<?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 Shruti Mandaokar on Medium]]></title>
        <description><![CDATA[Stories by Shruti Mandaokar on Medium]]></description>
        <link>https://medium.com/@shruti.mandaokar?source=rss-8994c596ef6e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*QG_nMwlxVg0G9677f2-9Mg.jpeg</url>
            <title>Stories by Shruti Mandaokar on Medium</title>
            <link>https://medium.com/@shruti.mandaokar?source=rss-8994c596ef6e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 02:45:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shruti.mandaokar/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[How Eating Chocolate in Grade 3 Led Me to a Dentist Appointment, and I am fixing my habits]]></title>
            <link>https://medium.com/@shruti.mandaokar/how-eating-chocolate-in-grade-3-led-me-to-a-dentist-appointment-and-i-am-fixing-my-habits-51dd921c14b1?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/51dd921c14b1</guid>
            <category><![CDATA[medicine]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[doctors]]></category>
            <category><![CDATA[adulthood]]></category>
            <category><![CDATA[body-image]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Wed, 09 Sep 2026 07:32:55 GMT</pubDate>
            <atom:updated>2026-09-09T07:32:55.821Z</atom:updated>
            <content:encoded><![CDATA[<p>I love my Smile and Dimples.</p><p>Found out Calculus is not just in Math also in Teeth and Gum.</p><p>Growing up, I had the bad habit of snacking and eating ice cream.</p><p>One fine evening, I ate Eclairs (chocolate), drank water, and they stuck to my teeth.</p><p>Before coming to the USA, a doctor helped me with my consultation and treatment of Teeth Cavity. Dental Costs in India are cheaper :) So I had a treatment when I went back to India last year, and the doctor put a silver cap.</p><p>I don’t enjoy visits to the doctor.</p><p>I think in my life, I have been to the hospital so many times.</p><p>Healthcare is nice and a bit expensive in US, hehe, caring for others too, but when it comes to finding solutions for myself, I’m not so good at it.</p><p>I’m like a tortoise.</p><p>chemistry and math: it&#39;s because of u all my reactions in my teeth and pimples on my face are happening because of that.</p><p>I brush twice daily; now I’m trying to follow a routine to fix my dental habits. If I floss, then my gums will bleed — that’s why I don’t.</p><p>I’m using Colgate Optic White Toothpaste and an Oral-B brush for now.<br>I change my toothpaste and brush too.</p><p>I think one problem that I have is eating hot and cold things one after another. I should not do that. My gums are slightly weak.</p><p>I lost a small part of a silver filling in my tooth yesterday. (RIP, my tooth)</p><p>Adulting hits hard, especially in the gums.</p><p>I am also noticing that I cannot eat huge portions of food because of that.</p><p>Most of my day, along with interview prep, is me jumping from one problem set to another. I can’t help but jump from one problem to another, like a kangaroo.</p><p>I make checklists and sometimes plan beforehand.</p><p>When I come back home, there are more things and a better strategy.</p><p>I noticed that sitting for long periods is not so good for the body — stress moves to my lower back, but it is relieved after going to the gym, stretching, and doing yoga exercises.</p><p>My focus is on getting a full-time job, listening to music, dressing up, interview prep, and taking care of myself, fixing my portfolio.</p><p>I am also working to accept my body and things where I don’t feel so beautiful. It is not easy and requires slowly acknowledging your feelings.</p><p>Phew! Phew! Phew! (Sigh!!!!)</p><p>Pimples are also forming a map on my face, bruh. I know it will go away.</p><p>I know about Health Insurance, learnt about Medicare — I’m a noob.</p><p>I change my bedsheets and clean my room too. Maybe it&#39;s also because it&#39;s so sunny in California, and imagine just a year ago I was in Fort Wayne, where it snowed (cool hehe), and my roommates and I made cookies at 4 am. Weather change. Our bodies are quite adaptable.</p><p>I love drinking chai — and I have been drinking it daily. I realise the chai is powerful, but maybe not so good for my teeth. So, I am shifting to smaller portions of food and getting good sleep.</p><p>My Dad and I are kinda similar — we put problems off till the last moment.</p><p>My Mom and my Brother are very proactive in taking care of themselves.</p><p>My Mom takes care of my Dad. and Vice-Versa.</p><p>So, when my Dad makes mistakes, Mom is there to look after.</p><p>My brother is looking for his other half.</p><p>There is some intersection.</p><p>Well, these are the things to take care of your teeth (self-directed)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1lfhVlEaBfwOrfXui8BBuA.png" /></figure><p>Adults have 32 teeth when they have a full set, which includes four wisdom teeth. ( Colgate said that !! )</p><p>Breakdown of Adult Teeth</p><ul><li><strong>8 Incisors:</strong> Front teeth used for cutting food.</li><li><strong>4 Canines:</strong> Sharp, pointed teeth used for tearing food.</li><li><strong>8 Premolars:</strong> Teeth located between canines and molars for crushing food.</li><li><strong>12 Molars:</strong> Back teeth used for grinding, which includes four wisdom teeth</li></ul><p>How do they look?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*56BOZC9-q3psDTkZV8yqIw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xqSLh246SYsg3CJYNpxCGQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1T4UYvdhA7mGV-gJ7uoQjg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1lfhVlEaBfwOrfXui8BBuA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cd-2c1DcV6AqctOOECZEwg.png" /></figure><p>A combination of sugar and acid, especially when found in sticky foods and frequently sipped drinks, destroys tooth enamel the fastest.</p><p>Steps to take care :</p><ol><li>Brush and Floss Daily</li><li>Watch out for Drinks</li><li>Read Home Remedies</li><li><a href="https://dentistpaloalto.com/effective-home-remedies-for-healthy-gums/">https://dentistpaloalto.com/effective-home-remedies-for-healthy-gums/</a></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gh6_zFaa0Eku12IsBjsY0w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v6TveRnzDjrioX_Uz-pyGw.png" /></figure><p>How many teeth do you have left?</p><p>~ Shruti</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51dd921c14b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Discussion: Do Robots Need More Than VLAs & World Models? ]]></title>
            <link>https://medium.com/@shruti.mandaokar/discussion-do-robots-need-more-than-vlas-world-models-7177912b83ed?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/7177912b83ed</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[multimodal]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 18:20:32 GMT</pubDate>
            <atom:updated>2026-09-02T18:20:32.994Z</atom:updated>
            <content:encoded><![CDATA[<p>Notes and open questions from an evening with the 90/30 ML Paper Group</p><p>On August 31, 2026, the topic was robotics — specifically, whether Vision-Language-Action models (VLAs) and world models are actually enough to get us to capable, general-purpose robots, or whether we’re still missing something more fundamental.</p><p>I walked in; the space is colorful and beautiful. Been there — 5 times already, saw existing faces from other hackathons, and events, people chilling, relaxing, and silent for a while on warm colored couches and chairs. I grabbed a Salted Caramel Chocolate and sat on the spot I spotted.</p><p>Did some homework before going in — read the paper a little before walking in. We had some great thought points — it was impromptu and open-ended. I walked out with a longer list of open questions than I came in with — which, in this group, usually means it was a good session.</p><p>Here’s what we talked about and what stuck with me.</p><p>How VLA Models Work</p><ul><li><strong>Vision:</strong> The robot uses a camera or video feed to look at its surroundings.</li><li><strong>Language:</strong> The robot reads or listens to a human instruction, like “pick up the red block”.</li><li><strong>Action:</strong> The internal network processes both inputs and outputs precise robot movements, such as moving a gripper or turning a wrist</li></ul><h3>VLAs are exciting, but “how far” is doing a lot of work</h3><p>A <strong>Vision-Language-Action</strong> model takes visual input and language instructions and maps them directly to robot actions. It’s an elegant idea, and the demos are genuinely impressive. But the room kept circling back to one question: <strong>how far is this realistically from working in the world, not just in a demo?</strong></p><p>The honest answer is: we don’t fully know yet. The field is fresh. The results are new. And “few new demos, still very unknown” is a very different claim than “this is close to solved,” which is what we started with.</p><p>That tension shows up in a bigger architectural debate too — <strong>four pillars versus end-to-end.</strong></p><p>Do we build separate systems for perception, planning, control, and learning and stitch them together (the “four pillars” approach), or do we try to learn the whole pipeline end-to-end and let a data flywheel do the heavy lifting? Nobody in the room thought this was settled, and I don’t think it is.</p><p>If you dig into the paper, the four pillars go into specifics a ton:</p><p>Here is the paper (hehe): <a href="https://arxiv.org/pdf/2606.06556">https://arxiv.org/pdf/2606.06556</a></p><p><strong>V-JEPA vs. VLA: what should a world model actually predict?</strong></p><p>V-JEPA — Meta’s joint-embedding predictive architecture — represents a different bet than VLA.</p><p>Instead of mapping directly from perception to action, it’s a <strong>world-action model</strong>: the idea is that a robot can learn to predict what the world will look like next, and use that predictive model to decide what to do.</p><p>Like — -&gt;<br>predicting what a robot would see going from one floor of a building to another or an unknown space at all.</p><p>The question that split the room: <strong>should a world model predict the next frame in pixel space, or something more abstract?</strong></p><p>Pixels feel like the “obvious” representation — they’re literally what a camera captures. But the world itself changes: lighting, occlusion, objects moving in and out of frame. If a world model is meant to generalize, is raw pixel prediction even the right target? Or do we need models that predict <em>representations</em> of the world rather than its literal appearance?</p><p>We also talked about cross-embodiment — how well a model trained on one robot’s data (or even transport/logistics data) transfers to a robot with a completely different body. This is where a lot of the theoretical promise of world models runs into practical limits.</p><h3>The data problem doesn’t go away</h3><p>No robotics conversation escapes without hitting the data wall, and this one didn’t either.</p><p>A few pieces of it:</p><ul><li><strong>Is there enough data for transformer-based scaling to keep working</strong> the way it has in language and vision? Robotics doesn’t have the internet-scale text corpus equivalent.</li><li><strong>Multimodal data isn’t just vision and language</strong> — it’s physical <em>and</em> chemical. Where exactly is the line between “physical boundary” a model needs to understand and information that’s out of scope?</li><li><strong>2D-to-3D conversion is slow and expensive.</strong> Getting good, usable episodes of 3D interaction data takes real time and real money — this isn’t a scraping problem, it’s a production problem.</li><li><strong>Sensing is the easy part. Interpretation is the hard part.</strong> We can collect sensor data reasonably well. The bridge from raw sensing to actual motor control — the biomechanics of turning perception into movement — is still far off.</li><li><strong>Every application has different data requirements</strong>, which makes collection, annotation, and labeling hard to standardize.</li><li><strong>Sensitive domains don’t share.</strong> Autonomous driving companies, in particular, are protective of their physical-world data — for competitive reasons, but also legitimate privacy ones. That scarcity slows down the whole field, not just one company.</li></ul><h3>Simulation is doing more work than “autonomy” suggests</h3><p>Here’s the part that felt most grounding: a lot of what gets called “autonomous” right now is running on simulators — NVIDIAs Isaac Sim came up repeatedly — paired with <strong>teleoperation</strong>. A human is still very much in the loop.</p><p>Game engines and physics-based, deterministic 3D simulators are genuinely useful for training and testing. But there’s a real gap between “performs well in a deterministic simulated environment” and “performs safely in the real, non-deterministic world.”</p><p>I thought, half-joking, “What is a Roomba?” I had to Google it up — and it was a good reminder that even our most ubiquitous “autonomous” robot is running an extremely narrow, well-bounded behavior, not general intelligence.</p><p>This raised a question I keep coming back to: <strong>instead of chasing general-purpose robots, what if we went narrow and deep?</strong> Build something that’s excellent in one specific domain, rather than mediocre everywhere. That might be a more honest near-term target than full generality.</p><h3>The feedback loop nobody’s measuring</h3><p>When you put an LLM in the driver’s seat of an agent, you introduce a new problem: how do you close the loop between the world model, real physics, and feedback from the environment? Physics is notoriously hard to encode faithfully — it’s messy, continuous, and unforgiving of approximation errors that a language model might not even notice it’s making.</p><p>And on the human side: if a human and a robot are working together, <strong>do we actually have good ways to measure whether that feedback loop is succeeding?</strong> Not clearly. That’s a research gap as much as an engineering one.</p><p>Related: what’s the actual handoff point? How much information does a robot need before a world model has “enough” to perform reliably? Nobody had a crisp answer — which felt honest.</p><h3>What I’m taking with me</h3><p>If there’s a theme across all of this, it’s that <strong>robots don’t just need better models — they need better bridges.</strong> Between pixels and representations. Between simulation and reality. Between sensing and motor control. Between one company’s data and another’s. Between demos and deployment.</p><p>VLAs and world models are real progress. But the 90/30 room’s skepticism felt earned: the interesting problems aren’t just “can we scale the model,” they’re “can we close these gaps at all.”</p><p>Still on my reading list: the four pillars framework in more depth, and <strong>ACT (Action Chunking Transformer) </strong>— a teacher-student imitation learning approach that chunks video frames for training, worth a look on Hugging Face if you’re curious.</p><p>ACT Link: <a href="https://huggingface.co/docs/lerobot/en/act">https://huggingface.co/docs/lerobot/en/act</a></p><p>Exchanged contacts with people in Trust, Safety, and Consulting and also did warm intros; met a Purdue student and asked if he is hiring</p><p>More soon. 🦋<br>- Shruti Mandaokar</p><p><em>Notes from the 90/30 Group meetup, August 31. If you were in the room and I mangled your point, tell me — I’d rather get it right. Feel free to suggest edits if you were in the room.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7177912b83ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My 30 Day Walking Statistics and Consistency ♡]]></title>
            <link>https://medium.com/@shruti.mandaokar/my-30-day-walking-statistics-and-consistency-c9aaffca8705?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9aaffca8705</guid>
            <category><![CDATA[life]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[exercise]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 05:09:30 GMT</pubDate>
            <atom:updated>2026-08-03T05:18:00.425Z</atom:updated>
            <content:encoded><![CDATA[<p>For me, Walking is therapy for my Mental Health. I need to walk outside and see nature every day. According to research, Bone Health deteriorates after the age of 30 in women. Leg muscles are one of the strongest, and they will help you lift weights when you are older. So, why not start today?</p><p>Here is my 30-Day Blueprint: I am a CS student; I love data points, hehe!</p><p>Okay, Story Time :</p><p>It’s Sunday evening, the weather is perfect, you’ve got your favorite playlist blasting in your headphones, and you crush a massive walk. You feel invincible. You look at your phone, see a huge step count, and think, <em>“This is it. This is the new me. I am going to do this every single day.”</em></p><p>I am in Technology and just finished my Master&#39;s in Computer Science. We are glued to our chairs for that debugging session. I am sure it is with you too: Tuesday hits. The emails pile up. You’re glued to your desk chair, staring at a screen for eight hours, and by the time you look down at your wrist, the reality check hits hard.</p><p>That was exactly my life until a few weeks ago. I was trapped in a classic, volatile fitness loop: riding waves of high weekend energy only to completely crash into sedentary habits the moment the workweek started.</p><p>But over a continuous <strong>25-day stretch</strong>, I ran an experiment to fix this.</p><p>I stopped trying to have “massive, heroic fitness days” and instead focused on building a mechanical, non-negotiable floor. Within a single week, something shifted. I didn’t just break my old habits — I unlocked a level of consistency that completely transformed my energy, my focus, and how I show up for myself every day.</p><p>I would like to share this with you and tell you the statistics on how consistency beats motivation. The first step to action always matters.</p><p>If you are tired of the exhausting start-stop cycle and want to see what happens when you track, adapt, and build a routine that actually sticks, let me walk you through the raw, daily receipts of my 25-day journey.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*QKZ5MMrDgAJIC2X7H_Fhlw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QROFDl40k2BMw_YYtQb50A.png" /><figcaption>June vs July, 2026 Comparison : shifting my entire month’s average by +1,942 steps a day just by changing my baseline framework.</figcaption></figure><h3>Phase 1: The Volatile “Before” Era (Days 1–6)</h3><p>Before you can build a system, you have to understand your starting point. My late June data was the perfect picture of an unintentional lifestyle: highly volatile weekend peaks followed by an immediate weekday slide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VXkr8eQTd455hgdvK5t--Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YPPgth6yom2IGdJ11uIzzw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ShSq8OZkspx1jXa7aGYUvQ.png" /><figcaption>Stats for Day 1 — 3</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*li8JHH0BPeAXQUC1h2kg1A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X5Il5sPyaIdNhSbldoJTzw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hhkCxn2Q4EjfU5nzEWvcUg.png" /><figcaption>Stats for Day 4–6</figcaption></figure><p>Here are the insights, the phases, and the growth pattern for my 6 days of consistency. I told myself to keep going — it was beautiful being in nature and people watching. Some people care about their fitness too; they were on their runs or with a cute pet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hQRhaN_BuVhD25qxf-D20w.png" /></figure><h3>Phase 2: The Catalyst &amp; The Reset (Days 7–12)</h3><p>Every transformation story needs an inciting incident. Mine happened on the very first day of July.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RUCtCEHMP3HAyCNlg801Eg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vt7LuYKOqubm45Mi_Bcizw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*laej9L9WBqL1ZN7lner7hw.png" /><figcaption>Stats for Day 7–9</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f-mwELdDDaEqK7UGk6DLXw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d4bk743H9IkGXv_Ozf6gAg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I-PUBa8z7gk3g4LAPNHu1w.png" /><figcaption>Stats for 10–12</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cXZZaS5y1cExAlt3MKLxaw.png" /></figure><h3>Phase 3: The Metronome Era (Days 13–18)</h3><p>If you take away anything from my journey, let it be this section. This is where willpower was replaced by a <strong>system</strong>. I mapped out a specific daily route, anchored it to fixed times, and executed it like clockwork. By this time, my body and brain were used to including walking in my routine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pvxze1j79Mq5fnImNipyOw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wlQ9hobrv9KSBEHU6cyqZg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mKwTFmJ5-HNhmCk5FzI6qA.png" /><figcaption>Stats for Day 13–15</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hv23uHIYciNFdflbfQzZvA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-RPhcJFukEedo6IHKI5FjA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XsvGCNPgHoNV_P55hapB4w.png" /><figcaption>Stats for Day 14–18</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F1TFdmT2SsitGXGuFGChGQ.png" /></figure><h3>Phase 4: Limits &amp; The Human Reality (Days 19–24)</h3><p>Here is the magic of a baseline shift: once your foundation is locked in, your capacity expands. But so does your need for recovery.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_fRV-Ky1J1DcnQ7GVkYKig.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hAAxFmlXz4laAprssMYPcA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x6vE4zxjkaqz8D0-BM1mtQ.png" /><figcaption>Stats for Day 19–21</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8fetnJ1EekOFvBhqnPunPA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h8jmJf1w1pQdiPwTHV_xxw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CR5SCXxQOblVNZUDhFaERA.png" /><figcaption>Stats for Day 22–24</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E5XNqUXBeUWwUN5skQa2sw.png" /></figure><h3>Phase 5: The Consistency(Days 25–30)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iYOOsRiPNi54eorASDjFIA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N_nVLejOO7UnAdRa450Rgw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*deqFkjgUk3BvvouJPu05kQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SP5E-03w4f_1mp5ED6vagg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cJqbbsj7pJF33T-ZnNwiUw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IQtoxQYu90PUcKAcmBk13w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CqDzseJL2LXdeRsOA3Jo_Q.png" /></figure><h3>Phase 5: The Consistency(Days 31–36)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EHgJ9JZZC4ubHuJpVcoIKw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l-BBXE4rJAiWHjH2i5zTtA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NSe7yJIWqm1D5Pxwt37dTg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NoeP4eWKm0ygTxW4Pbj0-g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bh95-8E3QqbFMCrOjk308w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*deiBBdRML98BbJQ8_Wxkxg.png" /></figure><h3>Phase 6: The Ongoing Journey (Days 37–39)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-ExDRtOcUr1EcCpbKDpdag.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5jrgJ0ybOjJNUjAcDAvfjw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/650/1*oc2dHTRotFnCsdbGcc_utw.png" /></figure><h3>The Takeaway:</h3><p>Health is very important, even more than your career. At the end of the day, consistency isn’t about attaining flawless perfection. It’s simply about moving your averages forward. By focusing on systems, choosing to flip that intentional switch, and allowing myself the human grace to rest when needed, my real-time pacing proves that I am fundamentally living at a higher standard than my old baseline ever allowed.</p><p>You don’t need a massive, exhausting life overhaul to feel better, think clearer, and move more. You need to look at yourself in the mirror today, decide what your new floor will be, and then show up for it — day after day, like clockwork. Cool, so what are you gonna do now?</p><p>Thank you for reading. <br>Feel free to check out my other articles.</p><p>~ Shruti Mandaokar</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9aaffca8705" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What a Room Full of Women in Tech Building AI Agents Taught Me ]]></title>
            <link>https://medium.com/@shruti.mandaokar/what-a-room-full-of-women-in-tech-building-ai-agents-taught-me-4a0950b44e25?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a0950b44e25</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[women]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 22:35:58 GMT</pubDate>
            <atom:updated>2026-07-31T22:36:57.641Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Notes from Women Tech Meetup at Entrepreneurs First, San Francisco</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FB7kWSKuT82UyUuivLxztg.png" /></figure><p>I walked into Entrepreneurs First expecting a panel about women in Tech productivity hacks. What I got instead was a much harder conversation — about the line between an <em>AI feature</em> and an <em>AI agent</em>, about which of our jobs agents are actually ready to take over, and about the very human side of anxiety of watching a field move faster than you can keep up with.</p><p>The panel:</p><ol><li><a href="https://www.linkedin.com/in/katesemizhon/"><strong>Kate Semizhon</strong></a> (Engineering Manager at Airbnb, author of a leadership interviews course)</li><li><a href="https://www.linkedin.com/in/cailee-moberg/"><strong>Cailee Moberg</strong></a> (Product at OpenRouter)</li><li><a href="https://www.linkedin.com/in/chandrikamaheshwari/"><strong>Chandrika Maheshwari</strong> </a>(co-founder &amp; CEO of Quivly AI, ex-Plaid, Google, Bain)</li><li><a href="https://www.linkedin.com/in/ivorytang/"><strong>Ivory Tang</strong></a> (founder of Contour, an AI OS for forward-deployed engineers)</li><li><a href="https://www.linkedin.com/in/dashashunina/"><strong>Dasha Shunina</strong></a>, founder of Women Tech Meetup.</li></ol><p>Four different vantage points on the same question:</p><blockquote><strong>what does it actually mean to build with agents right now, not in some hypothetical future.</strong></blockquote><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>1. An AI feature and an AI agent are not the same thing</h3><p>This distinction opened the night, and this is a good basic point to know.</p><blockquote><strong>An AI agent</strong> is like an employee: it can understand a goal, make decisions, take actions, learn from feedback, and adjust its approach.<br> <strong>An AI feature</strong> is like a tool: it performs one specific AI-powered task when asked but doesn’t independently plan or act toward a larger objective.</blockquote><p>It’s a small sentence with a big implication: a feature is a static output.</p><p>An agent is an ongoing process with built-in judgment. A lot of what gets marketed as “AI-powered” today is really just an AI feature — a single well-prompted call bolted onto a product. An actual agent has to know when it’s succeeding, when it’s failing, and what to do differently next time.</p><p>The panel’s bigger point:</p><p><strong>Building has gotten easier and cheaper. Finding the wedge hasn’t.</strong></p><p>Anyone can ship an AI feature this weekend. The hard part is still the oldest part of startups — finding a real problem, understanding how revenue models are shifting (subscription to usage-based pricing came up specifically), and figuring out whether customers are <em>actually</em> adopting what you built, not just whether they clicked “try demo.”</p><p>Their advice was refreshingly unglamorous: <strong>talk to your customers.</strong></p><p>Don’t build purely from your own read of their pain point — that’s how you end up perfectly solving a problem nobody asked you to solve.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>2. You don’t need the frontier model for everything</h3><p>Cailee’s OpenRouter perspective on model selection was one of the most practically useful parts of the night. The question was simple:</p><p><em>How do you pick which model to use for a task — do you always reach for the frontier model?</em></p><p>Her answer: <strong>no, and increasingly, less often than you’d think.</strong> Frontier models keep getting better, but there’s a lot of underutilized headroom in smaller, cheaper models for tasks that don’t need frontier-level reasoning.</p><p>She mentioned that the <em>share</em> of programming tokens going to top frontier models has actually been shrinking — from something like 90% down toward 15% for a given leading model — as more capable mid-tier options become “good enough” for a growing slice of real work.</p><p>Two mental models she offered for picking:</p><ol><li><strong>Classify by determinism.</strong> Break one giant prompt into discrete subtasks, then match the right model to each.</li></ol><p>Classification, translation, and code scanning are largely <strong>deterministic </strong>— they don’t need a frontier model.</p><p>Open-ended reasoning and complex code generation are <strong>highly variable</strong> — that’s where frontier capability actually earns its cost.</p><p><strong>2. Let routing decide for you.</strong> There’s now a whole layer of companies building <strong>model routers</strong> — systems that look at incoming work and dynamically pick the right model, so you don’t have to hand-build that logic yourself.</p><p>One pattern: a lightweight “advisor” model handles the semi-variable stuff and knows when to escalate to a more intelligent model.</p><p>Another, almost the inverse: a smarter model calls a simpler one to do that (she pointed to how Devin approaches this).</p><p>The framework she kept coming back to was the <strong>Pareto frontier of Cost vs. Intelligence</strong> — not every task deserves the smartest, most expensive model;</p><p>The goal should be matching the <em>minimum</em> intelligence a task actually needs to its cost, not maximizing intelligence by default. And underneath all of it: <strong>evals matter enormously.</strong> You can’t route intelligently if you can’t measure whether a model actually did the job well.</p><p>🔗 Worth digging into: <a href="https://openrouter.ai/">OpenRouter</a>, <a href="https://cognition.ai/">Devin (Cognition)</a></p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>3. The engineers struggling most with AI aren’t who you’d expect</h3><p>Kate’s answer to <strong>“who’s adapting well and who’s struggling”</strong> surprised a chunk of the room:</p><p><strong>It’s senior engineers who are struggling the most.</strong> There are early adopters among them, of course, but a meaningful group is resistant to changing how they work — and that resistance has a real cost.</p><p>If you’re not using agents to code today, your productivity is quietly falling behind the people who are. Meanwhile, junior engineers are picking this up faster and more comfortably, which is part of why companies are increasingly excited to hire junior AI engineers straight out of school — the incoming muscle memory is different.</p><p>Her advice for staying relevant wasn’t flashy: <strong>continuously invest in yourself.</strong> Tech has a real revolution roughly every ten years, and this is one of them.</p><p>Keep learning to debug at the systems level, not just the code level — because as agents take on more of the line-by-line work, the skill that matters is understanding how the pieces fit together.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>4. What should stay human, and what should go to agents</h3><p>This was the most structured — and most useful — part of the night.</p><p>The panel broke it into four buckets:</p><ol><li><strong>Human-only work.</strong> Think forward-deployed engineers running client meetings, setting priorities, building relationships. Trust built between people can’t be automated, full stop — this bucket stays human by design, not by limitation.</li><li><strong>Repetitive work that costs time or money.</strong> High-volume customer support is the obvious example. A human doesn’t need to be in that loop — agents can genuinely do it better here because of sheer bandwidth.</li><li><strong>Coordination and scale.</strong> Agents with memory can help humans coordinate across more moving parts than any one person could track — but the human still makes the actual decision. The agent extends judgment; it doesn’t replace it.</li><li><strong>Guardrails, always.</strong> Whatever bucket you’re in, the panel was clear that agentic workflows need explicit guardrails, not implicit trust.</li></ol><p>Ivory added a systems-thinking layer on top of this: <strong>every business problem is, underneath, a system design problem.</strong></p><p>Job titles are already shifting to reflect that — “customer success engineering” instead of plain customer success, for example — because more non-technical roles now need to <em>orchestrate</em> agents, not just use software.</p><p>Her practical framing for<strong> figuring out what to automate</strong>: look at your actual demo targets, your leads, and your revenue processes, and ask which of those steps could be <em>held</em> end-to-end by an agent instead of a person.</p><p>One detail that stuck with me: she pointed out that <strong>certain signals are trivially easy for a human to track and yet exactly where agents shine</strong> — things like noticing a customer’s declining product usage, a spike in usage, or a company fundraise announcement as the right moment to reach out.</p><blockquote>Humans are good at reading a room; agents are good at reading a thousand rooms at once and flagging the one that needs you.</blockquote><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>5. Everyone is building agents now, not just using them</h3><p>A theme that ran through the whole panel: the expectation has quietly shifted from “use AI tools” to <strong>“build your own agents, in natural language, as part of your actual job”</strong> — whether that’s product, customer acquisition, or ops.</p><p>Several panelists described this as roughly an even race between humans and agents at this point — not humans supervising agents from a distance, but both sides actively building and improving the same systems together.</p><p>The practical note that made me laugh: <strong>it’s genuinely fast to get an agent to 80% functional.</strong> The bottleneck isn’t the initial build — it’s that the rest of your company won’t actually adopt it unless that last 20% is right.</p><p>Building the agent is the easy part;</p><p>getting your teammates to trust and use it is the actual product problem.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>6. How do you keep a team curious without making them dependent?</h3><p>Someone asked the question I think every engineering lead is quietly wrestling with right now: <strong>how do you encourage AI adoption without your team losing the muscle that got them here?</strong></p><p>The panel’s answer had a clear exception carved out: <strong>encourage AI use broadly — except for new grads and interns.</strong> Coding with AI is faster, but if you’re still building foundational skills, writing everything with an agent from day one means you never build the judgment to know when the agent is wrong.</p><p>Some of the panelists run internal bootcamps that deliberately <em>discourage</em> full AI reliance for junior engineers early on — not because AI is bad, but because skill built without it transfers, and skill built entirely through it sometimes doesn’t.</p><p>The junior engineers who use AI to <em>learn</em>, rather than to skip learning, end up faster and more capable in the long run — with or without the tool in hand.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>7. Lightning round: what tools are you actually using?</h3><p>The audience Q&amp;A turned into a genuinely fun tools roundup, and I wrote down all of it:</p><ul><li><strong>WhisprFlow</strong> — came up more than once as a favorite for talking out loud and having it consolidate your rambling into clean paragraphs. Several panelists use it just for thinking.</li><li><strong>Granola</strong> — for meeting notes.</li><li><strong>Pound</strong> — described as a personal-assistant-style tool, particularly good at scheduling.</li><li><strong>Hermes</strong> — one panelist runs a shared agent across a household Slack space handling grocery ordering, home supplies, and scheduling with friends.</li><li><strong>Claude Enterprise</strong> — used specifically for writing performance reviews against a rubric that incorporates peer feedback.</li><li>An <strong>internal product-manager agent</strong> one team built pulls customer feedback from X, Discord, and Reddit alongside product metrics daily, and spins up a ticket in Linear for each pattern it finds — with a human-in-the-loop check every morning before anything ships. The team was clear that this doesn’t replace product strategy and vision — it clears the noise so a human can focus on the direction.</li></ul><p>There was also a sharper, security-focused thread from one panelist with a payments and routing background: different models have genuinely different strengths <em>and</em> different vulnerabilities when it comes to adversarial use — what she called AI offensive security.</p><p>Her point was that benchmark performance on security tasks varies as much between models as raw coding ability does, and that the same routing logic (autorouter vs. Pareto-style “minimum intelligence needed”) applies just as much to picking a model for a security-sensitive task as it does for picking one for a coding task.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>8. On the emotional side of all this</h3><p>The most honest moment of the night wasn’t about tools at all. S</p><p>omeone brought up that AI agent adoption has been notably slower in some regions — the Nordics were mentioned specifically — and that pushback has skewed differently by gender: some of the resistance the panel has observed has come from women navigating changed workflows, while some men have had more free time to experiment their way into comfort with the tools.</p><p>The panel didn’t offer a tidy fix. What they offered instead was permission to feel <strong>anxious, annoyed, or angry</strong> about the pace of change — and then a practical way through it:</p><ul><li><strong>Be patient with yourself.</strong> This is genuinely disorienting for everyone, not just you.</li><li><strong>Let curiosity be the motivator</strong>, not obligation.</li><li>Actually <strong>block an hour to practice</strong>, and sign up for a real class if that helps you commit to it.</li><li><strong>Be intentional with your time </strong>— a group of friends, a demo night, something with real stakes attached tends to stick better than solo tinkering.</li><li><strong>Do the parts that are actually fun</strong>. If it feels like homework, you’ll quietly stop.</li></ul><p>And a closing question from the founder of <strong>SouledSpace</strong>, an AI-powered emotional recognition platform, that felt like the right note to end the whole evening on, she asked:</p><p>As AI gets more powerful and more active in our lives, it’s on us to keep user trust and control central — and to be honest about something she called <strong>“AI psychosis”</strong>: the habit of going to AI for your <em>first</em> answer before you’ve thought it through yourself.</p><p>The advice and answer wasn’t to distrust AI — it was to take the time to understand your own principles before delegating your thinking to something else. So much of what we call “AI engineering” right now, she said, is really just human engineering, wearing a new name.</p><h4>— — — — — — — — — — — — — — — — — — — — — — — — — — — — —</h4><h3>The thread connecting all of it</h3><p>If I zoom out, this whole evening was really one long answer to a single question: <strong>as building gets cheaper and agents get more capable, what’s actually left for us to do?</strong></p><p>The panel’s collective answer wasn’t “nothing” — it was: the <strong><em>judgment</em></strong>. Knowing which model a task actually needs. Knowing which relationships can’t be automated. Knowing when your junior engineer needs to struggle a little before you hand them the shortcut. Knowing when to trust your own first instinct instead of an agent’s.</p><p>Building the agent, it turns out, was never really the hard part. Knowing what to build it for — and when to still do it yourself — still is. 🌸</p><p>Thank you for reading. Feel free to check out my other articles.</p><p>~ Shruti Mandaokar</p><p><em>This piece is based on notes from Women Tech Meetup’s “Do More With Less” panel, held July 29 at Entrepreneurs First, San Francisco.</em></p><p><strong>References &amp; further reading:</strong></p><ul><li><a href="https://www.linkedin.com/company/women-tech-meetup/">Women Tech Meetup</a></li><li><a href="https://www.linkedin.com/in/ivorytang/">Ivory Tang — Contour</a></li><li><a href="https://www.linkedin.com/in/chandrikamaheshwari/">Chandrika Maheshwari — Quivly AI</a></li><li><a href="https://www.linkedin.com/in/cailee-moberg/">Cailee Moberg — OpenRouter</a></li><li><a href="https://www.linkedin.com/in/katesemizhon/">Kate Semizhon — Airbnb</a></li><li><a href="https://www.youtube.com/watch?v=nJl2BWC60SY">Dasha Shunina — Talks With Dasha</a></li><li><a href="https://openrouter.ai/">OpenRouter</a></li><li><a href="https://www.joinef.com/">Entrepreneurs First</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a0950b44e25" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agents Need Infrastructure and Identity, Not Just Prompts ]]></title>
            <link>https://medium.com/@shruti.mandaokar/agents-need-infrastructure-and-identity-not-just-prompts-what-i-learned-at-the-data-ai-meetup-f6a47f1d7274?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6a47f1d7274</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[infrastructure]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 22:11:47 GMT</pubDate>
            <atom:updated>2026-08-01T02:00:32.010Z</atom:updated>
            <content:encoded><![CDATA[<p>Notes, questions, and a few “wait, that changes everything” moments from an evening at AWS Builder Loft at the Data + AI Meetup on July 22, 2026.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FpbxFgAZOzQ63QY7ldRtiQ.png" /></figure><h3>The Speakers :</h3><ul><li><a href="https://www.linkedin.com/in/krantiparisa/?utm_source=luma">Kranti Parisa</a>, Founder of <a href="https://www.linkedin.com/company/laserdata/?utm_source=luma">LaserData</a> and <a href="https://www.linkedin.com/company/apache-iggy/?utm_source=luma">Apache Iggy (Incubating)</a> PMC member, on Rust-based streaming.</li><li>​<a href="https://www.linkedin.com/in/bharos92/?utm_source=luma">Bharath Krishna</a> and <a href="https://www.linkedin.com/in/thoruto/?utm_source=luma">Akshay Thorat</a>, Senior Engineers at <a href="https://www.linkedin.com/company/roku/?utm_source=luma">Roku</a>: an AI agent that takes a request in chat and ships an approved merge request on its own. Every hop authenticated, zero hardcoded tokens.</li><li>​<a href="https://www.linkedin.com/in/vatsal-trivedi/?utm_source=luma">Vatsal Trivedi</a>, Founder of <a href="https://www.linkedin.com/company/runlog-atlas/?utm_source=luma">Runlog Atlas</a>: semantic collapse, what happens when retrieval mistakes noise for relevance, and how the brain’s trick of exciting and inhibiting memories inspired their Atlas architecture.</li><li>​<a href="https://www.linkedin.com/in/guy-lubovitch/?utm_source=luma">Guy Lubovitch</a>, Director of Customer Engineering at <a href="https://www.linkedin.com/company/falkordb/?utm_source=luma">FalkorDB</a>: Store Your AI Agent’s Memory and Context in a Knowledge Graph</li><li>​<a href="https://www.linkedin.com/in/markhoerth/?utm_source=luma">Mark Hoerth</a>, Product Lead at <a href="https://www.linkedin.com/company/datastrato/?utm_source=luma">Datastrato</a>: keeping a multi-model estate governed and queryable through one open catalog, <a href="https://www.linkedin.com/company/apache-gravitino/?utm_source=luma">Apache Gravitino</a>.</li></ul><p>I went into this meetup expecting the usual “which model should I use” conversation. Instead, the opening line from the talk flipped the whole night on its head:</p><blockquote><strong><em>Don’t ask “which model should I use?” Ask “how does data move between agents?”</em></strong></blockquote><p>That question — posed by <strong>Kranti Parisa</strong> (co-author of <em>Reliable ML</em> and a contributor to Apache Solr) — is the thread I want to pull on in this piece. Once you’ve built more than one agent, you stop caring about which LLM is smartest and start caring about something much less glamorous: Infrastructure. Context. Identity. Memory.</p><p>The boring infrastructure that decides whether your “multi-agent system” is actually reliable, or just five API calls wearing a trench coat.</p><p>Five speakers, five different angles on the same underlying problem. Here’s what I took away.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>1. Agents don’t need APIs. They need streams.</h3><p>Picture a customer support flow:</p><p><em>Refund my order #123.</em></p><p>That single request has to pass through a chain — <strong>Planner → Inventory → Payments → Shipping → CRM → Analytics</strong> — before a human ever sees it again.</p><p>Now imagine the customer texts back,</p><p>“Actually, it’s order #1234, sorry.”</p><p>In a normal request/response world (the kind built on REST APIs), that correction has to be manually re-propagated to every single agent downstream — inventory, payments, shipping — each of which may have already started working off the wrong number.</p><p>No components automatically “hear” the update or the change.</p><p>This is the core limitation Kranti was pointing at: <strong>Microservices exchange requests. Agents need to exchange context that changes between parts.</strong></p><p>APIs are built for one-shot request/response, not for a question that gets revised mid-process and needs to ripple through eight downstream systems at once.</p><p>His answer is a shift from APIs to S<strong>treams</strong> — specifically <strong>Apache Iggy</strong>, an open-source, Rust-based message streaming engine currently incubating at Apache. A few things make it interesting for agentic pipelines specifically -</p><ul><li>It’s built for <strong>very low latency</strong> — benchmarks put it dramatically faster than Kafka on comparable workloads, largely because of its thread-per-core, zero-copy architecture rather than being a layer bolted on top of existing infrastructure.</li><li>It has an <strong>embedded database</strong>, so you can write into the stream <em>and</em> query it at the same time — the data is materialized as it flows, instead of you having to wait for it to land somewhere else before you can ask questions about it.</li><li>It exposes primitives like <strong>key-value lookups</strong> and <strong>graph queries</strong> directly, so an <strong>“agent + embedded gateway”</strong> pattern becomes possible — Redis can be plugged in as the key-value layer alongside it.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pAZMTIG0PFqMFm5e9grYvg.png" /><figcaption>REST API vs Streaming</figcaption></figure><p>In the image above, the REST world (ask once, get an answer; a correction after that point never reaches the agent) is shown in the bottom panel, and the streaming world (agents subscribe once, and a correction propagates live to everyone downstream without anyone having to re-ask) is shown in the top panel.</p><p>Kranti also referenced something called the <strong>Agent Data Exchange Protocol (AGDX) </strong><a href="http://agdxprotocol.ai/">agdxprotocol.ai</a> — a proposed low-level protocol specifically for how agents exchange data with each other, rather than repurposing REST for a job it wasn’t designed for.</p><p>The honest, slightly uncomfortable takeaway: <strong>we may be rebuilding fundamental data infrastructure again, just for agents this time.</strong></p><p>Streaming, materialized views, event-driven architecture — these aren’t new concepts. What’s new is that <em>autonomous agents</em> are now the consumers, and they need context updates to propagate automatically, not on a human’s next curl command.</p><p>🔗 <a href="https://iggy.apache.org/">Apache Iggy — official site</a> · <a href="https://github.com/apache/iggy">GitHub</a></p><p>Laser SDK Repo : <a href="https://www.linkedin.com/safety/go?url=https%3A%2F%2Fgithub.com%2Flaserdata%2Flaser-sdk&amp;trk=flagship-messaging-web&amp;messageThreadUrn=urn%3Ali%3AmessagingThread%3A2-YmU1MWZjMjYtZDY4Ni00NGE5LWI3ODYtZTJiMjI1NWQzZDJjXzEwMA%3D%3D">github.com/laserdata/laser-sdk</a></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>2. Who is this agent, actually? The identity problem nobody solved yet</h3><p>The second talk, from <strong>Akshay Thorat and Bharat Krishna (Roku)</strong>, started with a question that stuck with me:</p><p><em>If an AI agent needs ten different data sources to answer one question, whose credentials is it using to get in?</em></p><p>Think about how many tools a single agentic workflow can touch — Trino, Hive, Looker, DataHub, Gravitino. In a human-only world, you log into each one separately, and IT can track exactly who accessed what. In an agent world, that model breaks down fast:</p><ul><li><strong>Secret sprawl</strong> — every new data source means another credential to mint, rotate, and secure.</li><li><strong>No real identity</strong> — is the agent acting as <em>itself</em>, or <em>on behalf of a user</em>? Most systems can’t tell the difference, which makes auditing nearly meaningless.</li></ul><p>Their answer was a project called <strong>AgentBox</strong> — essentially a <strong>single, unified authentication gateway</strong> that sits in front of every MCP server and data tool an agent might need. One token, once you’re authorized, gets you into the whole ecosystem — warehouses, metadata layers, BI tools — instead of a different login for each.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pAZMTIG0PFqMFm5e9grYvg.png" /><figcaption>AgentBox flow — human login → short-lived token → agent action → audit log → expiry, with the loop-back reminder that the agent never keeps standing access</figcaption></figure><p>What I appreciated is that they designed for <strong>three very different kinds of “users”</strong>:</p><ol><li><strong>People inside the company</strong> — standard OAuth login, refresh tokens, a CLI/proxy layer that routes everything through the AgentBox server.</li><li><strong>Non-human agents</strong> — authenticated via workload identity (think Google OIDC + Kubernetes workload identity), so a background agent has its own verifiable identity, not a shared service account.</li><li><strong>Agents invoked from Slack</strong> — the human logs in once, an AgentBox token gets generated, and from then on the Slack agent can act <em>on the human’s behalf</em> — with an explicit approve/deny step for anything destructive.</li></ol><p>The security principles were refreshingly simple: <strong>clients never hold raw backend credentials, everything is encrypted at rest, tokens are short-lived by default, and access follows least-privilege.</strong> And critically — every action gets logged back to a central audit trail, so “who did what” is answerable even when “who” is a piece of software.</p><p>There’s also a <strong>Swarm</strong> platform layer they described, where mentioning @kyubii in a Slack thread spins up an isolated, time-boxed session pod on Kubernetes to actually go fix the issue — complete with resource quotas, network policies restricting it to internal tools only, and a one-hour idle timeout so nothing lingers around unsupervised.</p><p>The underlying philosophy, stated plainly: <strong>agents shouldn’t have their own permanent identity with standing power — they should borrow a human’s identity for exactly as long as the task takes.</strong></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>3. More data doesn’t mean a smarter RAG system — it can mean a dumber one</h3><p>This is the talk that surprised me, from <strong>Vatsal Trivedi of Runlog AI</strong>.</p><p>We tend to assume that adding more documents to a retrieval system (RAG) makes it <em>more</em> capable. Vatsal’s research says the opposite can happen, and he called the failure mode <strong>semantic collapse</strong>.</p><p>The case study: a retrieval system started with 54 documents and hit around 75% retrieval accuracy. As the corpus grew to 1,128 documents, accuracy <em>dropped</em> to roughly 40% — even though, on paper, the system had far more relevant information available to it.</p><p>Why?</p><p>Because as the number of documents increases, the <em>confidence scores</em> of vector similarity keep climbing too — the system becomes more confident even as it becomes less correct.</p><p>Traditional RAG systems were never built to detect the point where that confidence stops being trustworthy — there’s no built-in alarm bell for “I’m now guessing, not retrieving.”</p><p>He referred to the sweet spot before this happens as the <strong>Goldilocks zone</strong> — not so little data that you lack context, not so much that everything starts blurring together.</p><p>To explain why, he reached for a neuroscience analogy that I loved: <strong>a single neuron fires only once excitatory and inhibitory signals cross a threshold — not gradually, but as a discrete event.</strong></p><p>Frequently-traveled neural pathways get reinforced into long-term memory, the way a well-used highway gets wider over time.</p><p>Vector databases, by contrast, don’t really do this — they don’t evolve their own structure as new data arrives; they just keep growing the same flat space, and eventually everything starts looking similar to everything else.</p><p>His proposed alternative is a framework called</p><p><strong>FECS</strong> — Data → <strong>F</strong>actions → <strong>E</strong>ntities → <strong>S</strong>ignatures → Cells:</p><ul><li><strong>Entity extraction</strong> (via an LLM) and <strong>entity resolution</strong> (via an algorithm) work together, backed by a mix of vector databases and <strong>LSH (Locality-Sensitive Hashing)</strong> for structural similarity — essentially, vocabulary rules for what “counts” as the same concept.</li><li><strong>Factions</strong> act like citable units — bundles of “cells” (stored as JSON) with clear provenance, so you can trace <em>why</em> a piece of information was retrieved, not just <em>that</em> it was.</li><li>A mechanism he called the <strong>SPINE algorithm</strong> governs excitation/inhibition — deciding what gets reinforced and what fades, echoing that neuron analogy directly.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fSwuS2FJ3P97vIzJreC_vA.png" /><figcaption>That’s the semantic collapse curve — accuracy climbs as the corpus grows from 54 to around 300 documents, then drops steadily as it scales past 1,000, with the goldilocks zone shaded where retrieval was most trustworthy.</figcaption></figure><p>The bigger idea underneath all of it: instead of dumping everything into one giant similarity space and hoping the “closest” match is the right one, <strong>go directly to where the context actually originates</strong> — and let the system update itself as a closed feedback loop, the same way repeated use strengthens a memory in a brain.</p><p>It’s a good reminder that “more data” is not automatically “better retrieval,” and that we probably need retrieval systems that know when to say <em>I’m not confident anymore</em>, instead of just getting louder as they get worse.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>4. Teaching agents what data actually <em>means</em> — not just where it lives</h3><p><strong>Mark</strong>, from <strong>Datastrato.ai</strong> (the team behind the open-source project <strong>Apache Gravitino</strong>), tackled a problem that’s easy to underestimate:</p><p>An agent can technically <em>access</em> a table and still have no idea what that table <em>means</em>.</p><p>Gravitino positions itself as a <strong>governance layer sitting on top of a wide range of data sources</strong> — spanning engines like Trino, Spark, and DuckDB, and sources from Teradata and Oracle to S3, GCS, and ADLS — built on Apache Iceberg to unify them into one queryable data lake.</p><p>Datastrato describes it, aptly, as a <strong>“catalog of catalogs.”</strong></p><p>He framed the agentic-era problem in five layers:</p><ol><li><strong>Modeling</strong> — what do these numbers actually <em>mean</em>? (Ontology, metrics, glossary)</li><li><strong>Governed access</strong> — identity and policy enforced right on the query path, not as an afterthought</li><li><strong>Identity</strong> — who is the actual principal behind this request?</li><li><strong>Federation</strong> — one catalog that spans many underlying catalogs</li><li><strong>Sources</strong> — the compute and storage systems underneath it all</li></ol><p>The part I found most conceptually interesting was the discussion of <strong>ontology</strong> — essentially a knowledge graph that captures how everything in an organization’s data relates to everything else, expressible in both directions (backward and forward).</p><p>For something like a financial organization, that means the graph doesn’t just store tables — it stores the <em>relationships</em> between customer accounts, finance records, transactions, and more.</p><p>Here’s where it got genuinely elegant: <strong>everything becomes a triple</strong> — Subject → Predicate → Object.</p><p>This is the classic <strong>RDF (Resource Description Framework)</strong> model, where:</p><ul><li><strong>RDF</strong> is the data model (everything is a triple)</li><li><strong>Turtle</strong> is a syntax for writing those triples down (.ttl files)</li><li><strong>OWL</strong> layers meaning and rules on top of the raw triples (also written in Turtle)</li><li><strong>SPARQL</strong> is the query language for asking questions of the graph</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rPvNiGEh2i8MHBdSM2t97Q.png" /><figcaption>RDF triple example builds on the container-with-rows pattern, so here’s the triple shown as subject → predicate → object, with the RDF/Turtle/OWL/SPARQL stack underneath it for context.</figcaption></figure><p>So instead of an agent blindly running SQL against a table it doesn’t understand, the flow becomes: <strong>ask in plain English → look up metadata via MCP → check the ontology → then, and only then, run a governed read</strong> — with enforcement (via policy-as-code tools like OPA) happening at the query engine itself, e.g. row- or column-level filtering applied directly by Trino before results ever reach the agent.</p><p>And critically: <strong>the agent borrows the human’s identity for the duration of the request.</strong> It doesn’t get to accumulate standing power of its own — a theme that echoed AgentBox’s approach from earlier in the night.</p><p>🔗 <a href="https://gravitino.apache.org/">Apache Gravitino — official site</a> · <a href="https://datastrato.ai/">Datastrato</a> · <a href="https://github.com/apache/gravitino">Gravitino on GitHub</a></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>5. Giving agents a memory that survives the conversation ending</h3><p>The closing talk, from <strong>Guy Lubovitch</strong>, addressed the problem every long-running agent eventually hits: <strong>the context window is short-term memory, and everything outside it just… disappears.</strong> No continuity across sessions, no real personalization, no sense of “we’ve talked about this before.”</p><p>His framework for agent memory rested on four pillars:</p><ul><li><strong>Graph-structured memory</strong> — storing not just facts, but the <em>relationships</em> between them</li><li><strong>Per-user (or per-graph) isolation</strong> — every user or agent gets their own memory graph, not one giant shared blob</li><li><strong>Memory evolution</strong> — support for updating facts and resolving conflicts as new information arrives</li><li><strong>Context-aware retrieval</strong> — semantic search over the graph via vector embeddings, so retrieval understands meaning, not just keywords</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d2kkquG6IU65s-yKx_Fbrg.png" /><figcaption>Memory graph — EdoMatrix as the central node, with Hiro Tanaka and Aiko Watanabe as person nodes carrying their roles, and Tokyo as the location node, exactly the kind of structure Cognee writes into FalkorDB as facts</figcaption></figure><p>The live demo used a genuinely fun stack: a <strong>Telegram bot</strong> connected to an agent (<strong>OpenClaw</strong>, running on Ubuntu) that used <strong>Cognee</strong> as its memory control plane, with <strong>FalkorDB</strong> — a Redis-based, in-memory graph database — actually storing the connected memory graph underneath.</p><p>Every fact typed into the chat became a node; every relationship between facts became an edge. Mention a person twice, and the graph reuses the existing node instead of duplicating it — which is a small detail, but it’s exactly the kind of consistency that makes an agent feel like it’s actually <em>remembering</em> rather than just logging.</p><p>One especially clean idea: giving two <em>completely separate</em> agents their own isolated graphs within the same FalkorDB instance, so their memories never bleed into each other — one Redis-backed system quietly running thousands of independent graphs at once.</p><p>A bit of lineage that I didn’t know before this talk: <strong>RedisGraph was originally a Redis module</strong>, and after Redis restructured as a company, that module effectively evolved into what’s now the independent <strong>FalkorDB</strong> project.</p><p>🔗 <a href="https://www.falkordb.com/">FalkorDB</a> · <a href="https://www.cognee.ai/">Cognee</a> · <a href="https://docs.falkordb.com/agentic-memory/">FalkorDB’s agentic memory docs</a></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>The thread connecting all five talks</h3><p>If I zoom out, every single talk that night was really about the same four things Kranti named right at the start: <strong>context, memory, coordination, and replay.</strong> Just from different angles:</p><ul><li><strong>Kranti</strong> said: context needs to <em>flow</em> between agents, not sit frozen in a request/response pair.</li><li><strong>Akshay and Bharat</strong> said: coordination breaks down without a real, auditable identity for every agent.</li><li><strong>Vatsal</strong> said: more context isn’t automatically better context — retrieval needs to know its own limits.</li><li><strong>Mark</strong> said: data needs <em>meaning</em> attached to it, not just access controls, or agents will misinterpret what they’re allowed to touch.</li><li><strong>Guy</strong> said: memory has to persist and evolve across sessions, or every conversation starts from zero.</li></ul><p>None of these are new problems in computer science.</p><p>Streaming, identity and access management, information retrieval, data governance, and stateful memory have all been solved — imperfectly, iteratively — for decades.</p><p>What’s new is that the <em>consumer</em> of all this infrastructure is now an autonomous agent instead of a human clicking through a dashboard, and that changes the stakes on every one of these layers.</p><p>An agent that misreads a table’s meaning doesn’t just get a wrong answer — it can act on it, downstream, without a human in the loop to catch the mistake.</p><p>We’re not inventing new infrastructure so much as we’re <strong>re-deriving the same fundamentals, under a much less forgiving set of constraints.</strong></p><p>That’s the part I keep sitting with. 🌸</p><p><em>Thank you for reading.</em></p><p>~ Shruti Mandaokar</p><p><em>Any errors in interpretation are my own — reach out if you were in the room and want to add nuance!</em></p><p><strong>References &amp; further reading:</strong></p><ul><li>Apache Iggy: <a href="https://iggy.apache.org/">iggy.apache.org</a></li><li>Apache Gravitino: <a href="https://gravitino.apache.org/">gravitino.apache.org</a> · <a href="https://datastrato.ai/">Datastrato</a></li><li>FalkorDB: <a href="https://www.falkordb.com/">falkordb.com</a></li><li>Cognee: <a href="https://www.cognee.ai/">cognee.ai</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6a47f1d7274" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[San Carlos is Sunny]]></title>
            <link>https://medium.com/@shruti.mandaokar/san-carlos-is-sunny-d5112b45f6fd?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d5112b45f6fd</guid>
            <category><![CDATA[life]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[relaxation]]></category>
            <category><![CDATA[california]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 17:46:39 GMT</pubDate>
            <atom:updated>2026-07-27T17:46:39.223Z</atom:updated>
            <content:encoded><![CDATA[<p>San Carlos is <em>very</em> sunny.</p><p>On most days, that’s a blessing. Blue skies, colorful flowers lining the sidewalks, quiet neighborhoods, and roads that make walking feel effortless. It almost invites you outside.</p><p>Lately, I’ve been trying to stay consistent with my daily walking goals. It’s one of those habits that clears my mind as much as it helps my body.</p><p>But sometimes, the California sun is… a little too enthusiastic.</p><p>While I know soaking up some Vitamin D is good for you, standing under the afternoon sun for too long also makes me think about taking care of my skin. Some days, the heat alone is enough to make you crave something cold every few minutes.</p><p>It reminds me of childhood.</p><p>Back then, the solution was simple: ask Mom or my <em>Mausi</em> if we could go out for ice cream. Those little outings weren’t just about the dessert — they were tiny celebrations tucked into ordinary days.</p><p>Now, adulthood has entered the chat.</p><p>Before buying that ice cream, I instinctively do a quick mental calculation:<br><em>“Do I really want this today? Should I save those few dollars instead? I have to catch a train, and I have 5$ — so let’s skip the ice cream”</em></p><p>It’s funny how priorities quietly change. You start budgeting without even realizing it. The spontaneous treats become intentional ones.</p><p>And honestly, that’s okay.</p><p>I’ve learned that not every beautiful day has to be spent outside. Sometimes the best decision is to stay indoors, relax, recharge, and let the evening decide the rest.</p><p>Maybe I’ll head out for a walk once the sun softens.</p><p>I’ll cook something comforting.</p><p>Or maybe I’ll simply enjoy doing… nothing.</p><p>I have to study, hehe!</p><p>Over the past few weeks, I’ve explored San Carlos, wandered through Redwood City, and spent time in San Francisco. Each place has its own personality.</p><p>As children, happiness often arrived in the form of an ice cream cone handed to us by someone we loved.</p><p>As adults, happiness sometimes looks like making a thoughtful choice, cooking dinner at home, watching the sunset from your window, and realizing that’s enough or meeting new people outside from far.</p><p>And perhaps that’s its own kind of sweetness.</p><p>~ Shruti ❤</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d5112b45f6fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Night Owl and Musings ❤]]></title>
            <link>https://medium.com/@shruti.mandaokar/night-owl-and-musings-be4bb034439d?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/be4bb034439d</guid>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 06:35:01 GMT</pubDate>
            <atom:updated>2026-07-27T06:35:01.511Z</atom:updated>
            <content:encoded><![CDATA[<p>It’s 4:05 AM on July 23, 2026. I have a few upcoming events and interviews to prepare for, so I chose to wake up early. Late nights and early mornings have always been my most productive hours — when the world is quiet, distractions fade, and my mind feels at its sharpest.</p><p>It reminds me of my school days, when my parents had enrolled me in Indian classical music classes. During exam season, my mom would gently wake me up at dawn so I could study while my mind was still fresh.</p><p>As much as I wanted to sleep. That routine stayed with me. Sometimes I would pull all-nighters before exams (not healthy! I don’t recommend it).</p><p>I’ve always been a visual and repetitive learner — I understand concepts best by drawing them out, writing on paper, reciting them aloud, and explaining them to myself or someone else. Speaking ideas out loud has never just been revision for me; it’s how I truly learn.</p><p>Another thing I love about late nights and early mornings is the silence. The world seems to pause for a while, leaving you with uninterrupted time to focus, create, reflect, or breathe.</p><p>Something is calming about watching the city lights shimmer outside your window while everyone else is asleep. And if you’re lucky enough to have a balcony, stepping outside to feel the cool breeze is a quiet reminder that not every moment has to be rushed. Sometimes, the stillness itself is enough to clear your mind.</p><p>One habit my mom instilled in me was to begin the day gently. Before rushing into work or studies, she’d encourage me to drink something warm, freshen up, and wash away the sleepiness from the night.</p><p>I’ve never been perfectly consistent with meditation, but I’ve found my own version of a mindful morning. Some days it’s a quiet walk after the world wakes up. Other days it’s journaling a few thoughts, or simply letting calm music play in the background while I organize my mind for the day.</p><p>I’ve realized that a peaceful morning isn’t about following the perfect routine. It’s about creating a few moments that make you feel grounded as the day unfolds.</p><p>For many of us, this is the beginning of a new day. But somewhere else, it’s the middle of someone’s workday.</p><p>Think about nurses caring for patients through the night, pilots flying across continents, truck drivers covering thousands of miles, emergency responders answering calls, road construction crews repairing highways while the roads are empty, or factory workers keeping essential services running. Their body clocks are completely different from ours.</p><p>While most of us associate the night with rest, for them it’s when they’re at their most alert, productive, and focused. It makes you realize that there’s no single “right” time to do your best work.</p><p>Whether you’re an early bird, a night owl, or someone whose schedule changes with every shift, what matters most is finding the hours when your mind and body work in harmony.</p><p>~ Shruti Mandaokar</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=be4bb034439d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What a Master’s in Computer Science Actually Teaches You (Beyond the Code) ❤]]></title>
            <link>https://medium.com/@shruti.mandaokar/what-a-masters-in-computer-science-actually-teaches-you-beyond-the-code-09196bf13d0a?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/09196bf13d0a</guid>
            <category><![CDATA[life]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Tue, 21 Jul 2026 21:25:59 GMT</pubDate>
            <atom:updated>2026-07-21T21:25:59.954Z</atom:updated>
            <content:encoded><![CDATA[<p>My take on finishing a Master&#39;s in Computer Science.</p><p>If you look at an M.S. in Computer Science strictly from the outside, it looks like a purely technical flex.</p><p>People see the credentials on paper: distributed systems, neural network architectures, high-dimensional data pipelines, and thousands of lines of refactored code. They assume graduate school is just two years of grinding through advanced syntax and learning bleeding-edge tech stacks.</p><p><strong>But here’s the thing I’ve been sitting with lately and would want to teach:</strong></p><p>The code you write in graduate school is just the surface layer. The tech stack you master today will evolve, pivot, or become obsolete in five years anyway. It comes with less certainty and more novelty and ambiguity.</p><p>The real architecture built during a Master’s degree isn’t in the repository -it’s in your brain and memory. It’s your own ability to break down a huge problem, connect abstract concepts, and design to improve an existing approach and innovate. It’s the invisible set of non-technical life skills that quietly rewire how you think, decide, and navigate complex problems.</p><p>Here is what finishing an M.S. in Computer Science actually teaches you.</p><p>A Computer Science degree teaches you to take risks and analyse tradeoffs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X6z0o01lWWe_495bjQ2YDw.png" /></figure><h3>1. Comfort with Ambiguity: The “Figure It Out” Muscle</h3><p>In undergraduate courses, problems usually come with clean boundaries: clear inputs, expected outputs, and a teaching assistant holding office hours to guide you. Teachers have an answer book and grading criteria</p><p>Graduate work hands you a blank page and a vague thesis statement:</p><p><em>“Here is an unsolved bottleneck. Figure it out.”</em></p><p>When you’re dropped into research or open-ended projects, there is no step-by-step map. You learn to sit with the discomfort of not knowing the immediate answer.</p><ul><li><strong>The Life Skill:</strong> You build a high tolerance for uncertainty. Whether you are navigating a career transition, launching a project, or making a personal pivot, you stop waiting for perfect instructions. You learn to state your hypotheses, run small experiments, and build forward momentum in the dark.</li><li><strong>Over time, this mindset cultivates patience for a long-term vision.</strong> You become comfortable investing effort today without needing immediate rewards tomorrow. Your brain adapts to think in longer time horizons, allowing you to stay with difficult problems for hours, weeks, or even months. As a result, your persistence strengthens, your decision-making becomes more resilient, and the pull of short-term gratification gradually loses its grip. It is like a Ladder to climb.</li></ul><h3>2. Trade-Off Thinking over Perfectionism</h3><p>In computer science, there is rarely a “perfect” system — only a series of deliberate trade-offs. Refinements occur, but there is a lot of persuasion involved. Some ideas are good on paper but don’t work in the real world. Some projects are projects but not a product yet.</p><p>In my opinion, most products start with being a project with a long-term vision. You stop working on that project, and then years later, you realise you learnt a new concept somewhere which could be an addition to your existing work that you have already built.</p><p>There are decision-making strategies that CS teaches you.</p><p>In any life decision or even while designing systems, should you optimize for memory or execution speed? Do you prioritize immediate simplicity or long-term scalability? You cannot have 100% efficiency across every parameter simultaneously. So, knowing the prioritization is important.</p><pre>[ Speed / Throughput ]<br>               /  \<br>              /    \<br>             /      \<br>[ Low Latency ] ---- [ Memory / Simplicity ]</pre><ul><li><strong>The Life Skill:</strong> You break out of binary, “all-or-nothing” thinking. In real life, every choice comes with an opportunity cost. Instead of agonizing over finding the single “correct” path, you learn to ask: <em>“What are my actual constraints, and which set of trade-offs am I willing to live with?”</em></li></ul><h3>3. Humility (Separating Knowing from Thinking You Know)</h3><p>There is a distinct moment in every Master’s program where your ego takes a hit. You walk into a seminar room, open a dense research paper, and realize you didn’t understand 80% of what you just read.</p><p>I have had a background in Electronics and Telecommunication in my Undergraduate Studies, but I was not confident in how CS works on top of Electronics — so I decided to pivot into CS. Electronics is way trickier and harder than CS. Both are so important to learn systems.</p><p>Graduate school forces you to confront the vast gap between <em>surface familiarity</em> and <em>true comprehension</em>.</p><ul><li><strong>The Life Skill:</strong> Intellectual humility. You stop pretending to understand things just to look smart. You want to dig deeper and ask ‘Why?’ — Some professors care about answering your questions. You have to be your own advocate and reach out to people who can answer your questions.</li><li>You learn to drop the defensive armor, ask sharper questions, and treat feedback as valuable data rather than a personal critique — Proactively.</li></ul><h3>4. Deep Focus and Cognitive Endurance</h3><p>We live in a world designed around short attention spans and rapid dopamine loops. An M.S. in CS demands the exact opposite: <strong>uninterrupted cognitive depth.</strong></p><p>Tracing a mathematical proof, isolating a silent concurrency bug, or reading academic papers requires hours of sustained, deep focus. It requires sitting in a chair with a single problem until the logic finally settles.</p><blockquote><em>That headache behind your eyes after hours of deep work isn’t weakness. It’s your neural architecture undergoing a forced remodel — at metabolic cost.</em></blockquote><ul><li><strong>The Life Skill:</strong> High focus capacity. The ability to lock in and engage in deep work for hours without reaching for your phone is becoming a modern superpower. Graduate school builds that focus muscle through pure repetition.</li><li>It’s not you who is going through it. All the graduate students near you are focused on making systems better and willing to be in this together.</li><li>The people you meet will build something for the world; it’s gonna be either improving a feature, building communities, or even a product — but everybody has the same aim to make XYZ better.</li><li>A CS degree teaches you to be open-minded, be open to ideas, take harsh constructive feedback and criticism, and makes you stronger. It teaches you step-by-step problem-solving and how to map real-world abstract ideas to a rapid product iteration.</li></ul><h3>5. Emotional Detachment from Sunk Costs</h3><p>Few things hurt like spending three weeks building an algorithm, running the benchmark evaluation, and realizing the core assumption was flawed.</p><p>You learn not everything anyone says about you is actually personal.</p><p>You stop taking things personally and see it practically.</p><p>In the early days, reaction is frustration. But over time, the mindset shifts:</p><ul><li><strong>Old framing:</strong> <em>“I just wasted three weeks of my life.”</em></li><li><strong>New framing:</strong> <em>“I successfully eliminated one invalid hypothesis. What does the data tell me to do next?”</em></li><li><strong>The Life Skill:</strong> Immunity to the sunk-cost fallacy. You learn to pivot quickly in life without taking failed attempts personally. You realize that a path that didn’t work out wasn’t “wasted time” — it was simply necessary data just to make a better decision ahead.</li></ul><h3>What Remains When the Code Changes</h3><p>If you look at my own journey — from studying electronics engineering to finishing a Master’s in Computer Science at Purdue — the most valuable things I brought with me weren’t just the algorithms or the frameworks.</p><p>It was the realization that <strong>we are built by the constraints we overcome and our skills are built on top of someone else’s systems.</strong></p><p>The degrees belong to us, but the real reward is the mental framework that remains after the terminal is closed: the patience to dissect a messy problem, the courage to embrace uncertainty, and the quiet confidence that no matter how complex the system is, you can break it down, understand it, and build something meaningful for the next generation. 🌸</p><p>~ Shruti Mandaokar (Feel free to read my other articles. ❤)</p><p>#ComputerScience #GraduateSchool #PersonalGrowth #LifeSkills #TechCareers #MastersInCS #MentalModels #SoftwareEngineering #StudentLife #TechMindset #Education #SelfImprovement #Technology #CareerAdvice</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=09196bf13d0a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Coderbyte #2: Bracket Combinations Challenge in Python]]></title>
            <link>https://medium.com/@shruti.mandaokar/coderbyte-2-bracket-combinations-challenge-in-python-43ab4aad2213?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/43ab4aad2213</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 20:11:27 GMT</pubDate>
            <atom:updated>2026-07-20T20:11:27.464Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FUwruVjXgvUCQeBA2ez_3w.png" /></figure><p>If you’ve dipped your toes into coding interviews or data structures, you’ve likely run into the ultimate algorithmic buzzword: <strong>Backtracking</strong>.</p><p>But in reality? Backtracking is just a fancy computer science term for playing <em>“Hot or Cold,”</em> or aggressively using the Ctrl + Z undo button on your life choices. May I go back and fix this?</p><p>One of the absolute best ways to understand this without your brain melting is a classic puzzle: <strong>Generating Valid Parentheses.</strong> Let’s break it down using plain English, zero pretension, and a little bit of Python.</p><p>Question : <a href="https://coderbyte.com/editor/Bracket%20Combinations:Python3">https://coderbyte.com/editor/Bracket%20Combinations:Python3</a></p><h3>The Ultimate Bracket Matching Game</h3><p>Imagine you are an inspector at a factory that builds bridges out of brackets. Your boss gives you a number, let’s say num = 3, and tells you to build every possible combination of 3 pairs of parentheses.</p><p>But there’s a catch: they must be <strong>well-formed</strong> (balanced).</p><ul><li>((()))? Chef’s kiss. Perfect. 🧑‍🍳</li><li>()()()? Classic. Clean.</li><li>)(())(? Absolutely illegal. Straight to coding jail. 🚨</li></ul><p>If you try to brute-force this by generating <em>every</em> possible combination of 6 brackets and checking them later, your computer’s fan will start sounding like a jet engine. Instead, we use backtracking to build them block-by-block while following two very simple rules.</p><p><em># Input : num = 3 means 3 opening brackets + 3 closing brackets</em></p><p><em># Pair 1 : () () ()</em></p><p><em># Pair 2 : () ( () )</em></p><p><em># Pair 3 : ( () ) ()</em></p><p><em># Pair 4 : ( ( () ) )</em></p><p><em># Pair 5 : ( () () )</em></p><p><em># Input: num = 2 means 2 opening and 2 closing brackets</em></p><p><em># Pair 1 : () ()</em></p><p><em># Pair 2 : ( () )</em></p><h3><em># Intuition: Backtracking?</em></h3><ol><li><em>Imagine a tree of choices. At the start, my string is empty “”.</em></li></ol><p><em>2. I can add ( -&gt; so now the string becomes string (</em></p><p><em>3. I can choose to add anothe opening or closing bracket ( or a )</em></p><p><em>4. The code explores one choice all the way to the end.</em></p><p><em>5. Once it hits a dead end or a success, it backtracks (steps backward) to the previous choice and tries the other option.</em></p><h3>The Two Golden Rules of the Factory</h3><p>To keep our brackets from breaking the laws of physics, we only need two gatekeepers:</p><ol><li><strong>The Opener Rule:</strong> You can always throw down an opening bracket ( as long as you haven&#39;t run out of them (i.e., you&#39;ve used less than num).</li><li><strong>The Closer Rule:</strong> You can <em>only</em> use a closing bracket ) if you currently have <strong>more</strong> opening brackets than closing brackets in your hand. Otherwise, you’re creating an orphan closer like)(, and we don&#39;t do that here.</li></ol><h3>Final Solution in Python</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XDOMcgTQ6MxxY3l6z7NbSA.png" /></figure><h3>What on Earth is (current + &quot;(&quot;, open_count + 1, close_count)?</h3><p>This line looks like a math textbook sneezed on your screen, but it’s actually just a <strong>data snapshot</strong> being handed off to the next step.</p><p>Think of it like a relay race where you pass a sticky note to your teammate:</p><ul><li><strong>current + &quot;(&quot;</strong>: You take the text you’ve built so far and slap a ( onto the end. If your string was&quot;(()&quot;, it now becomes &quot;(()(&quot;.</li><li><strong>open_count + 1</strong>: You whisper to your teammate, <em>&quot;Hey, I just used another opener, so the tally went up!&quot;</em></li><li><strong>close_count</strong>: You leave this number exactly as it was, because you didn&#39;t touch a closing bracket in this step.</li></ul><blockquote>Backtracking isn’t just a tool to survive technical interviews; it’s a superpower for whenever you need to explore a maze of choices (like solving a Sudoku puzzle or navigating a literal labyrinth). Write out your rules, keep track of your counters, and let recursion do the exhausting work for you.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*adl8U0C2tBYaegrFQ5Jd4g.png" /></figure><p>~ Shruti Mandaokar (Thanks for Reading :) )</p><p>I’m literally using two brackets haha!</p><h4>#Python #Programming #Algorithms #SoftwareEngineering #CodingInterviews</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43ab4aad2213" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Coderbyte #1: Codeland Username Validation Challenge in Python]]></title>
            <link>https://medium.com/@shruti.mandaokar/coderbyte-1-codeland-username-validation-challenge-in-python-927bde88c450?source=rss-8994c596ef6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/927bde88c450</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Shruti Mandaokar]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 19:20:42 GMT</pubDate>
            <atom:updated>2026-07-20T19:20:42.207Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SHRmwfXQRxZXO2SKRsyp-g.png" /></figure><p>I started preparation for my Tech Interviews. One of the first coding challenges I encountered on platforms like Coderbyte is <strong>Codeland Username Validation</strong>. Although it looks simple, it tests several important Python concepts in a single problem: string manipulation, regular expressions, conditional statements, and built-in string methods.</p><p>Let’s solve it step by step.</p><h3>Problem Statement</h3><p>Write a function that returns &quot;true&quot; if a username is valid; otherwise, return &quot;false&quot;.</p><p>A valid username must satisfy all of the following conditions:</p><ul><li>Be between <strong>4 and 25 characters</strong> long.</li><li>Start with a <strong>letter</strong>.</li><li>Contain only <strong>letters, numbers, and underscores (</strong><strong>_)</strong>.</li><li><strong>Not end</strong> with an underscore.</li></ul><h3>Understanding the Requirements</h3><p>Instead of writing code immediately, let’s translate each requirement and condition individually first into Python. Always break the problem into smaller steps first.</p><h3>1. Check the Length</h3><p>The username should have at least 4 characters and at most 25. We can use a comparison parameter here.</p><pre>4 &lt;= len(strParam) &lt;= 25</pre><p>Python allows chained comparisons, making this very readable.</p><h3>2. Check the First Character</h3><p>The username must begin with a letter.</p><p>Python provides the built-in function isalpha(), which is used to check whether a string consists <strong>only of alphabetical characters</strong>. It returns a Boolean value: True if all characters in the string are letters, and False otherwise</p><pre>strParam[0].isalpha()</pre><p>Examples:</p><pre>&quot;Shruti&quot;.isalpha()   # True <br>&quot;A&quot;.isalpha()        # True<br>&quot;9&quot;.isalpha()        # False<br>&quot;_&quot;.isalpha()        # False</pre><h3>3. Allow Only Letters, Numbers, and Underscores</h3><p>This is where <strong>regular expressions (regex)</strong> become useful.</p><p>The Pattern which is perfect here :</p><pre>^[A-Za-z0-9_]+$</pre><p>Let’s break it down. Don’t be scared — it is not that hard! I am saying identify capital letters A-Z, small letters a-z, numbers 0–9, and underscore _</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qs0hk9gZz_vZNdSSOcFrBA.png" /><figcaption>Regex Pattern Recognition</figcaption></figure><p>This ensures that <strong>every character</strong> in the username is in the allowed set.</p><p>Example:</p><pre>import re<br><br>pattern = &quot;^[A-Za-z0-9_]+$&quot;<br>bool(re.match(pattern, &quot;Shruti_99&quot;))</pre><p>Output:</p><pre>True # as Shruti_99 has letters, digit or underscore</pre><h3>4. Ensure the Username Doesn’t End with “_”</h3><p>Python lets us access the last character using negative indexing. The last element is always accessed with -1. Pretty cool, right :)</p><pre>strParam[-1] != &quot;_&quot;</pre><p>Examples:</p><pre>&quot;Shruti_&quot;</pre><p>Returns:</p><pre>False # Because at the end there is a underscore</pre><p>while</p><pre>&quot;Shruti99&quot;</pre><p>Returns:</p><pre>True # Because no underscore, right?</pre><h3>Final Solution: Joining the Pieces</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7ej3KZPIRZscc_0oS-kMPg.png" /></figure><h3>Time and Space Complexity</h3><p>Each validation step scans the username at most once.</p><ul><li>Length check: <strong>O(1)</strong></li><li>First character check: <strong>O(1)</strong></li><li>Last character check: <strong>O(1)</strong></li><li>Regular expression match: <strong>O(n)</strong></li></ul><p>Overall Time Complexity: <strong>O(n)</strong></p><p>where n is the length of the username.</p><p>Overall Space Complexity: <strong>O(1)</strong></p><h3>Key Takeaways</h3><p>This challenge is a great introduction to combining multiple Python concepts in one solution.</p><p>By solving it, you practice:</p><ul><li>String indexing</li><li>Conditional logic</li><li>Regular expressions</li><li>Boolean expressions</li><li>Built-in string methods like isalpha()</li></ul><p>While the problem itself is straightforward, the problem-solving method uses breaking a problem into small validation rules and checking each one systematically, and it is a technique you’ll use repeatedly in technical interviews and real-world software development.</p><p>Small validation problems like this build the foundation for larger systems such as form validation, authentication, and input sanitization.</p><p>Do you know problem-solving will keep your brain healthy! :)</p><p>~ Shruti Mandaokar</p><p>#Python #PythonProgramming #CodingChallenge #Coderbyte #ProblemSolving #SoftwareEngineering #LearnToCode #Programming #PythonTips #RegularExpressions #Regex #Algorithms #CodingInterview #TechEducation # #Developer #ComputerScience #WomenWhoCode #CodeNewbie #TechCommunity #ProgrammingJourney</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=927bde88c450" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>