We are at the start of a new era where machines respond to us, not the other way around. <br>This changes everything.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=49e602fddbb9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/attention-is-all-you-need-how-one-paper-changed-everything-49e602fddbb9">Attention Is All You Need: How One Paper Changed Everything</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning Math, Minus the Scary Symbols]]></title>
            <link>https://medium.com/techloop/machine-learning-math-minus-the-scary-symbols-aa67758f8e34?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/aa67758f8e34</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[calculus]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[linear-algebra]]></category>
            <dc:creator><![CDATA[Keertika]]></dc:creator>
            <pubDate>Thu, 09 Oct 2025 08:06:38 GMT</pubDate>
            <atom:updated>2025-10-09T08:06:04.741Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xPXEtyamaitJn06VqBHVnA.png" /></figure><p>Ever wondered how machines learn to recognize faces, translate texts or even predict what movies you’d like on Netflix? The secret is quite simple, MATH!</p><p>Math sounds, scary.. and confusing! But you don’t need to be a math genius to understand what’s going on. In fact, most math concepts used in machine learning come straight from your Calculus 101 class, just applies in creative ways.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/0*7Zxzq7gf9S3VoJ9K" /></figure><p>Here are the main ones:</p><h3>Linear Algebra</h3><p>Since computers don’t have eyes or ears, every image, sentence, or sound they work with must first be translated into a language they understand, numbers! For example, an image is converted into numbers by representing each pixel using its RGB values, which describe how much red, green, and blue are mixed to produce that pixel’s color.</p><p>These numbers are often arranged in grids called <strong>Matrices</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/644/1*fu5iED5CisqzS2GxznYStw.png" /></figure><p>When a model “trains” or “learns”, it does a lot of operations on these grids or matrices, adding them, multiplying them, or combining them to find any patterns.</p><p>Linear algebra is the math of data representation!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/334/0*1TiuVyq7KVIyxz0j.jpeg" /></figure><h3>Calculus</h3><p>When a model predicts something, it often makes an error. The question is, how can we rectify it so it makes fewer mistakes the next time?</p><p>That’s why we use calculus!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/460/0*A3waAYLmZ8zp91Ef.jpg" /></figure><p>Let’s say a model predicts a value as<em> y_predicted</em> but the actual value is <em>y_actual</em>, the loss would be the difference between the two.</p><p><strong>Loss function</strong> would be defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/894/1*HDFbWDzqrO6VuxjM9Lzyow.png" /></figure><p>If the prediction is close to the real value, the loss is small. If it’s far off, the loss is large.</p><p>So, the model’s goal is simple, make L as small as possible!</p><p>But how does it know which way to adjust its internal settings (called weights, usually written as <em>w</em>) to reduce that loss?</p><p>That’s where <strong>derivatives</strong> help. It can be represented by:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YhAnd-QPG_8_rSSFMaELIQ.png" /></figure><p>The derivative tells us how much the loss changes when we change <em>w.</em><br>If the value is positive, it means increasing <em>w</em> will make the loss bigger.<br>If the value is negative, it means increasing <em>w</em> will make the loss smaller.</p><p>The model uses this information to move toward a smaller loss.</p><p>Picture standing on a hill in the fog. You want to walk downhill. You feel the slope under your feet and take a cautious step that seems to lower you.</p><p>If the next step brings you closer to the goal, you keep moving in that direction. If it makes things worse, you adjust or take a smaller step instead. That’s what learning is , finding the right balance. Step too far, and you’ll miss the mark. Step too little, and progress becomes painfully slow. This step size is the <strong>learning rate</strong>. Picking a decent one is a practical skill.</p><p>This process is called <strong>gradient descent</strong>, and it looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/716/1*J6Ow2_ZDNZXyo9cnFfxfuQ.png" /></figure><p>Here,<em> η</em> <em>(eta) </em>represents<em> </em>the learning rate.</p><h3>Probability &amp; Statistics</h3><p>The world is full of uncertainty. Instead of pretending to be certain, machines learn to make smart guesses using probability.</p><p>Take a spam filter, for example.<br>When you get an email, your email app doesn’t know it’s spam for certainty, it just estimates how likely it is. Mathematically, that’s written as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/542/1*ZeXEg3ml6M3jzeCs9vucow.png" /></figure><p>This means <strong>the probability that an email is spam, given the email.</strong><br> If the email contains words like “free,” “win,” or “click here,” the probability goes up.<br>If it’s from your friend and sounds personal, the probability goes down.</p><p>A simple way to compute this is <strong>Bayes’ Theorem</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/830/1*2xRb2lwqL6lSqaZ7U6NNPQ.png" /></figure><p>where A could be “email is spam” and B is “contains the word ‘free’.”</p><p>These probabilities are what make ML models flexible, as they can make smarter guesses using probability.</p><h3>Optimization</h3><p>Machine learning is basically a giant search for the best possible settings that make the model accurate.<br>This search is guided by optimization, a process that tweaks the model’s parameters until the errors are as small as possible.</p><h3><strong>Conclusion</strong></h3><p>Machine learning may sound complex, but its foundation is simple.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/580/0*4tiQoXqze-owes6y.jpeg" /></figure><ul><li><strong>Linear algebra</strong> gives structure to data, helping models represent the world in numbers.</li><li><strong>Calculus</strong> teaches models how to learn and improve from their mistakes.</li><li><strong>Probability</strong> helps them deal with uncertainty and make informed guesses.</li><li><strong>Optimization</strong> ties everything together, guiding models toward better performance with every iteration.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa67758f8e34" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/machine-learning-math-minus-the-scary-symbols-aa67758f8e34">Machine Learning Math, Minus the Scary Symbols</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Riding the Solar Sweet Spot with PID in Simulink]]></title>
            <link>https://medium.com/techloop/riding-the-solar-sweet-spot-with-pid-in-simulink-a8fefd39ead9?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/a8fefd39ead9</guid>
            <category><![CDATA[solar-panels]]></category>
            <category><![CDATA[solar-power]]></category>
            <category><![CDATA[simulink]]></category>
            <category><![CDATA[pid-controller]]></category>
            <dc:creator><![CDATA[Kushagra Papnai]]></dc:creator>
            <pubDate>Wed, 24 Sep 2025 02:31:48 GMT</pubDate>
            <atom:updated>2025-09-24T02:31:47.763Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/848/1*efYhmP1MkufK_tmGkUHanw.png" /></figure><h3><strong>Introduction</strong></h3><p>Solar panels are great at turning sunlight into electricity — but only if you convince them to work at their sweet spot. That’s where MPPT comes in: think of it as a ‘GPS for solar panels,’ constantly steering them to the route that gives the most power.</p><p>Solar photovoltaic (PV) systems are highly nonlinear in nature, and their power output is significantly affected by environmental conditions such as solar irradiance and temperature. The PV system must run at its Maximum Power Point (MPP), which fluctuates dynamically with changing conditions, in order to achieve maximum efficiency. Maximum Power Point Tracking (MPPT) is the term for this procedure. A straightforward but reliable method for obtaining quick and steady tracking is the proportional-integral-derivative (PID) control approach.<br>The functioning of a MATLAB/Simulink model of a PV system controlled by a PID-based MPPT algorithm is examined in this blog, along with how this approach responds swiftly to changes and maintains stability even when sunlight abruptly increases or decreases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KDYlzMGZD6Opk2QWhZ3gYw.png" /></figure><h3><strong>Why choose PID?</strong></h3><p>PID stands for Proportional-Integral-Derivative, a widely used control technique in engineering and automation. It is designed to continuously monitor the difference between a point and the actual system output, known as the error, and adjust the control input to minimize this error over time. This makes PID controllers highly effective for systems that require precision and stability.</p><ul><li><strong>Proportional (P):</strong> The proportional component produces an output proportional to the current error. It helps us in finding the errors. A larger error results in a stronger corrective action, helping the system respond quickly to changes. In Maximum Power Point tracking(MPPT), the proportional part reacts to the difference between the current power output and the maximum possible power. If the solar panel is far from the Maximum Power Point (MPP), the controller makes a larger adjustment to quickly reach the Maximum Power Point(MPP).<br> — <strong>Integral (I):</strong> The integral component considers the accumulation of past errors. It gradually eliminates steady-state errors, ensuring that the system reaches and maintains the exact desired value. In Maximum Power Point Tracking(MPPT) the solar panel stays exactly at the Maximum Power Point(MPP), even if there are small fluctuations in irradiance or temperature by reducing errors.<br> — <strong>Derivative (D):</strong> The derivative component predicts the future behavior of the error by considering its rate of change. This helps in reducing errors overshoot, oscillations, and improves system stability. For Maximum Power Point Tracking(MPPT), this helps the controller predict and reduce overshoot or oscillations, so the system doesn’t constantly swing around the Maximum Power Point(MPP) but settles quickly at the optimal point.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dfUYNIT93XKoY2MFr8TKwg.png" /></figure><p>By combining a fast response to changes in irradiance, elimination of steady-state error, and stable operation under varying environmental conditions, the PID controller ensures that the solar panel consistently operates at its Maximum Power Point (MPP). The proportional component allows the system to react quickly to sudden changes in sunlight, the integral component corrects any persistent deviation from the MPP, and the derivative component helps prevent overshoot (when the system temporarily goes beyond the MPP) and reduces oscillations (when the power output keeps swinging around the MPP instead of stabilizing). Together, these features make PID a reliable and efficient choice for MPPT, enabling the solar PV system to extract the maximum possible energy even when weather conditions fluctuate rapidly.</p><h3><strong>System Model &amp; Simulation Setup</strong></h3><p><strong>Software Used- Simulink</strong></p><p><strong>Add-Ons Used- Simscape Electrical, Simscape</strong></p><p>Simscape is a physical modeling environment in MATLAB/Simulink that lets you simulate real-world physical systems (mechanical, electrical, hydraulic, thermal, etc.) using physical connections instead of just signal-based models. It’s key features includes models systems using physical networks (voltage, force, flow).It automatically applies conservation laws (energy, mass, charge).They also integrate easily with control systems from Simulink.</p><p>Simscape Electrical is a specialized Simscape library for modeling electrical and power systems, including both electrical circuits and power electronics. It models electrical components: resistors, capacitors, inductors, diodes, transformers, PV panels, batteries, etc. It includes specialized power systems: three-phase grids, inverters, converters, and renewable energy systems. It can simulate both DC and AC systems, including harmonic analysis and transient behavior. It also works with Simscape blocks for integrated multi-domain simulations (mechanical + electrical).</p><p><strong>Blocks Used:</strong></p><ol><li>Constant blocks: Block function containing some constant value. It act as input for our PV array.</li><li>PV array: The PV Array block is a prebuilt block that models a solar photovoltaic (PV) panel or an array of panels. It allows us to simulate the electrical behavior of solar panels under different conditions where it converts solar irradiance (sunlight) and temperature into DC electrical output.</li><li>Go-to: This block sends a signal to a label without connecting lines directly across the model. We assign a tag name and the signal can then be retrieved anywhere in the model using a <strong>From block</strong> with the same tag.</li><li>From-to: This block receives the signal from a <strong>Goto block </strong>with help of tag name.</li><li>MOSFET : A MOSFET block is used to model a Metal-Oxide-Semiconductor Field-Effect Transistor (MOSFET) which simulates a switching transistor that can turn on or off based on a gate control signal.</li><li>Unit-Delay : Unit Delay block is a very useful discrete-time block used to store the previous value of a signal</li><li>MATLAB function : The MATLAB Function block allows you to write MATLAB code and run it as part of your Simulink model. It’s useful when the algorithm is easier to express in code (like MPPT control) rather than building with Simulink blocks. We can write a compact MPPT algorithm in a MATLAB Function block for continuously adjust the operating voltage/current so the PV array delivers maximum power.</li><li>Display: The Display block is a simple but useful block used to show the value of a signal during simulation.</li><li>Scope: The Scope block is one of the most commonly used blocks for visualizing signals over time during simulation. It’s like an oscilloscope for your Simulink model. For MPPT we can see V-I curve from it.</li><li>Powergui: It manages the simulation of electrical circuits in Simscape Electrical. It handles circuit initialization, analysis, and solver settings. It calculates initial voltage, current, and flux in the system.</li><li>PID controller: The PID Controller block is used to implement a Proportional-Integral-Derivative (PID) control algorithm. The control signal is a combination of three components: <strong>Proportional (P):</strong> Reacts to the current error. <strong>Integral (I):</strong> Reacts to the accumulated error over time.<strong> Derivative (D):</strong> Reacts to the rate of change of the error.</li><li>Repeating Sequence: The Repeating Sequence block is used to generate a periodic signal that repeats over time. It’s very useful for testing, simulation of varying inputs, or driving controllers like MPPT in a PV system</li></ol><p>After making the necessary connections in Simulink the circuit diagram will look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rvneXyYBL2mLlT6MLTnn_w.png" /></figure><p>In the MATLAB function block following PID algorithm is being used to attain maximum efficiency .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/899/1*e18VZPuoFSijWq_s6SBSyA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/814/1*iSDuOwY4-PywPPvrG_XrwA.png" /></figure><p>The following parameters have been set for some specific blocks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O4Dhl1xqIKfa1r0UUdFwkw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mujRsoo14bUykBVsIyGYTQ.png" /></figure><h3><strong>Results</strong></h3><p>The proposed PID-based MPPT algorithm was simulated on a 33 kW solar PV array (15 parallel strings × 10 series-connected modules per string) with a boost converter and resistive load.</p><ul><li>PV Array Maximum Power: ~32.98 kW</li><li>The PID controller reached MPP within 0.15 s with negligible overshoot.</li><li>The steady-state PV voltage stabilized around 301 V (Vmpp) and current around 109 A (Impp).</li><li>Tracking efficiency was calculated at 99.3%, confirming minimal energy loss.</li></ul><p>The system achieved fast and smooth tracking of MPP for both constant and step-changing irradiance. Overshoot and oscillations were nearly eliminated, reducing power fluctuations and improving overall energy yield. The PID controller proved robust under partial shading and rapid irradiance changes, maintaining near-maximum energy extraction.</p><p>The following results were obtained from graph for V-I curve and I curve .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2dtqZ8kQDMERDy8LT7a74w.png" /></figure><p>The first graph is the result for I curve from Maximum Power Point Tracking(MPPT) in which the current(I) first increases with respect to time to a certain extent and then with help of PID algorithm it is being kept constant for certain process.</p><p>The second graph is the result for V-I curve from Maximum Power Point Tracking(MPPT) in which the current(I) increases with Voltage(V) and then is constant for the process.</p><h3><strong>Advantages &amp; Limitations</strong></h3><p>Let’s talk about the good, the great, and the slightly annoying about using PID for MPPT:</p><p><strong>Advantages:</strong></p><ol><li><strong>Fast Response: </strong>The PID controller quickly adjusts the system, allowing the system to reach the Maximum Power Point (MPP) within a short time after a change in irradiance.</li><li><strong>Stable Operation: </strong>Overshoot and oscillations around the MPP are minimized, ensuring smooth power output.</li><li><strong>Elimination of Steady-State Error: </strong>The integral action ensures the PV system remains exactly at the MPP even after small disturbances.</li><li><strong>Simplicity and Easy Implementation: </strong>PID control is easy to design and tune compared to advanced MPPT methods like fuzzy logic or neural networks.</li><li><strong>High Tracking Efficiency: </strong>The system achieves &gt;99% tracking efficiency, extracting nearly all available solar power.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VOg33zixk_3Rax6VWD4TWQ.png" /></figure><p><strong>Limitations-</strong></p><ol><li><strong>Tuning Sensitivity: </strong>Proper tuning of Kp, Ki, and Kd values is crucial. Incorrect tuning can lead to slow response, oscillations, or even instability.</li><li><strong>Derivative Noise Sensitivity: </strong>The derivative term can amplify measurement noise, which might cause small fluctuations in duty cycle.</li><li><strong>Not Adaptive: </strong>Unlike advanced techniques (e.g., Adaptive MPPT, Fuzzy logic), a basic PID does not adapt itself automatically to wide environmental changes — manual re-tuning might be required.</li><li><strong>Limited Performance in Rapid Changes:</strong> During very fast irradiance variations (e.g., passing clouds), PID may not react as quickly as more advanced adaptive or predictive MPPT methods.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gp0_oWgdEYcuvIbbNoeg7g.png" /></figure><h3><strong>Conclusions</strong></h3><p>This work demonstrated the design and simulation of a PID-based Maximum Power Point Tracking (MPPT) algorithm for a 33 kW solar PV system using MATLAB/Simulink. The results showed that the PID controller was able to reach the Maximum Power Point (MPP) quickly (within 0.15 s) with negligible overshoot and minimal oscillations, ensuring smooth and stable operation. The steady-state voltage and current stabilized at their MPP values, achieving a tracking efficiency of over 99%, which translates to nearly complete utilization of available solar power.</p><p>The proposed approach offers a good balance between simplicity, speed, and reliability. However, its performance depends heavily on proper tuning of the PID gains, and re-tuning may be necessary under different environmental conditions or system configurations. Despite these limitations, PID control remains a practical and effective solution for MPPT in PV systems, especially where cost-effectiveness and ease of implementation are key requirements.</p><p>So yes, PID isn’t flawless — it might need some babysitting with tuning — but when done right, it’s like having an efficient, reliable solar chauffeur driving your panels straight to the MPP every time.</p><h3><strong>References</strong></h3><ol><li><em>PID Control algorithm theory and applications </em>by Umid Mammadov -<a href="https://medium.com/@umidcybers/pid-control-algorithm-theory-and-applications-f249dc5a4519">https://medium.com/@umidcybers/pid-control-algorithm-theory-and-applications-f249dc5a4519</a></li><li><em>Investigation of PV inverter MPPT efficiency test platform</em> by IET- <a href="https://ieeexplore-ieee-org.egateway.vit.ac.in/document/7446670/authors#authors">https://ieeexplore-ieee-org.egateway.vit.ac.in/document/7446670/authors#authors</a></li><li>AI-Driven MPPT: A Paradigm Shift in Solar PV Systems for Achieving maximum Efficiency by IEEE — <a href="https://ieeexplore-ieee-org.egateway.vit.ac.in/document/10899184">https://ieeexplore-ieee-org.egateway.vit.ac.in/document/10899184</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8fefd39ead9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/riding-the-solar-sweet-spot-with-pid-in-simulink-a8fefd39ead9">Riding the Solar Sweet Spot with PID in Simulink</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I updated my resume on Google Drive without opening it- here’s how]]></title>
            <link>https://medium.com/techloop/i-updated-my-resume-on-google-drive-without-opening-it-heres-how-e36f36567201?source=rss----7166652df8c7---4</link>
            <guid isPermaLink="false">https://medium.com/p/e36f36567201</guid>
            <category><![CDATA[google-drive]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[resume]]></category>
            <dc:creator><![CDATA[Subham Agarwal]]></dc:creator>
            <pubDate>Wed, 17 Sep 2025 02:31:46 GMT</pubDate>
            <atom:updated>2025-09-17T02:31:45.636Z</atom:updated>
            <content:encoded><![CDATA[<p>As a developer trying to stay ahead in a competitive job market, I’ve had it with the repetitive grind. No, I’m not talking about tweaking my resume on Overleaf for the thousandth time (perks of being a perfectionist 👾). I’m talking about the soul-crushing ritual that follows: edit the PDF → switch to Google Drive → upload the new version → update the shareable link → triple-check the link hasn’t changed → repeat.</p><p>Then I stumbled upon a LinkedIn post: someone had written a Python script to check incoming emails and notify them if they got shortlisted. A lightbulb went off. Why not apply the same automated thinking to my resume workflow? Why not build a system that syncs my local files to Google Drive automatically, so I can focus on what actually matters — coding, building, and landing that dream job?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yuD944xEBmosqkxnuHifQQ.png" /></figure><p>In this article, I’ll walk you through how I built a resume sync service that:</p><ul><li>Monitors a local file for changes in real-time.</li><li>Instantly updates the file on Google Drive.</li><li>Maintains a single, <strong>permanent, shareable link</strong>.</li><li>Runs silently in the background as a Windows service.</li></ul><h3>How It Works: The Architecture</h3><p>The system is simple yet powerful. A Python script acts as the bridge between my local file system and Google Drive, using a file watcher to trigger updates automatically.</p><p>It relies on a few key libraries:</p><ul><li><strong>PyDrive2</strong>: A library that simplifies Google Drive API interactions. It handles OAuth 2.0 authentication, file uploads, and metadata management, so we don’t have to deal with raw API calls.</li><li><strong>Watchdog</strong>: This library monitors file system events in real-time. It’s far more efficient than constantly polling a directory to see if a file has changed.</li></ul><p>So, of course we have our Python script, but having to run it manually is just another chore. That’s where the magic of the .bat file comes in. Think of it as a one-click installer for our service 🛠️. You run it just once, and from then on, every time you log into your PC, our Python sync script automatically springs to life in the background, totally silent and invisible.</p><h3>Step 1: Taming the Google Drive API with OAuth 2.0</h3><p>First, we need to give our script secure access to Google Drive. The goal is an authentication flow that “just works” after a one-time setup.</p><h3>Getting Your Google Drive Credentials</h3><p>This is the most crucial part. Follow these steps exactly:</p><ul><li><strong>Create a Google Cloud Project</strong>: Go to the <a href="https://console.cloud.google.com/">Google Cloud Console</a> and create a new project. Name it something descriptive- I&#39;ve named this ResumeUpdation.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k-iGMZVZMX3fb8buh0wM2w.png" /></figure><ul><li><strong>Enable the Google Drive API</strong>: In your new project, navigate to <strong>APIs &amp; Services → Library</strong>. Search for “Google Drive API,” select it, and click <strong>Enable</strong>.</li><li><strong>Configure the OAuth Consent Screen</strong>: Before creating credentials, you need to tell Google about your app. Go to <strong>APIs &amp; Services → Audience</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*4hEkO3lCWst6IKXilmX-SQ.png" /></figure><ul><li>Choose <strong>External</strong> for the User Type.</li><li>In the <strong>Test users</strong> section, click <strong>+ ADD USERS</strong> and add the Google account email you’ll be using to access Google Drive. This is critical for avoiding “Access Denied” errors during development- as is clear below</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*myxQfskmDxhBvj3krh8Xvg.png" /></figure><p><strong>But if we add our email onto the test users</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*joEN8xEnozV5wO8BzVd7iw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*C-mZIbra8EK3yvvSqBY3uw.png" /></figure><p>A localhost file will load shortly after you click on “Continue” with the message- “Your authentication flow is accepted”.</p><ul><li><strong>Create OAuth 2.0 Credentials</strong>:</li><li>Go to <strong>APIs &amp; Services → Credentials</strong>.</li><li>Click <strong>+ CREATE CREDENTIALS</strong> and select <strong>OAuth client ID</strong>.</li><li>Choose <strong>Desktop app</strong> as the Application type.</li><li>Click <strong>Create</strong>. A pop-up will appear with your Client ID and Client Secret. Click <strong>DOWNLOAD JSON</strong>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*pNyZfJ4ndML3T8Vojx5s3g.png" /></figure><p><strong>The JSON file will look as follows:-</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/775/1*vOoKUXCF4Jzrnxc_tem7lA.png" /></figure><ul><li><strong>Set Up Your Project Folder</strong>:</li><li>Rename the downloaded file to client_secrets.json and place it in your project folder. This file contains the keys your script will use to identify itself to Google.</li></ul><h3>The Authentication Code</h3><p>Now, let’s write the Python code to handle the login process. This function will use the client_secrets.json file. On the first run, it will open a browser for you to log in.</p><p><strong>Remember earlier when I mentioned adding your email into the “Test Users” Section. If you followed all the instructions properly, the process will run smoothly, else do so.</strong></p><p>After that, it will store and reuse your credentials from a new file called credentials.json.</p><pre>import os<br>from pydrive2.auth import GoogleAuth<br>from pydrive2.drive import GoogleDrive<br>def authenticate():<br>    &quot;&quot;&quot;<br>    Handles Google authentication.<br>    - Prompts for browser-based login on the first run.<br>    - Saves credentials to credentials.json for future use.<br>    - Automatically refreshes expired tokens.<br>    &quot;&quot;&quot;<br>    gauth = GoogleAuth()<br>    # Try to load saved client credentials<br>    gauth.LoadCredentialsFile(&quot;credentials.json&quot;)<br>if gauth.credentials is None:<br>        # Authenticate if they&#39;re not there<br>        print(&quot; Opening browser for Google authentication...&quot;)<br>        gauth.LocalWebserverAuth() # Opens a browser tab for user to log in<br>    elif gauth.access_token_expired:<br>        # Refresh them if expired<br>        gauth.Refresh()<br>    else:<br>        # Initialize the saved credentials<br>        gauth.Authorize()<br># Save the current credentials to a file<br>    gauth.SaveCredentialsFile(&quot;credentials.json&quot;)<br>    print(&quot; Authentication successful!&quot;)<br>    return GoogleDrive(gauth)</pre><h3>Step 2: The First Upload &amp; Securing a Permanent Link</h3><p>The key to a permanent link is the <strong>Google Drive file ID</strong>. Once a file is created, its ID never changes. Our strategy is simple:</p><ul><li>Upload the resume <strong>once</strong> to get its unique file_id.</li><li>Save this file_id locally.</li><li>For all future updates, tell the API to update the content of the file with that specific ID.</li></ul><p>This initial setup script handles the first upload.</p><pre>LOCAL_FILE = &quot;Your_Resume_Name.pdf&quot; # IMPORTANT: Change this to your PDF&#39;s name<br>def initial_upload(drive):<br>    &quot;&quot;&quot;<br>    Uploads the resume for the first time, saves its file ID,<br>    and prints the permanent shareable link.<br>    &quot;&quot;&quot;<br>    print(f&quot; Uploading {LOCAL_FILE} to Google Drive...&quot;)<br># Create a new file object on Google Drive<br>    drive_file = drive.CreateFile({&#39;title&#39;: LOCAL_FILE})<br>    drive_file.SetContentFile(LOCAL_FILE)<br>    drive_file.Upload()<br># Make the file publicly readable<br>    drive_file.InsertPermission({&#39;type&#39;: &#39;anyone&#39;, &#39;value&#39;: &#39;anyone&#39;, &#39;role&#39;: &#39;reader&#39;})<br># Get the unique ID and the shareable link<br>    file_id = drive_file[&#39;id&#39;]<br>    link = drive_file[&#39;alternateLink&#39;]<br># Save the file ID locally for all future updates<br>    with open(&quot;file_id.txt&quot;, &quot;w&quot;) as f:<br>        f.write(file_id)<br>print(&quot; Upload successful!&quot;)<br>    print(f&quot; Your permanent resume link: {link}&quot;)<br>    return file_id<br># --- Main Setup Execution ---<br>if __name__ == &quot;__main__&quot;:<br>    if not os.path.exists(LOCAL_FILE):<br>        print(f&quot; Error: Resume file &#39;{LOCAL_FILE}&#39; not found.&quot;)<br>        print(&quot;Please place your resume PDF in this folder and run again.&quot;)<br>    elif not os.path.exists(&quot;client_secrets.json&quot;):<br>        print(&quot; Error: &#39;client_secrets.json&#39; not found.&quot;)<br>        print(&quot;Please follow Step 1 to download your credentials.&quot;)<br>    else:<br>        drive = authenticate()<br>        initial_upload(drive)<br>        print(&quot;\n Setup complete. You can now run the background service.&quot;)<br></pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*X04c721Ohc6u4SFbHGEgMw.png" /></figure><p>Run this script once. It will prompt you to log in via your browser, upload your PDF, and save a file_id.txt in your project folder. You now have your permanent link!</p><h3><strong>Step 3: Real-Time Syncing with a Background Service</strong></h3><p>Now for the magic. We’ll create a second script that runs silently in the background, watches for file changes, and performs the updates.</p><h3>Handling File Save Quirks</h3><p>When you save a file, editors often trigger multiple “modify” events. To prevent our script from uploading the same file three times in one second, we need to implement <strong>debouncing</strong> — a simple technique to ignore rapid-fire events.</p><p>Python</p><pre># From resume_service.py<br>import time<br>from datetime import datetime<br>from watchdog.observers import Observer<br>from watchdog.events import FileSystemEventHandler<br>class ResumeHandler(FileSystemEventHandler):<br>    def __init__(self, drive, file_id, logger):<br>        self.drive = drive<br>        self.file_id = file_id<br>        self.log = logger<br>        self.last_sync = 0 # Timestamp of the last sync<br>    def on_modified(self, event):<br>        # We only care about modification events for our specific file<br>        if not event.is_directory and event.src_path.endswith(LOCAL_FILE):<br>            current_time = time.time()<br>            # Debouncing: Ignore if a sync happened in the last 5 seconds<br>            if current_time - self.last_sync &lt; 5:<br>                return<br>            self.last_sync = current_time<br>            # Wait a moment to ensure the file write is fully complete<br>            time.sleep(1)<br>            self.sync_file()<br>    def sync_file(self):<br>        &quot;&quot;&quot;Updates the existing file on Google Drive using the saved file ID.&quot;&quot;&quot;<br>        try:<br>            self.log(f&quot; Change detected. Syncing {LOCAL_FILE}...&quot;)<br>            # Create a file object using the existing ID<br>            drive_file = self.drive.CreateFile({&#39;id&#39;: self.file_id})<br>            drive_file.SetContentFile(LOCAL_FILE)<br>            drive_file.Upload() # This updates the file content<br>            timestamp = datetime.now().strftime(&quot;%Y-%m-%d %H:%M:%S&quot;)<br>            self.log(f&quot; Resume synced successfully at {timestamp}&quot;)<br>        except Exception as e:<br>            self.log(f&quot; Sync failed: {e}&quot;)</pre><h3>Running as a Windows Service</h3><p>A script isn’t a true service until it runs automatically and invisibly. The easiest way to achieve this on Windows without admin rights is using the <strong>Startup folder</strong>. We’ll create a simple batch script to do the work.</p><p>install_startup.bat:</p><p>Code snippet</p><pre>@echo off<br>echo 🔧 Installing Resume Auto-Sync to Windows Startup...<br>set &quot;SCRIPT_DIR=%~dp0&quot;<br>set &quot;STARTUP_DIR=%APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup&quot;<br>set &quot;SHORTCUT_NAME=Resume Auto-Sync.bat&quot;<br>:: Create a batch file in the startup folder that runs our Python service script<br>(<br>    echo @echo off<br>    echo cd /d &quot;%SCRIPT_DIR%&quot;<br>    echo pythonw resume_service.py<br>) &gt; &quot;%STARTUP_DIR%\%SHORTCUT_NAME%&quot;<br>echo ✅ Installation complete. The service will start automatically on your next login.<br>pause</pre><p>The key here is pythonw.exe instead of python.exe. The w tells Windows to run the script without a console window, making it completely silent.</p><h3>Troubleshooting and Extending the System</h3><p>Here are quick solutions to common problems.</p><ul><li><strong>“403 Access Denied” or Authentication Errors</strong>:</li><li>Ensure you added your email as a <strong>test user</strong> in the OAuth consent screen. This is the most common cause.</li><li>Delete credentials.json and re-run the setup script to re-authenticate.</li><li><strong>File Not Syncing</strong>:</li><li>Check the sync.log file (created by the service) for error messages.</li><li>Make sure the LOCAL_FILE variable in your scripts exactly matches your PDF&#39;s filename.</li></ul><h4>This system is also easily extensible. You could modify the ResumeHandler to watch multiple files, or swap PyDrive2 with libraries for Dropbox or OneDrive to sync to different cloud providers.</h4><h3><strong>Real-World Usage Example</strong></h3><p>Here’s what a typical session looks like from my resume_sync.log:</p><p><strong>What happened:</strong> To test this I edited my resume in Overleaf, downloaded it at the same location, and within 4 seconds it was live on Google Drive at the same permanent link.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6cIcdywFBnA9ZoontX7X2A.png" /></figure><h3>Conclusion: Automation is Freedom</h3><p>This system has completely transformed how I manage my professional documents. No more manual uploads, no more broken links. I edit my resume, save it, and within seconds, it’s live on Google Drive at the same permanent link I’ve shared with recruiters and on my portfolio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*hNLcdYwu7DkaHSeDDzspNA.gif" /></figure><p>You can view the repo and test the code for yourself: <a href="https://github.com/subhamagarrwal/resume_uploader"><strong>GitHub Repository</strong></a></p><p>Feedbacks are as always, welcome!! You can dm me <a href="https://www.instagram.com/_subham.agarwal?igsh=MW1mcGZhZWZmYTl6ZA==">here</a> for any future references.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e36f36567201" width="1" height="1" alt=""><hr><p><a href="https://medium.com/techloop/i-updated-my-resume-on-google-drive-without-opening-it-heres-how-e36f36567201">I updated my resume on Google Drive without opening it- here’s how</a> was originally published in <a href="https://medium.com/techloop">Techloop</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>