<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Neeraj Sharma on Medium]]></title>
        <description><![CDATA[Stories by Neeraj Sharma on Medium]]></description>
        <link>https://medium.com/@neeraj716sharma?source=rss-f8625ab779a7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*inQPbsgAiedrJuQ79FAWUQ.jpeg</url>
            <title>Stories by Neeraj Sharma on Medium</title>
            <link>https://medium.com/@neeraj716sharma?source=rss-f8625ab779a7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 03:28:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@neeraj716sharma/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[So Apparently Songs Have Fingerprints Too]]></title>
            <link>https://medium.com/@neeraj716sharma/so-apparently-songs-have-fingerprints-too-6c957d404cb6?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c957d404cb6</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[math]]></category>
            <category><![CDATA[music]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Mon, 25 May 2026 16:57:20 GMT</pubDate>
            <atom:updated>2026-05-25T16:57:20.787Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/225/0*KJVbzTXCGi-AJ_s7" /><figcaption>Spectrogram</figcaption></figure><h3><strong>Inspiration</strong></h3><p>So my smartphone has this app called Now Playing which basically listens 24x7 to everything around me (Google speciality apparently 💀) and whenever a song plays nearby, it instantly tells me what song it is. Quite effective if you ask me.</p><p>And honestly we all hear random songs everywhere now, in reels, gyms, cafés and especially salons 🫡(the undisputed kings of unexpected bangers).</p><p>As a normal techie with borderline maniac levels of curiosity , I naturally started wondering, “How does this thing even know the song from just a few seconds of audio?”</p><p>That question became the starting point of this project.</p><h3><strong>The Theory</strong></h3><p>So I started exploring the original Shazam, (the actual song recognition engine ) , and quickly found out that it heavily relies on Fourier Transformation, one of the most hated and difficult topics for me back in college.😿</p><p>But this time it somehow felt intuitive and surprisingly fun to learn. Maybe because there was no exam pressure involved, or maybe because of those magical <a href="https://www.youtube.com/watch?v=spUNpyF58BY">3Blue1Brown </a>explanations that make even scary math feel beautiful.</p><p>So I started learning a bit about FT, IFT, sound waves, amplitudes, spectrograms and all that signal processing stuff which I would have probably ignored instantly if this was a classroom lecture.</p><p>And the more I studied how Shazam works, the more I started appreciating the simplicity and genius behind it.</p><p>Before this project I genuinely thought apps like these probably needed some massive AI model running on NASA level GPUs doing billions of calculations every second just to recognize one song playing in a noisy gym.</p><p>But the core idea is actually just math at its absolute best. Simple (well atleast for computers ), efficient and super elegant.</p><p>The basic idea is that every song gets converted into a unique audio fingerprint based on its frequency patterns. Then when you record a small audio clip, the system generates fingerprints for that clip too and tries to find matching patterns from the database. If enough patterns align with the same timing offsets voila song identified.</p><p>Yes its that simple but so so effective and almost fool proof.</p><h3><strong>The Implementaion</strong></h3><p>So without further ado, I jumped right in.</p><p>I decided to go module by module instead of pretending I understood the whole pipeline at once like some signal processing professor.</p><p>The first thing I started exploring was spectrograms. I took around 3–4 songs and started analyzing how they looked visually. And honestly, this was the point where things started becoming really cool.</p><p>Different regions in the spectrogram represented different instruments, drums had these sharp energetic spikes, guitars had cleaner patterns, vocals looked different altogether, and suddenly music started looking less like audio and more like some secret scientific heatmap.</p><p>Then, for research purposes (and definitely not to damage my own confidence), I recorded myself singing and generated a spectrogram for that too.</p><p>The results were concerning 🫠.</p><p>Unlike actual songs which had relatively stable and uniform frequency patterns, my singing looked like a stock market crash. Random ups, random downs, zero consistency and overall just visually sad. It looked as if spectrogram was also crying.</p><p>But weirdly enough, that comparison helped me understand the concept even better. Real songs tend to have identifiable and repeatable frequency structures, which is exactly what makes recognition possible.</p><p>After staring at spectrograms for unhealthy amounts of time, I moved to the next step: collecting strong frequency peaks.</p><p>Now obviously songs contain thousands upon thousands of frequencies at every moment, and storing all of them would be absolute overkill especially for my low end PC . The goal was not to remember the entire song, but to remember the important parts, kind of like how our brains remember the chorus of a song.</p><p>So instead of storing everything, I used a local maxima approach.</p><p>In simple words, for every small neighborhood in the spectrogram, I checked which points had the strongest energy. If a frequency peak was stronger than its nearby neighbors, it survived.</p><p>This step was surprisingly genius because it removed a massive amount of noise while keeping only the most dominant and unique parts of the song.And these surviving peaks became the foundation for generating audio fingerprints later on.</p><h3><strong>Hashing My Way Into Audio Recognition</strong></h3><p>At this point the peaks basically looked something like frequency points with their timing information attached to them. Instead of storing raw spectrogram data directly into the database, which would be massive and inefficient, the smarter approach was to convert these relationships into compact hashes.</p><p>I started pairing nearby peaks together and generating tiny signatures using their frequency positions and time differences. So a fingerprint would roughly contain information like:</p><p>frequency 1 → frequency 2 → time difference.</p><p>Then these triplets were hashed using SHA1, turning them into small unique identifiers that looked less like music data and more like some encrypted government records. (if they are encrypted highly doubtful. These hashes became the actual identity of a song.</p><p>And the really clever part is that even if the audio quality changes slightly, or there’s background noise (or you may sing yourself instead of playing recording of song) , many of these frequency relationships still survive. Which means the hashes still match.</p><p>After generating these fingerprints, I stored them in a local SQLite database along with their corresponding song name and time offsets. So now the database basically contained multiple songs and thousands of tiny audio fingerprints associated with them.</p><p>For the matching module, the process was almost the same. Give it a query audio clip, generate fingerprints from it, and then search the database for matching hashes.</p><p>(And yes, most code parts were written with AI because I believe in working smarter, not suffering unnecessarily.)</p><p>So for every fingerprint generated from the query audio, the system searched the database for identical hashes. Whenever matching hashes were found, the corresponding song names and offsets were collected.</p><p>If multiple hashes from the query aligned consistently with the same song at roughly the same timing offset, the confidence score for that song increased massively.( I know story is boring no masala but it is what it is , it is so powerfully simple that no hiccups or issues came up).</p><h3>Making It Look Slightly Less Horrible</h3><p>Finally, after the core logic was working, I attached a low-to-mid tier frontend using the holy trinity of HTML, CSS and JavaScript.( was working just fine with terminal but added anyhow )</p><p>Nothing fancy, no cinematic animations, just enough frontend so the project stopped looking like a terminal-only crime scene.</p><p>Then I converted the Python scripts into FastAPI endpoints so instead of manually running scripts, I could upload audio directly from the browser and let the backend handle fingerprint generation and matching.</p><p>And BOOM.</p><p>I had officially built my own tiny version of Now Playing/Shazam.</p><p>(For like 10 songs 🎶but still counts.)</p><p>I tested it multiple times with random clips, different timestamps and even low quality recordings, and surprisingly it worked with almost near 100% accuracy.</p><p>Although realistically speaking, with such a tiny dataset it was probably overfitting. Still, seeing it correctly identify songs from just a few seconds of audio felt ridiculously satisfying.</p><p>Of course there are a lot of areas where this can be improved.</p><p>Database storage can be optimized properly, fingerprint lookup can become faster, matching algorithms can be made more scalable, hashing strategies can become more robust and the entire pipeline can be made production ready for larger datasets.</p><p>But the beautiful thing is that the core idea remains the same.</p><p>Convert songs into unique fingerprints, store them efficiently, and later match patterns from tiny audio clips.</p><p>Simple, elegant and genuinely clever.</p><p>(Attaching some ugly screenshots for refrences )</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r6Aj7gSy_tvsJrbv4lYzbA.png" /><figcaption>Add song ui 😔</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/752/1*SMJ_WJON757K5I11z_gccg.png" /><figcaption>Match Song UI</figcaption></figure><p>Thanks to Python and all the amazing supporting libraries that made signal processing feel way less scary than it actually is, ChatGPT for helping me understand concepts whenever my brain started buffering, and all the other articles, videos and resources that quietly carried this project in the background.</p><p>Feel free to reach out if anyone’s interested in discussing the project further, although realistically it sometimes feels like I’m talking to myself because probably only 2–3 people read beyond this point while the rest skim through the article like it’s Terms and Conditions 👓</p><p>( Low attention span generation, can’t even blame them honestly )</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c957d404cb6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Most Overengineered Way to Track your health]]></title>
            <link>https://medium.com/@neeraj716sharma/the-most-overengineered-way-to-track-your-health-d9dc8069654a?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9dc8069654a</guid>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[telegram]]></category>
            <category><![CDATA[n8n-workflow]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Thu, 21 May 2026 18:57:28 GMT</pubDate>
            <atom:updated>2026-05-21T18:57:28.958Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Inspiration</strong></h3><p>I have a smartwatch that tracks almost everything my sleep, heart rate, SpO₂, stress levels, steps and ships all of that data somewhere far away ( China ofc ) to probably help train their llm models.</p><p>So I thought: <em>why not use my own data for my own benefit instead?</em></p><p>And since I was already deep into the n8n rabbit hole, combining automation with wearable data felt too tempting to ignore. Yeah i am literally thinking every way possible to use this n8n 🧠.</p><p>One more thing I didn’t want to send ultra-personal health data to external AI APIs. Sure, cloud models are better analyzers, but they also end up knowing way too much about your life. So I decided to build a completely local pipeline instead.</p><p>No grand architecture. No proper planning.</p><p>Just curiosity, music, and “let’s see if this works. (F around and Find Out) ”</p><h3><strong>How do i process the DATA</strong></h3><p>The major issue was that my smartwatch lives inside a closed ecosystem (Hyper OS you can guess the brand now) . It happily tracks my sleep, heart rate, stress, oxygen levels and everything else but exporting that data? Nope. Not happening.</p><p>So I had to find a workaround.</p><p>After some research, I found <a href="https://play.google.com/store/apps/details?id=com.google.android.apps.healthdata&amp;utm_source=chatgpt.com">Health Connect by Google</a>, which is surprisingly more open when it comes to data sharing. (Yeah Google and data sharing 🙀 shocker for me too)<br> So the flow became:</p><p><strong>Watch → Health Connect</strong></p><p>But there was still one problem.<br> Health Connect itself doesn’t provide an easy automatic export system. 🫠</p><p>Which led me to another hidden gem of the internet — an app with barely ~1k downloads:</p><p><a href="https://play.google.com/store/apps/details?id=com.teqxnology.healthdataexport&amp;hl=en_IN&amp;utm_source=chatgpt.com">Health Data Export</a> ( unpaid Shout out )</p><p>And honestly, this app solved the biggest headache of the project.</p><p>It basically pulls data from Health Connect and pushes it directly into Google Sheets. Suddenly I had my sleep data, workouts, calories, heart rate, steps everything sitting neatly inside spreadsheets.</p><p>So the actual pipeline became:</p><p><strong>Watch → Health Connect → Health Data Export → Google Sheets</strong></p><p>Funny enough, i was building the pipeline before n8n even entered the picture. Now I finally had data flowing into sheets but I still had absolutely no idea how I was going to make with this data in sheets.</p><h3>What to do with DATA now?</h3><p>With data filling in the Excel sheet, I still had absolutely no idea what to build with it. 😕</p><p>Then I remembered those Marvel movie scenes where JARVIS keeps warning Tony in real time:</p><p><em>“Suit overheating.”<br> “Critical damage.”<br> “Power levels dropping.” <br> “Captain America on your left.”</em></p><p>And I thought wait, why can’t I have that too .</p><p>So I started exploring n8n nodes.<br> First thing I needed was a Google Sheets trigger node.<br> (Also, setting up Google Cloud deserves its own rant someday. That UI tested my patience, my soul, and probably my blood pressure. Google made sure they get me even with Health Connect help)</p><p>After somehow surviving Google Cloud setup, I finally got the sheet triggers working.</p><p>Now every new row added became an event:</p><ul><li>Heart rate polling every minute</li><li>Sleep/activity updates every hour</li><li>Workout logs getting pushed automatically</li></ul><p>Polling sheets and asking the question</p><p><em>new row added? new row added? new row added? every min or hr as per config.</em></p><p>Then I added alert logic on top of it:</p><ul><li>High heart rate detected → “sit down brother”</li><li>Low activity this hour → “go walk”</li><li>Critically low sleep detected →<br> my personal speciality ( I can do this all day everyday 💘)</li></ul><p>Now the question was:<br> how do I actually receive these alerts on my phone/watch?</p><p>And then entered my old friend</p><h4>The Holy Almighty Telegram 🛐</h4><p>I’ve made way too many Telegram bots before, so this part felt like home ❤</p><p>Easy setup, clean UI, great APIs, and with the blessings of the legendary BotFather 👐, I quickly wired everything up.</p><p>Now the pipeline could send almost real-time alerts directly to my phone. And surprisingly everything worked smoothly no hiccups no mess ups no issues. 🤞🏻</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*GnF1UklcvsCiqlgzL37mtQ.png" /><figcaption>Legs Day is always painful 😀</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/493/1*B5QSsZiOMXbmZUpjollZrw.png" /><figcaption>Bro was expecting me to finish 10k steps at 8 am itself</figcaption></figure><h3><strong>Whats more to Add ?</strong></h3><p>With the alert system working fine i was getting impatient seeing so much data at my disposal. No way i was letting all this data go to waste. So i thought why not introduce AI into the loop too. Let him also enjoy what he likes the most DATA. ( Lots of personal training data )</p><p>So i made another pipeline inside the project which gathers the complete weekly data, cleans it and sends it to an LLM to analyse my life (which sounds sad af honestly).</p><p>Now my CPU somehow survived the last experiment so as any good parent would do I decided to give him more pain 💓. I downloaded a 9GB LLM model this time , ran it and boom instant failure 💀</p><p>Turns out my 16GB RAM was already split between Ollama and n8n (8GB each) and smart me tried loading a 9GB model into 8GB RAM. Thankfully the hardware survived another day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mx2f0NszEmRddib5xu497A.jpeg" /><figcaption>🫡</figcaption></figure><p>So i switched to somewhat smaller models and tried qwen2.5:3b-instruct. (upgrade from last time ) The plan was simple give weekly data, ask AI to analyse it and send insights on Telegram every week.</p><p>But the model hallucinated like crazy initially. It invented dates, made up workout stats and basically generated fan fiction about my health 😭</p><p>After tweaking prompts many times, reducing temperature and restructuring the data properly, it finally started giving decent analysis. Not perfect, but honestly pretty impressive for a tiny local model running on shared RAM (inside container ) and owner’s stubbornness.</p><p>And obviously once it started workin I added another Telegram trigger for weekly AI reports too</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/577/1*2HOA5HG28I7HqvRohrAjog.png" /><figcaption>Bros asking me to do cardio 😆</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*3P32EciJ1eFCOnSsWPo68g.png" /><figcaption>ignore the dates llm is already doing overtime</figcaption></figure><h3><strong>Hmm Still DATA was there 🤣</strong></h3><p>So this was already a perfect point to stop the project gracefully but there was still time-series data lying around looking at me and my inner “data scientist who watched 3 dashboard reels on Instagram” woke up.</p><p>So naturally I made even more unnecessary nodes to convert raw data into visual graphs, because apparently human civilisation peaked at turning numbers into colorful lines and calling it “insights”. Whats life without turning raw data into meaningful action items.</p><p>So yeah :) here are some beautiful graphs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/632/1*RyOX3WJuiutgkO1WjdFpyA.png" /><figcaption>Sleep</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*31QK5ciosRvdU-Q9vxiWqw.png" /><figcaption>Steps</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/507/0*91Ony-LMD3zHp25o.png" /><figcaption><a href="https://thedatascore.substack.com/p/the-roi-of-data-revisited">https://thedatascore.substack.com/p/the-roi-of-data-revisited</a></figcaption></figure><p>As always i would like to thank the tech tools and community n8n , health connect , health data export , ollama , telegram , and quickchart.</p><p>Feel free to contact me for pipeline schema , code etc.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9dc8069654a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My Laptop Plans My Trips Now]]></title>
            <link>https://medium.com/@neeraj716sharma/my-laptop-plans-my-trips-now-251db4901599?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/251db4901599</guid>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[reddit]]></category>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[n8n-workflow]]></category>
            <category><![CDATA[docker]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Thu, 14 May 2026 17:26:11 GMT</pubDate>
            <atom:updated>2026-05-14T17:26:11.173Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The WHY ?</strong></h3><p>Me and my friend spend an unhealthy amount of time planning trips (we rarely execute any anyway :) ).</p><p>Every single time, it starts the same way googling destinations, opening random sponsored blogs, reading paid reviews pretending to be “hidden gems”, and somehow ending up with 20 tabs open and zero actual plans.</p><p>And the ads ; don’t even get me started 😈</p><p>Search for one destination just once, and suddenly YouTube starts recommending travel vloggers, Instagram becomes full of it, and every other website wants to sell you some “exclusive” package. Google does its magic so well that it genuinely feels like even your neighbours know where you’re planning to go. Privacy on the internet honestly feels like a joke at this point.</p><p>To avoid all these biased and sponsored blogs, we usually end up on Reddit because the discussions there feel far more real and honest. Real people sharing real experiences, not SEO-optimized articles calling every place a “must visit hidden gem”.</p><p>But even then, the same privacy issue still exists. So I thought, why not build something local that does the job for us?</p><h3><strong>The WHAT?</strong></h3><p>So the next obvious question was: what exactly should I build?</p><p>The goals were pretty simple. I wanted everything to be as local as possible and CPU-only. Mainly because I’m broke. Buying a GPU was not happening, using my company laptop’s GPU was definitely not an option, and I absolutely did not want to spend my life fighting with <a href="https://colab.google/?utm_source=chatgpt.com">Google Colab</a> sessions. That thing feels less like a development environment and more like a patience test. 🫠 ( i add these emoji manually these are not ai generated )</p><p>So the plan became clear: keep it local, keep it lightweight, and keep Reddit at the center of everything.</p><p>That’s when I decided to try <a href="https://n8n.io/?utm_source=chatgpt.com">n8n</a>. I had never used it before I’d only heard about it from one of my tech enthusiast friend who somehow discovers every new tool before the rest of humanity and talks about it as if he invested in their seeding round.</p><p>And since I already established my financial situation, there was only one deployment strategy possible: <strong>Docker ⚓</strong>. The poor man’s VM.</p><p>I pulled the Docker image for n8n, ran it, and boom it just worked.</p><p>Suddenly I was thrown into this entirely new world of automation workflows. I started experimenting with random nodes, triggers, conditions, code blocks, and whatever else I could click without breaking my laptop. Honestly, it was way more fun than I expected.</p><p>But anyway, back to the actual problem.</p><p>Now that the automation pipeline was in place, I got my hands on the Reddit API and thought, “Alright, this should be simple now.”</p><p>Well…</p><p>It wasn’t.</p><p>Let me break down the problems I ran into in the next section.</p><h3><strong>The ISSUES?</strong></h3><p>With the Reddit API finally in my hands, I genuinely felt like Thanos collecting the Infinity Gauntlet. 💪🏻</p><p>So I built the first version of the automation. The workflow was simple: enter a destination, send it to the Reddit API, fetch posts, and let the magic happen. ( HEHE inncoent me thought it would work )</p><p>And boom… I got a massive amount of data.</p><p>The only problem? Most of it was absolute garbage.</p><p>Random posts with zero upvotes, people asking for honeymoon suggestions, “fun trip ideas with friends”, completely unrelated discussions, and occasionally someone fighting in the comments for no reason at all. The actually useful travel insights were probably just 5% of the total data.</p><p>But of course, as an overconfident developer, I thought:<br> “Who cares? I’ll just give everything to AI. Let that guy figure it out.” 😎</p><p>Now since this was a low-budget setup, I needed a small local model that could somehow survive on CPU-only hardware. After some searching, I found Phi-3 Mini and naturally ran that in Docker too (inside ollama ofc).</p><p>Then, like a completely innocent child with no understanding of consequences, I dumped all the Reddit data into the model with a “good prompt” and waited for intelligence to emerge.</p><p>Instead, my CPU entered survival mode.</p><p>The laptop fans started spinning like they were preparing for takeoff, CPU usage looked illegal 600% , and the entire system sounded like it was fighting for its life. I had to kill the container before my laptop achieved spiritual enlightenment. 🛐</p><p>At one point I literally kept the laptop upside down under the AC for five minutes just to cool it down.</p><p>But honestly, that disaster taught me an important lesson: all those boring lectures about data cleaning and preprocessing in college were actually useful.</p><p>So I went back, added more preprocessing nodes in <a href="https://n8n.io/?utm_source=chatgpt.com">n8n</a>, filtered out junk posts, cleaned the data properly, and only then sent it to the local LLM instead of trying to nuke my CPU with raw Reddit chaos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7c8EOOcgzFFN1MGudV4vJg.png" /><figcaption>the CPU abuse</figcaption></figure><h3><strong>The HOW?</strong></h3><p>After multiple failed attempts, broken workflows, questionable prompts, and enough CPU torture to qualify as hardware abuse, I finally started understanding what actually worked.</p><p>The first few versions of the pipeline were honestly terrible. The summaries were bad, irrelevant posts kept slipping through, useful information got buried under random noise, and sometimes the model would confidently summarize complete nonsense (and dont ask abt hallucinations ) . I kept tweaking things again and again until I eventually stopped pretending the “one big AI call” approach was going to magically solve everything.</p><p>So instead of throwing raw Reddit chaos directly into the LLM, I broke the problem into smaller stages.</p><p>The final workflow ended up looking something like this:</p><p>A webhook first accepts the travel destination as input. Instead of directly searching just “Goa”, I added a query generation step that creates multiple travel-focused searches like “Goa itinerary”, “Goa best places”, “Goa solo travel”, “Goa scams”, “Goa budget trip”, and similar variations. This alone massively improved the quality of Reddit results because people discuss travel differently depending on context.</p><p>These generated queries are then passed to the Reddit API to fetch relevant posts and discussions.</p><p>Now comes the important part: filtering. 🧹🧹</p><p>Not every Reddit post deserves AI attention. Some posts had zero engagement, some were irrelevant, some were memes pretending to be travel advice, and some were just two people arguing in comments for absolutely no reason. So I added filtering nodes, limits on the number of posts, and basic preprocessing before anything reached the model.</p><p>Then I introduced another AI layer using Qwen. Instead of directly summarizing everything, I first used the model to rate the usefulness of each post. Posts with low relevance scores were rejected, while genuinely useful travel experiences were retained.</p><p>Only after all this filtering, scoring, cleaning, and aggregation did the final summarization step happen.</p><p>And honestly… the difference was huge. 🤯</p><p>The final summaries started feeling less like random internet noise and more like an actual traveler sharing consolidated experiences from dozens of Reddit threads. It could highlight common scams, underrated places, realistic budgets, transport tips, solo travel advice, and things people usually only learn after making mistakes themselves.</p><p>The funniest part is that when you look at the final workflow in <a href="https://n8n.io/?utm_source=chatgpt.com">n</a>8n, it looks ridiculously complicated for something that started with:<br> “Bro let’s just summarize Reddit posts.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sr10b5pcq2dNHwuffooGWA.png" /><figcaption>I know , no one will click and see this in detail</figcaption></figure><h3><strong>Was it Worth it?</strong></h3><p>Well, the thing I built is still far from production-level, but considering the adversity of running everything locally on a CPU, I’d say I ended up building a pretty decent little system.</p><p>More than anything though, it was genuinely fun. 💘</p><p>I learned a lot, experimented with a bunch of tools, broke things multiple times, overheated my laptop at least once, and somehow still ended up with something useful at the end of it. And honestly, it was still a far better use of time than doomscrolling for hours.</p><p>So yes, it was absolutely worth every second.</p><p>This pipeline will probably keep evolving as I learn more and continue improving it, but for a first iteration, I’m pretty happy with where it landed.</p><p>Big thanks to <a href="https://n8n.io/?utm_source=chatgpt.com">n8n</a>, <a href="https://ollama.com/?utm_source=chatgpt.com">Ollama</a>, the Reddit API, and of course my CPU for holding on to dear life throughout this entire experiment.</p><p>And if anyone actually wants the code, prompts, or workflow resources… feel free to reach out.</p><p>(I know nobody will, but still.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/883/1*3WtjOYFlnXHHWXF2DNtxIQ.png" /><figcaption>Final Result which i am proud of (dont look for mistakes in this)</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=251db4901599" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Quiet Math Behind Authenticator Apps]]></title>
            <link>https://medium.com/@neeraj716sharma/the-quiet-math-behind-authenticator-apps-b38c9a266b51?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/b38c9a266b51</guid>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Sat, 03 Jan 2026 11:12:30 GMT</pubDate>
            <atom:updated>2026-01-03T11:12:30.404Z</atom:updated>
            <content:encoded><![CDATA[<h3>A Small Moment During a Routine Login</h3><p>On a very cold Saturday morning, I was doing what I’ve done countless times before, trying to connect to my company VPN. Nothing unusual. Open the client, approve the request, enter the passcode from Duo (yes, I still use passcodes). I’ve done this so many times that it’s almost muscle memory.</p><p>But this time, something felt off. My phone’s internet was flaky. Mobile data hadn’t connected yet. Still, without really thinking about it, I opened Duo anyway, out of habit more than confidence.</p><p>And the code was already there. Counting down. Calm. Unbothered.</p><p>That’s when it clicked. This app clearly wasn’t asking anyone for permission. It wasn’t reaching out to some remote Duo server to fetch a passcode. Yet it knew exactly what number to show.</p><p>And that made me pause.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*b1JLjfAp8twoF1IX.jpg" /></figure><h3>Curiosity Took Over</h3><p>I didn’t connect to the VPN right away. Instead, I did what any curious engineer would do. I went straight to reading how this thing actually works.</p><p>I was expecting something complex, background syncs, encrypted API calls, maybe some clever offline fallback logic.</p><p>What I found was surprisingly simple. 😱 Almost humbling, actually.</p><p>There was no hidden network trick. No last-known-code cache. Just a well-thought-out idea that has been around for years, quietly doing its job.</p><h3>So What’s Actually Happening Under the Hood</h3><p>Once you look past the UI and the countdown timer, authenticator apps like Duo are doing something very deliberate and very minimal.</p><p>They follow the <strong>Time-based One-Time Password (TOTP) standard (RFC 6238)</strong>, which is built around one idea. Both sides should be able to compute the same answer without ever talking to each other again ( basically the healthiest long-distance relationship in security )</p><h3>It All Starts With a One-Time Trust Exchange</h3><p>During setup the first time, the QR code you scan isn’t just a convenience step, it is the foundation of the entire system.</p><p>Encoded inside it is a randomly generated <strong>secret key</strong>. That key is stored in two places:</p><ul><li>On Duo’s server</li><li>On your phone</li></ul><p>Once this exchange happens, the setup phase is over. The secret never needs to be sent again. From this point on, both sides are expected to independently generate the same result using the same inputs.</p><p>This is why reinstalling the app or switching phones forces you to set it up again. The secret is intentionally tied to the device.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*G3Rj8ktVfyAwaSm4" /></figure><h3>Where That Secret Lives on Your Phone</h3><p>Curiosity made me want to peek at the secret ( atleast in my phone since duo won’t allow me to sneek in their server) 🥷🏻. But it isn’t saved as a file or preference that an app or a user can casually read.</p><p>Authenticator apps rely on <strong>OS-level secure storage</strong>:</p><ul><li>Android uses the <strong>Android Keystore</strong>, often backed by hardware (TEE or Secure Element)</li><li>iOS uses the <strong>Keychain</strong>, protected by the <strong>Secure Enclave</strong></li></ul><p>In both cases, the secret is:</p><ul><li>encrypted at rest,</li><li>bound to the device hardware, and</li><li>inaccessible to other apps</li></ul><p>Even the authenticator app itself cannot freely extract it. The operating system controls access and enforces strict boundaries.</p><p><strong>Security note:</strong> If someone compromises your device, the secret can be exposed. That is why TOTP apps are tied to hardware and require device-level authentication.</p><h3>Time Is All You Need</h3><p>Once the secret is stored, the authenticator app does not need much else. The only variable input is <strong>time</strong>.</p><p>Initially, I assumed time zones would make this messy. My phone is in IST, the server could be anywhere in the world (us-east-1 🫠). Surely that would cause problems. But TOTP sidesteps this by using <strong>Unix time</strong>, the number of seconds elapsed since January 1, 1970 (UTC). Time zones are just a display concern.</p><p>The Unix timestamp is divided into fixed 30-second windows. Each window maps to exactly one passcode. As long as both sides land in the same window, they will generate the same result.</p><p>If a window is missed, nothing breaks. Both sides simply move forward to the next one. Servers usually allow a small buffer to handle minor clock drift, keeping the system stable even if your phone’s clock is not perfectly synced.</p><p>Time, in this system, replaces coordination. Instead of talking to each other, both sides independently agree on when they are, and that turns out to be enough.</p><h3>Turning Time Into a Passcode</h3><p>Once time is split into fixed windows, the app combines the current window with the shared secret key and feeds them into an <strong>HMAC function</strong> (usually SHA-1).</p><p>The output is a binary hash, which is too long and messy for humans. The algorithm then performs <strong>dynamic truncation</strong>, extracts a 31-bit value, and reduces it modulo 10^6 to produce the familiar six-digit passcode.</p><p>This process is fully deterministic. Same secret, same time window, same code. The server performs the exact same calculation independently. No network, no guessing, just precise math.</p><p>Six digits may not look impressive, but behind them is a carefully constrained pipeline designed to be predictable, synchronized, and resistant to replay attacks, all without requiring a single network call from your phone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lm0CU12UCTR8LKaQ33CnbQ.png" /></figure><h3>Why Offline Generation Is a Feature, Not a Side Effect</h3><p>Once the secret is stored, the app can generate passcodes completely offline. No server calls, no syncing, no internet required.</p><p>This is not a lucky coincidence. In the ever evolving internet where even checking something small requires an internet connection ,offline generation reduces network dependency, minimizes attack surfaces, and ensures the system works even in flight mode or poor coverage. Fewer moving parts mean fewer things that can break or be exploited.</p><h3>Bonus: How Push Notifications Differ</h3><p>( For most NPC users who still use push notifications )</p><p>Unlike TOTP codes, push-based MFA actually needs the server. When you try to log in, the server sends a notification to your device asking, “Is this you?” Your phone approves or denies in real time, and the server waits for the response.</p><p>It is more convenient, but also more dependent on connectivity and server availability. TOTP, by contrast, works quietly offline, old-school math and synchronized time doing all the heavy lifting without calling anyone. ( my kind of guy )</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b38c9a266b51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Designing a Simple Rate Limiter]]></title>
            <link>https://medium.com/@neeraj716sharma/designing-a-simple-rate-limiter-7a1a72bfe3d1?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a1a72bfe3d1</guid>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Sun, 21 Dec 2025 15:15:49 GMT</pubDate>
            <atom:updated>2025-12-21T15:15:49.081Z</atom:updated>
            <content:encoded><![CDATA[<p>This is a quick write-up of how I built a simple rate limiter over the weekend.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8MM3HshwJDZ9h6Sb.png" /><figcaption>429 Error Code</figcaption></figure><h3>Why Rate Limiting Matters 🤔</h3><p>Most backend systems don’t break because of complex bugs, they break because someone sends too many requests. Imagine launching a simple API over the weekend and waking up on Monday to find your server struggling because one client kept hitting the same endpoint in a loop. This is where rate limiting quietly saves the day.</p><p>We’ve all seen it in real life: login pages that lock you out after a few wrong password attempts, or OTP services that don’t let you request a new code every second. That’s rate limiting in action. Interestingly, while writing backend services in Node.js and Python, I never really experienced this problem firsthand, mostly because my local apps were never exposed to real traffic (just 3–4 users max 🙃).</p><p>The reality hit when I started scraping data from public APIs like GitHub, Twitter, or Amazon. After a few rapid requests, my calls were suddenly blocked. That’s when it clicked — these systems have to protect themselves. Without rate limiting, a single user or bot could easily overwhelm an API and degrade the experience for everyone else.</p><p>At its core, rate limiting isn’t about saying “no” to users. It’s about making sure the system stays responsive, fair, and alive for everyone.</p><h3>Understanding the Problem &amp; Design Decisions 📚</h3><p>Before writing any code, I first turned on some music <a href="https://youtu.be/MlXTbqptX_8?si=5CuBjuQ5Mu5-Fxb5">🎧</a>, then I clarified what kind of rate limiter I actually wanted to build. This would be a server-side API rate limiter, since client-side limits are easy to bypass and don’t really protect the backend.</p><p>I wanted the limiter to be flexible, capable of throttling requests based on different properties like IP address or user ID, depending on the endpoint. I studied different types of properties — how we can allow some IPs (admin) to bypass this system altogether or block some IPs (those who try doing continuous DDoS). Since it was a weekend project, I decided to go with user ID–based limiting. This was based on user ID, which is expected to be passed with each API request as a token or proof that you are an authenticated user.</p><p>After this came the part where I needed to study different available algorithms for the same. This was a down-memory-lane walk since we studied this briefly in our Computer Networking course in college. Some of the algorithms which are often used are:</p><p><strong>Token Bucket</strong><br> Tokens are added to a bucket at a fixed rate. Each request consumes a token. If the bucket is empty, requests are rejected. This approach allows short bursts while still enforcing an average rate.</p><p><strong>Leaky Bucket</strong><br> Requests are processed at a constant rate, like water leaking from a bucket. Excess requests are dropped when the bucket overflows, making traffic smooth but sometimes too strict for bursty use cases.</p><p><strong>Fixed Window Counter</strong><br> Requests are counted within a fixed time window (e.g., 100 requests per minute). It’s simple to implement but suffers from boundary issues, where users can send a burst of requests at window edges.</p><p><strong>Sliding Window Log</strong><br> Stores timestamps for each request and counts only those within the last time window. This is very accurate but can be memory-intensive at high traffic.</p><p><strong>Sliding Window Counter</strong><br> A hybrid approach that approximates the sliding window using counters from adjacent windows. It offers better accuracy than fixed windows while being more efficient than storing every request timestamp.</p><p>(For reading in more depth about each algorithm, I recommend <a href="https://bytebytego.com/courses/system-design-interview/design-a-rate-limiter">🔗</a>)</p><h3>My Choice of Algo and Tech Stack 🤓</h3><p>I chose the sliding window counter approach because it offers a good balance between accuracy and efficiency. Unlike fixed windows, it avoids burst traffic at window boundaries, and unlike sliding window logs, it doesn’t require storing every request timestamp. This makes it practical for high-throughput systems. And this was the first time using a sliding window outside of LeetCode questions 🤩.</p><p>For the backend, I used TypeScript, mainly for better type safety and cleaner APIs as the codebase grows (I don’t know Go or Rust as of now). For the backend, a simple Express server was a good option. To store rate-limiting state, I used Redis as an in-memory data store. Rate limiting is extremely latency-sensitive, and Redis provides fast atomic operations and built-in expiry, which are ideal for counters.</p><p>Why not a simple SQL or NoSQL database? That was what I thought. Using a traditional database like PostgreSQL would add unnecessary latency and write load for something that is highly volatile and short-lived. Rate-limit data doesn’t need long-term persistence — it just needs to be fast and accurate — which makes Redis a much better fit. Redis operates in memory (RAM) with sub-millisecond latency, while PostgreSQL queries typically take several milliseconds, making Redis far better suited for latency-sensitive, high-frequency operations like rate limiting.</p><h3>Implementation: What Actually Happened 🪛</h3><p>The implementation started simple, fun, vibing on music — and then reality kicked in.</p><p>I used a free Redis Cloud instance (30 MB) for storage, which sounds generous until you actually start debugging rate-limit keys and expirations. Memory constraints quickly force you to be intentional, which, in hindsight, was a good learning experience. The backend itself was built using Node.js with Express, keeping things lightweight and easy to reason about.</p><p>Then came the first unexpected boss fight. Since I was working on a company laptop, Redis just refused to connect. After nearly three hours of debugging, I finally discovered the culprit: the company VPN was silently blocking Redis’s port (15468). Apparently, unusually large port numbers don’t get along well with corporate networks. I learned this the hard way, tweaked a few things (details best left out here 😄), and moved on.</p><p>With connectivity sorted, I built a working prototype — only to run straight into a race condition. This is where my OS classes finally justified their existence. Multiple requests were incrementing counters at the same time, leading to incorrect limits. After digging through Reddit threads and leaning on AI-assisted debugging (and almost quitting 🫠), I found the solution: Redis Lua scripts 🙏🏻. Executing the logic atomically inside Redis solved the issue cleanly.</p><p>Of course, the bugs didn’t stop there. A few more small issues surfaced, as they always do, but eventually the limiter was ready to be tested. That introduced another problem: I can’t physically make 10 requests per second (yes, I tried it). Thankfully, Postman’s Collection Runner came to the rescue and helped simulate realistic traffic. I then fixed a few TTL errors in Redis to avoid the 30 MB bottleneck.</p><p>In the end, I had a working sliding window counter rate limiter. It’s far from perfect and still has plenty of open ends — things like location-based filtering, more flexible rule definitions, and better observability. But for a weekend project, it did what it was supposed to do: teach me how real systems behave once they leave the comfort of localhost.</p><h3>Conclusion 👨🏻‍⚖️</h3><p>The next time I see a <strong>429: Too Many Requests</strong> error, I’ll know exactly what’s happening behind the scenes. Building this rate limiter made system design concepts feel real (pain), especially the parts you don’t learn from diagrams, like race conditions, network issues, and testing challenges.</p><p>If you want to go deeper, use an AI agent to brainstorm, but don’t follow it blindly. Keep asking <em>how</em> and <em>why</em> at every step. That’s where real understanding comes from.</p><p>The solution isn’t perfect and has many open ends, but for a weekend project, the learning alone made it worth it.</p><p><strong>Recommended Reading → </strong><a href="https://bytebytego.com/courses/system-design-interview/design-a-rate-limiter"><strong>🔗</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7a1a72bfe3d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Backend-as-Service]]></title>
            <link>https://medium.com/@neeraj716sharma/backend-as-service-dce33b4a9d07?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/dce33b4a9d07</guid>
            <category><![CDATA[backend-as-a-service]]></category>
            <category><![CDATA[firebase]]></category>
            <category><![CDATA[backend-development]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Thu, 23 Mar 2023 19:12:51 GMT</pubDate>
            <atom:updated>2023-03-23T19:12:51.176Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*w-EiAeBIYDVr72xe9cq28w.png" /></figure><p>Backend as a Service (BaaS) is a cloud computing service model that allows developers to build and deploy the backend infrastructure of their applications without having to worry about the server-side programming, infrastructure management, or other complexities associated with traditional backend development like creating your own backend inNode Js.</p><p>BaaS providers offer pre-built backend components such as databases, authentication, push notifications, and other features that developers can access via API calls or SDKs. This allows developers to focus on building the frontend of their application, while leaving the backend infrastructure to the BaaS provider.</p><p>Here I am discussing some of pros and cons of BaaS.</p><p>Developing an application without a BaaS provider can also be compared to building a house without a construction company. Just like building a house requires a lot of planning, preparation, and effort, building an application also requires a lot of backend infrastructure work such as setting up servers, databases, and APIs.</p><p>Now imagine if there was a service that took care of all the construction work, including hiring workers, managing the construction schedule, and handling all the necessary permits and approvals. This would be similar to using a BaaS provider, which takes care of all the backend infrastructure work, allowing developers to focus solely on building the frontend of their application.</p><p>BaaS generally offers an interactive UI where the developer can see actual data stored in form of database, see users getting authenticated in real time this can be very overwhelming as compared to running custom backend in terminal environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5exsdSveRuvyXRM0b1g49A.png" /><figcaption>Real Time Analysis using Firebase (BaaS)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/0*_ws4FJbJ4oI_ghCP.jpg" /><figcaption>Snippet of custom backend.</figcaption></figure><p>BaaS services also include hosting at early stages so the developer can use that to host the website and later push updates in it.</p><p>Nowadays more and more startups are using BaaS model for building their dream projects. It is because of its simplicity and scalable nature that atracts more and more people towards it.</p><p>Some of key features of Backend as a Services (BaaS) are:</p><p>🎯Time Saving : BaaS can save developers significant time and resources by providing pre-built, ready-to-use backend infrastructure, which eliminates the need for developers to create and maintain their own backend systems. This can reduce the time and cost required to develop and deploy applications. With this saving time company can spend time in other fields like marketing and design.</p><p>🎯Scalability: BaaS providers often offer scalable infrastructure, which allows developers to quickly scale up or down their applications as needed without worrying about managing the underlying infrastructure.</p><p>🎯Easy Integration: BaaS often provides easy integration with popular third-party services, such as payment gateways, social media platforms, and messaging services. This can simplify the development process and reduce the time required to integrate these services.</p><p>🎯 Easy Hosting : Nearly all BaaS providers offers their own hosting services which is very much GUI driven and easy as compared to manually deploying on command line.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fHCC8BkKqKQUPTi-V_PiNw.png" /><figcaption>Google Firebase Services.</figcaption></figure><p>But all that shines isnt gold. All these luxuries of BaaS comes with some conditions.</p><p>🎯Limited Functionality in free version: You can’t expect all the functionality in the base plan. Eg: Google Firebase is one of the most popular BaaS out there. Google smartly offers many of its services in base plan. But as the number of users of the app increases the developers had to buy the spark plan which offers more wider range of functions but it comes with a cost. With app so much scaled developers have no other option but to take the spark plan with causes a hole in pocket.(Atleast for startups in India , due to Dollar Rupee conversion rate 🥲).</p><p>🎯 Unpredictable costs: The main pain in any starup is cash flow. There is literally no regulation on pricing of these BaaS in the market. They can choose to increase price whenver they want and existing users have to agree to their terms to avoid risk of losing thier hard earned user base.Eg: AppWrite another famous BaaS provider increased its price.In January 2022, AppWrite announced a price hike for its paid plans. The platform’s pricing for paid plans increased by about 30% across the board. The company cited increased operational costs as the reason for the price increase.</p><p>🎯Limited Customization : With a basic backend layout already laid down developers dont have much say in redefining their backend . There might be cases when a special backend is needed for a particular project in that case custom backend is the only solution.</p><p>🎯Huge Dependancy: No matter how big a company is there are always server downtime and that will directly affect your project because of complete dependancy on BaaS.</p><p>Some of BaaS to explore are</p><ol><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#ApacheUsergrid"><strong>Apache Usergrid</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#AWSAmplify"><strong>AWS Amplify</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#Back4App"><strong>Back4App</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#Backendless"><strong>Backendless</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#builtioBackend"><strong>built.io Backend</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#Couchbase"><strong>Couchbase</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#Firebase"><strong>Firebase</strong></a></li><li><a href="https://www.hitechnectar.com/blogs/top-14-backend-as-a-service-providers/#Kii"><strong>Kii</strong></a></li><li><strong>AppWrite</strong></li></ol><p>Conclusion: Any BaaS is really flexible and can be inserted into just about any app with a little javascript, or used as a stand-alone solution.</p><p>For a simple application where you’re serving static web assets (HTML and JS) and just need an easy way to store and access data and certain other assets (JSON, Images, etc…) go for BaaS. No point in wasting valuable dev time implementing your own solution.</p><p>If you need some custom back end logic, you can use node.js for that.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dce33b4a9d07" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bing AI Search vs ChatGPT]]></title>
            <link>https://medium.com/@neeraj716sharma/bing-ai-search-vs-chatgpt-1840d3b68ded?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/1840d3b68ded</guid>
            <category><![CDATA[bing-search-engine]]></category>
            <category><![CDATA[future-of-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Sun, 12 Mar 2023 08:38:55 GMT</pubDate>
            <atom:updated>2023-03-12T08:38:55.725Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/474/0*Hy7nQJjdAFvw7HqY" /></figure><p>So after a long wait for its arrival Bing AI tool is finally rolled out for usage. I had huge expectations from it because of the hype Microsoft created about it. Here is a brief comparison i made with the existing chatGPT .</p><ol><li>UI: Bing’s UI is much better than that compared to normal Bing Search or ChatGPT . It offers clean full page UI with lot of white spaces and the plus point is it offers many themes unlike chatGPT which has only light/dark mode only.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/627/1*sOAy-3_c_VtnlP3sWDlJug.png" /></figure><p>2.The response time for Bing AI is somewhat slower than ChatGPT . However this is due to its ability to incorporate live events as well unlike chatGPT which simply responds like “ I dont have knowledge about events post 2020”. But still this response time of Bing AI should be little bit better.</p><p>3. The answers given by Bing AI are not as much clear and elaborative as compared to chatGPT , Bing AI gives the basic answers and in the footer it gives the link to many sites to read further which isnt what expected from an AI model (we can directly web search if we want refrences from websites )</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sTtbLWO0Gkjud1uOHQ5CKQ.png" /></figure><p>4. Another unusual thing I noticed about Bing AI is that it isnt very much comfortable talking about itself . And out of blue it stopped responding to the above questions and made me change the topic. This is not the case with chatGPT one can talk to it about above questions any no of times.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*geRV90nXHiI89XIIZBUWRg.png" /></figure><p>5. It does not respond to coding requests as diligently as chatGPT .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r5xmzDbrVyEGGPWIOpu8zw.png" /><figcaption>Bing AI Response</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xaApD897iMFsKZvRRe89iw.png" /><figcaption>ChatGPT responded with literally the whole code.</figcaption></figure><p>6. Bing AI is too protective and doesnt responds to any controversial topics. I think its good to stay away from any controversies but even non targeted general queries are not responded.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v6ZTyaRyrZkwagPdGJ4Zcg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pmT2UHoAy6V2006ff8QlQg.png" /><figcaption>I don&#39;t support or verify any claims or statements made by chatGPT in this poem.</figcaption></figure><p>7. This one is nice , with chatGPT we often get many wrong answers to maths or physics questions and chatGPT is too gullible that if we just say isnt it wrong to even a correct solution it dillegently accepts its failure and recompute the result, Bing AI is very effective in this term and sticks to its answers unless we provide any concrete proofs of its failure</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-niEfxlRm48j6QmvgxnA-w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EsH3moO5IfXO6hoHu9LSTg.png" /></figure><p>(PS: Bing AI got the ans right in very first time and chatGPT took two attempts to get to the ans + if i say its wrong ChatGPT will again go and find some other answer ).</p><p>8. Currently Bing AI only allows 10 questions per topic i.e. after 10 questions user need to sweep all the questions and start over again . This isnt the case with chatGPT which can chat on same topic for many many questions ( i dont know the exact limit but i tried till 20–25 questions).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pDBX69D6fBQuprENzTRlww.png" /></figure><p>In conclusion i think both models are in early stages of development and will be improved much further with passage of time.</p><p>Bonus: I made both of them write a thank you for reading message tell me which one you liked more.</p><p>Bing AI: <em>I am grateful that you took the time to read and comment on my article. Your feedback and support mean a lot to me. You have given me valuable insights and suggestions that will help me improve my writing. I hope you enjoyed the article and learned something new from it.</em></p><p>ChatGPT: I wanted to take a moment to express my sincere gratitude for taking the time to read my article. Your support and interest in my work is truly appreciated, and it motivates me to continue creating content that is informative, engaging, and valuable to you.</p><p>Thank you for your attention, and I look forward to sharing more with you in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1840d3b68ded" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[@BITS PILANI]]></title>
            <link>https://medium.com/@neeraj716sharma/bits-pilani-873955904828?source=rss-f8625ab779a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/873955904828</guid>
            <category><![CDATA[hackathons]]></category>
            <dc:creator><![CDATA[Neeraj Sharma]]></dc:creator>
            <pubDate>Fri, 25 Nov 2022 17:25:26 GMT</pubDate>
            <atom:updated>2022-11-25T17:25:26.388Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>BITS Pilani Postman API Hackathon 2.0 (co-powered by PIEDS)</em></strong></p><p>So I have to write about our experience about the Hackathon. Frankly speaking if you asked us a week ago how will we manage time for this hackathon with End Sem Exams waiting for us , Practical File submission , Class Tests and 75% Attendance , our answer would be “We won’t be able to compete in this Hackathon”. Cut to the chase we have completed our project a day before extended deadline. I think its all in one’s mentality, we shifted our mind from this idea of non participating to “Lets Try maybe we can make it “ and yes we did! I think it is rightfully said Where there’s will There’s a Way.</p><p><strong>&gt; The Idea Selection</strong></p><ul><li>After a brief WhatsApp discussion on all the topics pro’s and con’s we landed our attention on the topic “ Animal Care : Help Adoption of animals and Keeping an eye on Endangered Species“. We all agreed on it because we all can very well relate to the problems faced by the stray animals we see on everyday roads. And with our Prime Minister releasing Endangered Species of Cheetah in Madhya Pradesh , Endangered Animal Care is indeed a hot topic . We wanted to help the animals who love everyone uncondionally.</li></ul><p><strong>-&gt;The Problems</strong></p><p>The problems we mainly had to encounter was time management , with 5–6 hrs of rigorous classes and a mandatory 75% attendance 😞 our college did challenged us to push our limits. Nevertheless, we tried and managed coding at every free time we get especially late night coding. With Harsh maintaing repository on GitHub we all managed to work parallelly on the problem statement. We roughly divided the roles amongst ourselves and provided constant feedback amongst ourselves on each other’s work. The problems we tried to tackle are : Lack of Knowledge about Endangered Species , Platform for Adoption , Blogs from top Vet docs. on various issues, Help find shelters for Animals and some more features like News on Animals around the globe. We tried our best to implement each and every thing in a user friendly manner so everyone can use them even without a Readme File.</p><p><strong>Technologies and References</strong></p><p>We used some of the latest technologies in successful deployment of our project.</p><p>Some of them include : NPM, Express.js, EJS, REST , API’s,MongoDB, Mongoose , React.js,MongDB Atlas , HTML 5 ,CSS 3, Bootstrap 4.0, Javascript ES6, Jquery ,Git , GitHub . Node.js.</p><p>We took refrences from Resources: <a href="https://www.adoptapet.com/public/apis/pet_list.html">AdoptAPet</a>, <a href="https://www.petfinder.com/developers/">Petfinder</a>, <a href="https://userguide.rescuegroups.org/display/APIDG/API+Developers+Guide+Home">RescueGroups</a>,Wikipedia.</p><p>Images were sourced from <a href="https://www.freepik.com/free-photos-vectors/dog">https://www.freepik.com</a> , <a href="https://pixabay.com/vectors/search/dog/">https://pixabay.com</a>, <a href="https://www.vectorstock.com/">https://www.vectorstock.com</a> .</p><p>We are very obliged to all the community and websites for providing such a great <strong>plethora<em> </em></strong>of amazing resources. <strong>🙏 </strong>💖</p><p><strong>Overall Experience</strong></p><p>It was a great learning experience altogether , this being our second ever Hackathon , we learned a ton from it. We learned important aspects of Team Building , Team Coordination , Managing time, Management of personal ego’s, Working on common grounds and suggestions And these small skills we learned on the way are way more valuable than any other prize any Hackathon has to offer.</p><p>It brought us close to how animals give unconditional love and in return we almost remain deaf ears most of the time. This hackathon changed our perspective towards animals . We now become more empathetic towards the nature’s beautiful creation.</p><p>We know there can be no such thing as perfect success story but as far as we’re concerned from thinking not to participate to finish the project , thinking of quitting midway and finally finishing it was a success altogether.</p><p>Thanks to the organizers for helping new Dev Geeks like us to explore more of opportunities.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=873955904828" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>