<?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 Rishabh on Medium]]></title>
        <description><![CDATA[Stories by Rishabh on Medium]]></description>
        <link>https://medium.com/@rishabhbarnwal?source=rss-17e30ed92c36------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*lkEuHkrU9LYZNwjp_94HOQ.jpeg</url>
            <title>Stories by Rishabh on Medium</title>
            <link>https://medium.com/@rishabhbarnwal?source=rss-17e30ed92c36------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 08:21:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rishabhbarnwal/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 Do You Build a Score That a Model Cannot Game?]]></title>
            <link>https://rishabhbarnwal.medium.com/how-do-you-build-a-score-that-a-model-cannot-game-2925ddaa8296?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/2925ddaa8296</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[reward-hacking]]></category>
            <category><![CDATA[hacking]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 13:01:55 GMT</pubDate>
            <atom:updated>2026-11-19T13:01:55.612Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QdllHOLrIL_gDT253oXL3Q.jpeg" /></figure><p>In my last 28 hours of learning about reinforcement learning (RL), I came across a concept that helped me name something I previously only felt. It is called <strong>reward hacking: a </strong>model finds a way to get a high score without doing what we actually wanted.</p><p>The question that stayed with me is simple to ask and hard to answer: <strong>How can we build a score that a model cannot game, while still accurately measuring whether it achieves the goal?</strong></p><p>This article follows that question step by step, using real systems and numbers from published papers. We first start with how RL works. Then where the score comes from, how it gets gamed, how teams catch it, how the checker is tested, and what it means for someone who manages AI data work.</p><p>I have tried to explain every term when it first appears, and sources are listed at the end.</p><h3>The loop: attempt, score, adjust, repeat</h3><p>RL is a way of training in which a model learns by trying and receiving a score for each attempt. The cycle has four steps.</p><ol><li>The model attempts something</li><li>A scorer rates the attempt</li><li>The model adjusts toward whatever scored well</li><li>and the cycle repeats.</li></ol><p><em>A model that chooses actions in this way is often called an ‘agent’.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RA9vyuWqQ9k5AYvHOejwlg.jpeg" /></figure><p><strong>Example:</strong> In 2022, a research team working with Trane Technologies, a building management system provider, published a paper on using RL to control the cooling of commercial buildings. The agent set targets for equipment in the ‘chiller plant’ (the machine that produces cold water for a building) every 5 minutes. Its score was the negative of the total energy the chiller plant used, <strong>so less energy meant a higher score</strong>.</p><p>The team compared it with the existing controllers, which follow rules of thumb written by engineers (heuristics), in an ‘A/B test’, where two versions run side by side. Each test ran for 3 months, at two sites. The RL system saved about 9% and 13% energy.</p><p>Now think about what a score on energy alone would reward: a plant that stops cooling <strong>would get</strong> the highest reward. The paper does not rely on the score alone, and the goal also included keeping the air temperature and humidity in a desired range for the people in the building. So the agent works within ‘constraints’, which were hard limits on its behaviour.</p><p><em>Hence, this means that whoever designs the scoring system ultimately influences what the model learns.</em></p><h3>Where does the Score come from?</h3><p>In the cooling system case, an energy meter gave the score. For language models, there is no meter for ‘helpful’ or ‘correct’, so teams build the score in one of two common ways.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fx6hs7mS9WRBmMyqajqNiA.jpeg" /></figure><p><strong>RLHF: RL + Human Feedback.</strong> People rank the answers a model gives, and the model is tuned toward the answers people prefer. Because people cannot rank every attempt during training, a ‘reward model’ is trained on the rankings first. It is a second model that learns to predict which answer <strong>people would prefer</strong>, and it then gives the score.</p><p>The <strong>InstructGPT paper</strong> used this approach. Its authors report that outputs from their 1.3B parameter model were preferred to outputs from the 175B parameter GPT-3, despite having 100 times fewer parameters. (Parameters are the adjustable numbers inside a model, usually larger model = more parameters).</p><p><strong>RLVR: RL + Verifiable Rewards.</strong> Here, a ‘verifier’ (which is a program) checks the result with fixed rules. Code either passes its tests or it does not. A math answer either matches or it does not.</p><p><strong>DeepSeek-R1-Zero</strong> used a rule-based reward with two parts: accuracy rewards, which check whether the answer is correct, and format rewards, which require the reasoning to be written inside set tags. During training, the model’s average pass@1 score on AIME 2024 rose from 15.6% to 77.9%. (AIME is a United States high school mathematics competition. Pass@1 is the average share of problems solved on the first try).</p><p>The difference between the two is where the <strong>judgement</strong> sits. In RLHF, it sits in a learned model of human preference. In RLVR, it sits in a program. Each can fail in its own way, which we will discuss later.</p><h3>How does the model update?</h3><p>Once every attempt has a score, one more question remains: how should the model change? There needs to be a relative number to tell what is better and what is worse, right?</p><p>Two common methods are PPO and GRPO.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PmELfI5GCX1DGYmnh4CPuQ.jpeg" /></figure><p>PPO (proximal policy optimization) alternates between collecting attempts and training on them, and it is designed to keep each model update controlled. <br>It uses a ‘critic’, a second model that predicts the score (expected score) for an attempt. <strong>The gap</strong> between the <strong>actual score</strong> and that <strong>prediction</strong> is called the ‘advantage’, which means how much better or worse the attempt did than expected. A positive advantage pushes the model toward that attempt, and a negative one pushes it away.</p><p>GRPO (group relative policy optimization), in the authors’ words, “foregoes the critic model, instead estimating the baseline from group scores”. <br>The model writes several answers to the same question, and each answer gets a score. The average score of the group becomes the ‘baseline’. An answer above the group average gets pushed up, and an answer below gets pushed down. <strong>With no critic to train, it uses less memory</strong>.</p><p>One consequence is worth noting. If every answer in a group gets the same score, the advantage is zero for all of them, and that group teaches the model nothing. A task that is always passed, or never passed, gives no signal. That matters for the next section, because <strong>someone has to build tasks that are neither too easy nor too hard</strong>. This is why the task builder matters.</p><p>At the end of the day, whichever method is used, <strong>the model only follows the score.</strong></p><h3>Where does the model practice?</h3><p>For the model to practice, it needs a place where it can <strong>act</strong> and <strong>be</strong> <strong>scored</strong>. That place has three parts.</p><ul><li><strong>Environment:</strong> the place where the model acts (a sandbox). It holds the task and the ‘state’, which is the current condition of everything the model can change.</li><li><strong>Harness:</strong> the software that connects the model to the environment. It gives the model access to the controls, runs the attempt, and records the result.</li><li><strong>Verifier:</strong> the program that checks the result and produces the score.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UwGidkxNKhameJ7QVmWKHA.jpeg" /></figure><p><strong>Example:</strong> SWE-bench (published in 2023) is a test for AI coding assistants. It uses 2,294 real problems from 12 popular software projects. Each problem started as a real request on GitHub. Developers had reported that something was broken, and the team had later fixed it.</p><p>The model gets the same job:</p><ol><li>read the bug report,</li><li>look at the code,</li><li>and change the code so that the bug is fixed.</li></ol><p>Here is how the practice setup looks in this case (mapping each term with the example):</p><ul><li>Environment: a copy of the software project, exactly as it was when the bug was reported (before the fix).</li><li>Harness: the program that takes the model’s output (change), applies it to that copy, and runs the checks.</li><li>Verifier: the checks themselves, called ‘unit tests’. Each one ends in pass or fail (a unit test is a small program that checks one thing the software should do, e.g., whether saving a file keeps its contents).</li></ul><p><strong>The model counts as successful only if all the checks pass</strong>. Two kinds of checks matter:</p><ol><li>The checks that failed before the real fix must now pass, which shows the bug is gone.</li><li>The older checks must keep passing, which shows the fix did not break something else.</li></ol><p>One more detail shows how much care this takes. A problem was kept only if at least one check turned from fail to pass after the real fix. If every check already passes before the fix, even a model that changes nothing would get full marks. So that problem cannot be scored.</p><p>But what if the score itself has a flaw?</p><h3><strong>What is reward hacking?</strong></h3><p>When a model finds a way to get a high score without doing what the score was meant to measure. The model is not being clever on purpose. It is doing exactly what it was trained to do: <strong>raise the score.</strong></p><p><strong>Example 1:</strong> CoastRunners (OpenAI, 2016), a boat racing game. The score comes from hitting targets along the track, but the intent is to finish the race with the highest score.<br>The agent (the AI playing the game) found an isolated sector where 3 targets keep reappearing. It circled there, hit them again and again, and caught fire and crashed. <strong>It still scored about 20% higher than human players</strong>, but it never finished the race.</p><p><strong>Example 2:</strong> coding agents (OpenAI, 2025). Researchers watched coding models that were rewarded when unit tests passed. The models found shortcuts:</p><ol><li>Exit the program early, so the tests never get to fail.</li><li>Change the code itself that the tests depend on, so the tests pass without real work.</li><li>Replace the checking function with one that always says “true”.</li></ol><p>Every one of these gets a high score. None of them fixes the bug (similar to shortcuts students sometimes try).</p><p>This pattern has a name. Goodhart’s law (in Marilyn Strathern’s wording):<br>“When a measure becomes a target, it ceases to be a good measure”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NhxEacXUsYd4Ami8HIJREQ.jpeg" /></figure><h3><strong>Where can you catch it?</strong></h3><p>There are three places to look.</p><ol><li><strong>Output:</strong> check the final result + actions the model took. A monitor (a second AI that reads what the model did) that checked results + actions caught 60% of the hacks in the 2025 study.</li><li><strong>Reasoning + Output:</strong> also read the model’s step-by-step thinking (chain of thought). The same kind of monitor + reasoning check caught 95% of hacks.</li><li><strong>Independent check:</strong> a separate measure of the real goal that is never used in training. The model never gets to learn it, so it does not share the score’s blind spots.</li></ol><p>Why does the Independent check matter?</p><p>A study from 2022 tested this with two scorers. Picture each one as a list (example) of how judges would rate answers.</p><ul><li><strong>List 1 (the “proxy”):</strong> an approximate score built from a limited set of human ratings. The model is trained on this list. It has blind spots.</li><li><strong>List 2 (the “gold”):</strong> a stronger score with fewer flaws, standing in for real human judgment. The model was never trained on this list. It was only used to measure the output.</li></ul><p>As the model pushed its List 1 score higher, the List 2 score first rose, then fell. The model had found blind spots in List 1 and aimed at them. List 1 rewarded those answers, but List 2 did not.</p><p><strong>Another caution:</strong> the 2025 study also put the reasoning monitor inside the reward. With light pressure, hacking dropped. But with heavy pressure, the models still hacked and also started hiding their intent in the reasoning.</p><p>The researchers call the cost of keeping reasoning readable the “monitorability tax”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IjqC-FUp9sQh0hUTL_CAdw.jpeg" /></figure><h3>Who checks the checker?</h3><p>This heading is derived from ‘Quis custodiet ipsos custodes?’ (Juvenal), or ‘Who guards the guards?’, a very famous line used in many socio-political discussions.</p><p>Till now, we have been trusting our verifiers to give correct scores, and catching reward hacks showed us where their blind spots were. But what if the verifier itself is wrong? It can be too lenient and pass a wrong answer, or too strict and fail a correct one. Either way, the model learns from wrong scores, and nobody notices. <strong>So the checker needs to be tested before it is trusted.</strong></p><p>There are three simple tests:</p><ol><li>Give it an answer that is known to be right. It must pass.</li><li>Give it an answer that is known to be wrong. It must fail.</li><li>Ask a human expert to review a sample of its decisions, and compare. The expert’s view is the reference (“the gold”). The model never sees this during training.</li></ol><p>The first two use a “control set”, which is a small set of answers where the correct result is already known (known passes and failures).</p><p><strong>Example:</strong> SWE-bench Verified. Remember the SWE-bench problems from earlier? In 2024, OpenAI asked 93 experienced Python developers to review 1,699 random problems from it. The reviewers were not testing the models. They were testing the problems and their checks.</p><p>This is what they found:</p><ol><li>38.3% of the problems had an unclear bug report. Even a skilled developer could not tell what exactly was expected.</li><li><strong>61.1% had checks that could mark a correct fix as wrong.</strong></li><li>Some problems had both issues.</li></ol><p>Overall, 68.3% were filtered out. 500 problems were kept and released as SWE-bench Verified. <br>Why does this matter? A check that rejects a correct fix punishes the model for being right (a correct answer marked wrong). The model then learns to avoid good behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y6xqkngeE4UAX8opizfJog.jpeg" /></figure><h3>What does this mean for project managers?</h3><p>A model improves by hill-climbing, which means it makes small changes that raise its score, step by step. But that only works if the score is honest. If the score has gaps, the model climbs toward the gaps.</p><p>In the AI teams I know, the project manager is the single point of contact (SPOC) for everyone around the work: Product, Engineering, Quality Analysts, Operations Managers, Sales, Business Analysts and, finally, Clients. So questions about the score reach this person from every side. They do not need to write the checker. <strong>They need to know what to ask about it.</strong></p><p><strong>Here are five questions (according to me) to ask about any score:</strong></p><ol><li>What is the real goal (human intention), and what is the score measuring instead (proxy)?<br>(in CoastRunners, the goal was to win the race, and the score was hitting targets)</li><li>Where can the model score high without reaching the goal? <br>(the exit-early trick to not fail the test cases)</li><li>Which of the three places would catch cheating here: the reasoning, the output, or both?</li><li>Is there an independent check that the model never trains on?</li><li>Who tested the checker, and how?<br>(known right answers pass, known wrong answers fail, and an expert reviews a sample)</li></ol><p>So, how do you build a score that a model cannot game?</p><p>Honestly, you cannot make one that is perfect. What you can do is keep asking these questions, keep catching the gaps, and keep testing the checker. <strong>A score that is questioned regularly stays useful for much longer than one that is trusted blindly.</strong></p><h3><strong>Sources:</strong></h3><ol><li>Jack Clark and Dario Amodei, <a href="https://openai.com/index/faulty-reward-functions/">“Faulty reward functions in the wild”</a>, OpenAI, 21 December 2016.</li><li>Jerry Luo et al., <a href="https://arxiv.org/abs/2211.07357">“Controlling Commercial Cooling Systems Using Reinforcement Learning”</a>, 2022.</li><li>Long Ouyang et al., <a href="https://arxiv.org/abs/2203.02155">“Training language models to follow instructions with human feedback”</a> (InstructGPT), 2022.</li><li>DeepSeek-AI, <a href="https://arxiv.org/abs/2501.12948">“DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning”</a>, 2025. Figures are from version 2 (4 January 2026).</li><li>John Schulman et al., <a href="https://arxiv.org/abs/1707.06347">“Proximal Policy Optimization Algorithms”</a>, 2017.</li><li>Zhihong Shao et al., <a href="https://arxiv.org/abs/2402.03300">“DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models”</a>, 2024 (introduces GRPO).</li><li>Carlos Jimenez et al., <a href="https://arxiv.org/abs/2310.06770">“SWE-bench: Can Language Models Resolve Real-World GitHub Issues?”</a>, 2023.</li><li>Bowen Baker et al., <a href="https://arxiv.org/abs/2503.11926">“Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation”</a>, OpenAI, 2025.</li><li>Leo Gao, John Schulman and Jacob Hilton, <a href="https://arxiv.org/abs/2210.10760">“Scaling Laws for Reward Model Overoptimization”</a>, 2022.</li><li>Marilyn Strathern, “‘Improving ratings’: audit in the British University system”, <em>European Review</em>, 1997 (the wording of Goodhart’s law used in this article). Background: <a href="https://en.wikipedia.org/wiki/Goodhart%27s_law">Goodhart’s law</a>, Wikipedia.</li><li>OpenAI, <a href="https://openai.com/index/introducing-swe-bench-verified/">“Introducing SWE-bench Verified”</a>, 2024.</li><li>Juvenal, <em>Satires</em>, VI (“Quis custodiet ipsos custodes?”).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2925ddaa8296" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What should you ask before trusting an AV validation metric?]]></title>
            <link>https://rishabhbarnwal.medium.com/what-should-you-ask-before-trusting-an-av-validation-metric-f0d7bdf7a336?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0d7bdf7a336</guid>
            <category><![CDATA[autonomous-vehicles]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 10:10:11 GMT</pubDate>
            <atom:updated>2026-11-01T10:12:43.767Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What a 2019 driving dataset and a simulation study taught me while exploring multimodal data for Physical AI</em></p><p>Physical AI systems that sense the real world and then act in it start with <strong>perception</strong>. I was exploring multimodal sensor data (cameras, LiDAR, radar) and wanted to understand how models behave against annotated datasets. So I went back to a 2019 paper, <a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Caesar_nuScenes_A_Multimodal_Dataset_for_Autonomous_Driving_CVPR_2020_paper.pdf">nuScenes</a>.</p><p>I did not start with the question in the title. I came for the sensors, and realised <strong>scoring was the bigger problem</strong>.</p><h3>Why nuScenes?</h3><p>Per its authors, it was the first public driving dataset with not just cameras and LiDAR but also RADAR (and it included night and rain scenes).</p><p>Quick sense of the three sensors, as the paper describes them:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qJwd5WjZ_9fC8d6a8VpJRw.png" /><figcaption>(here the LiDAR was not FMCW)</figcaption></figure><h3>“Correct” is a choice, and it affects the ranking</h3><p>A detector draws a box around each object. The score compares those boxes with human annotated boxes. So you need a rule for what counts as a match.</p><ul><li>The common rule is box overlap (called <strong>IoU</strong>, intersection over union).</li><li>nuScenes matches detections by the <strong>distance between box centers</strong> instead.</li></ul><p>Why it matters: small objects like pedestrians and bikes can get zero overlap from a small position error, and camera methods tend to be less precise on localisation. Center distance is fairer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pJfa32lG-7nVg3-TJ5uH5Q.png" /></figure><h3>Data volume can change the ranking</h3><p>With a small amount of training data, a LiDAR and a camera-based method looked alike. With more data, the LiDAR method pulled ahead. A ranking on a small dataset may not hold on a bigger one.</p><h3>Time and speed are in the benchmark</h3><ul><li><strong>Time:</strong> nuScenes tracks how long a model takes to first detect an object and the longest gap between two detections.</li><li><strong>Speed needs history:</strong> nuScenes lets a model look back 0.5 seconds (10 LiDAR sweeps). Feeding <a href="https://docs.nvidia.com/tao/archive/5.3.0/text/point_cloud/pointpillars.html">PointPillars</a> those past sweeps cut its speed error to about a quarter.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KMRRhiOi_X8jCAkOj-HUaA.png" /></figure><h3>How the combined score (NDS) works</h3><p>Half of the nuScenes detection score (NDS) is detection: did it find the object?</p><p>The other half is five quality checks, 10% each: where it is, how big, which way it faces, how fast it moves, and its attributes. Tracking over time isn’t in NDS at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9rzZ0QlSBxh3nCtnBIJfyw.png" /></figure><h3>A good score isn’t a safe car</h3><p>Next, I searched for simulation studies that referenced nuScenes and found <a href="https://ijcai.org/proceedings/2020/0483.pdf">Modeling Perception Errors towards Robust Decision Making in Autonomous Vehicles</a>.</p><p>The authors <strong>swapped real perception for an error model inside a driving simulator</strong> to test where standard metrics fall short.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j6x9XNDjFxzDo1fnvZkjWA.png" /></figure><ul><li>More detections helped only up to about a 75% detection rate. How long the car was blind mattered far more.</li><li>With a jaywalker and perfect detection in every frame, repeated track loss meant no speed estimate and no braking. Position error size barely mattered.</li></ul><p>*These are preliminary results from a simulation with hand-made error models and one driving system, so I read them as a warning.</p><p>The authors do suggest nuScenes’ distance threshold. <strong>But a score alone can’t show how the car responds to the detections.</strong></p><h3>Still pondering</h3><ul><li><strong>Where did the RADAR go?</strong> In nuScenes, radar-only wasn’t promising. But radar measures speed directly, and speed is what the car couldn’t estimate in the jaywalker test. My hunch: worth testing radar as a speed backup when tracking fails because neither paper tested that.</li><li><strong>Is there LiDAR that measures speed directly (FMCW)?</strong> Next goal.</li><li><strong>Do these questions hold beyond driving, for other Physical AI systems?</strong> I would like to find out.</li></ul><h3>Finally, before you trust a metric, ask</h3><ol><li>What counts as a “correct” detection, and at what threshold?</li><li>Is it one metric or a composite, and can I see the parts?</li><li>What’s in the data (place, weather, rare objects)?</li><li>What does it prioritise (detection over tracking, semantics over localisation)?</li></ol><h3>My take</h3><p>A headline detection score can look great while track loss goes unmeasured because it only tells you how a model did on a benchmark. Not what an AV will do next.</p><p><strong>Sources</strong></p><ul><li>Caesar et al., <em>nuScenes: A multimodal dataset for autonomous driving</em>, CVPR 2020. [<a href="https://openaccess.thecvf.com/content_CVPR_2020/papers/Caesar_nuScenes_A_Multimodal_Dataset_for_Autonomous_Driving_CVPR_2020_paper.pdf">link</a>]</li><li>Piazzoni et al., <em>Modeling Perception Errors towards Robust Decision Making in Autonomous Vehicles</em>, IJCAI-20, pp. 3494–3500. [<a href="https://ijcai.org/proceedings/2020/0483.pdf">link</a>]</li></ul><p>(all images/screenshots are generated using Claude)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0d7bdf7a336" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I built a newspaper for myself]]></title>
            <link>https://rishabhbarnwal.medium.com/i-built-a-newspaper-for-myself-f13de5329676?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/f13de5329676</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 13:25:44 GMT</pubDate>
            <atom:updated>2026-10-11T15:51:38.699Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A small daily tool that reads the news for me, keeps only what’s mine, and never shows me the same story twice.</em></p><p>There’s a colleague of mine who always seems to know where our industry is heading, and what our clients care about. They’ve simply been around both much longer than I have.</p><p>That made me ask a question: <strong>do I need to spend all those years too, or can I get there faster by being more disciplined?</strong></p><p>Part of the answer was something I’d noticed but never named. What makes the work yours is understanding the people on the other side of it. Nobody taught me that in a meeting; I caught it from the people around me. And that kind of understanding comes from a habit, like reading a morning paper: a few sections you always read, a little every day, and slowly you build the context that people around you already have.</p><p><strong>I wanted that habit, so I built it into a paper.</strong></p><p>This post is about how it works, what broke along the way, and where it’s going next. It stays in plain English, and the technical parts are marked so you can skip them.</p><p><strong>How it works</strong></p><p>Every morning I press one button. A few minutes later, I have a short front page with hyperlinked headlines and summaries. Here’s what happens in between:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ca1XnKJ1-i15EGx1bqF8rA.png" /></figure><ol><li><strong>It goes looking.</strong> A news search service is asked the same handful of personalised questions every day: who raised money, who partnered with whom, what new thing was built, what research came out, what rules are changing. Those questions live in a plain settings file that I write by hand, so I decide what the paper is about.</li><li><strong>It reads.</strong> For every story it finds, an AI works out who it’s about, which section it belongs in, and how important it is to someone following this field. It also writes a two-line summary.</li><li><strong>It cuts.</strong> Anything below my bar is dropped, and each section is capped at five stories. On a normal day, around 170 stories could go in, and fewer than 20 might come out.</li><li><strong>It hands me the page.</strong> What’s filtered till here lands on a small review page. I tick what’s useful, cross what isn’t, and skip what doesn’t matter.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bae_j_YP3OUN2MazOZXVfw.png" /></figure><p><strong>Then something quietly broke</strong></p><p>I allocate about thirty minutes a day for reading a newspaper. I skim the headlines, cherry-pick a few, and dive in.</p><p>One day I realised some stories I would have picked weren’t there. Good ones, too! The kind that should have scored higher than what I was looking at. There was no error and no warning, just a page that was thinner than it should have been.</p><p>So I dug into the saved runs, and the numbers spoke.</p><p><strong>What I found:</strong></p><ul><li>On busy days, hundreds of stories went to the AI in one big pile</li><li>In one run, <strong>609 stories went in, and 45 came back.</strong> The rest simply vanished/dropped</li></ul><p><strong>What I changed:</strong></p><ol><li>Smaller piles: 25 stories at a time.</li><li>A counter that checks every story that went in has come back with a verdict</li><li>If any are missing, it retires those stories once more</li></ol><p>After the fix, a 100-story test returned <strong>100 out of 100</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1e7tciZla_PqKph-qxV4ng.png" /></figure><p>And it turned out there was a second thing the page was breaking at.</p><p><strong>The system kept repeating itself</strong></p><p>The second problem was easier to see once I looked for it. The page read like a stuck record.</p><p>For example: When a startup raises money, ten outlets write about it, each with a slightly different headline. To a search tool, those are ten different articles. To me, they’re one story, told ten times. The tool also returns the same link again and again, because different searches find the same page (overlaps).</p><p><strong>I implemented three gates, cheapest first</strong>:</p><ol><li><strong>Same link?</strong> Already seen, so it’s dropped. Free.</li><li><strong>Same headline?</strong> Another outlet used the same words, so it’s dropped. Free.</li><li><strong>Told in new words?</strong> Same event written differently; the AI that’s already reading the article spots it in the same pass while scoring. No extra wait.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YTW_m1gwU8c8zKycWWufrA.png" /></figure><p><strong>Where does the memory live?</strong> The AI doesn’t carry any of this. It’s a new call each time, and a simple database does, for free. The AI only gets a short list of recent headlines to spot retellings, and that list stays the same size. So the paper can remember a year of stories without the AI getting slower or more expensive. I wanted the AI biased by only one thing, my settings, and not by a pile of old conversation.</p><p>But where is all this leading to?</p><p><strong>The part that makes it mine!</strong></p><p>Everything so far is about getting a <em>good</em> paper. This part is about getting <em>my </em>paper.</p><p>Every day, the review page asks for three small things:</p><ol><li>a tick for <strong>useful</strong></li><li>a cross for <strong>not for me</strong></li><li>a skip for <strong>no view</strong></li></ol><p>It takes seconds, but here is the start of a record of what I actually care about. I could have borrowed a ready-made dataset of “good articles”, but it would know what’s popular, not what’s useful to me. For that, the dataset has to come from me.</p><p>Today, my ratings are only stored. They don’t change what appears on the page.</p><p><strong>What I’m building next</strong></p><p>This is a work in progress, and I use it every day. After a few weeks, that record will be big enough to say something. The next idea is to let it show me patterns:</p><ul><li>which topics keep earning my tick,</li><li>and which searches keep bringing things I cross out</li></ul><p>Then it can <em>suggest</em> changes to my settings, while I stay in charge of what actually changes. I don’t know yet how well it will work, which is why I’m calling it a work in progress.</p><p>One thing I’ve already learned: missing a day doesn’t matter; it’s about the habit that survives.</p><p>So, a question for you. What are you building that helps you grow <em>outside</em> your job description? I would love to hear it in the comments.</p><p>Thank you for reading.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f13de5329676" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 7 am Baingan ka Bharta That Made me miss the Gym]]></title>
            <link>https://rishabhbarnwal.medium.com/the-7-am-baingan-ka-bharta-that-made-me-miss-the-gym-dac2eae3d63d?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/dac2eae3d63d</guid>
            <category><![CDATA[homecooking]]></category>
            <category><![CDATA[cooking]]></category>
            <category><![CDATA[baingan-ka-bharta]]></category>
            <category><![CDATA[indian-recipe]]></category>
            <category><![CDATA[recipes-for-cooking]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Tue, 17 Mar 2026 13:14:17 GMT</pubDate>
            <atom:updated>2026-03-17T13:14:58.503Z</atom:updated>
            <content:encoded><![CDATA[<p>I am not the person who wakes up and decides to cook. Especially at 7 am in the morning. But today was odd, I slept early last night for no reason, woke up without an alarm, and suddenly I had time. Proper time. Not something I found between things.</p><p>And I had a Baingan (Eggplant, Aubergine) sitting in the fridge. I had picked it up from the market 2 days ago, and today felt like the right time. I didn’t think too much cause I had confidence in my skills here :)</p><p>Well, not just skills but my roots too. I am from Bihar. This dish has always been a favourite of mine for its unique texture and taste. This was home food for me. You don’t plan home food, you just cook it when you feel like it.</p><p>I knew I was getting late for the Gym when I started preparing.</p><p>Still did it!</p><p>Serves — 2</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-uBYEt0ZLBr17h0ia4SNow.jpeg" /></figure><h3><strong>Ingredients</strong></h3><ol><li>Baingan (a big one)</li><li>Tomato (whole) — 2 Nos</li><li>Onion (large) — 1 Nos</li><li>Garlic — 5/6 Cloves</li><li>Green Chillies — 1/2 Nos</li><li>Jeera (Cumin) — 1 Tsp</li><li>Red Chilli Powder — 1 Tsp</li><li>Haldi (Turmeric) — 0.5 Tsp</li><li>Garam Masala — 2 Tsp</li><li>Coriander powder — 1 Tsp</li><li>Hing — a pinch</li><li>Salt — as per mood</li><li>Lemon — 1 (0.5 sliced + 0.5 juice)</li><li>Coriander — handful</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HFaIblOh9haguXydvag6IA.jpeg" /></figure><h3>The Journey (aka don’t rush this)</h3><p><strong>Step 1: Roast the World</strong></p><p>Slice your Baingan from 4 directions. Don’t cut it open, just make slits to create pockets. Similarly, on tomatoes. Put a steel grill on a flame that will support your vegetables while roasting. Add whole Garlic that will be used later. Unlike other ingredients here, which will take time to cook, you need to remove the garlic as soon as it starts giving off a strong smell.</p><p>For Baingan and Tomato, once the skin starts coming off and the insides seem softer, it’s done.</p><p><strong>Step 2: The chop chop part</strong></p><p>Let these roasted ingredients cool, and trust the chop-chop process. Dice your onions and garlic, slice your chillies, and chop your coriander. Keep them aside.</p><p><strong>Step 3: The annoying part</strong></p><p>Unless you are faster, this time would be enough to cool your roasted vegetables so they can be peeled by hand.</p><blockquote>Pro tip: Use a bowl of water to keep the juices away that make your fingers slippery while peeling the burnt part of the vegetable skin.</blockquote><p>It takes time. You don’t have to skin 99.999% of it. 99% is also fine. At the end of the day, it was an edible part just roasted to hell.</p><p><strong>Step 4: Build the base</strong></p><p>Heat ~1 Tbsp of Mustard Oil (any other oil is betrayal) on low-medium heat. As it starts to smoke, add the Jeera. Then add your chillies. Then add your onions and garlic. And keep tossing your vegetables so they don’t stick.</p><p>Once the onions and garlic turn Golden Brown, add in your roasted tomato and lower the heat so the juices don’t stick and burn. Mash the tomatoes, let the juice out, and once that’s done, add in your spices. Keep stirring until it dries up.</p><p>Add salt as per taste :)</p><p>Add 0.5 Cups of water to make a slurry. Add 2–3 slices of lemon. Let it dry up on medium to high flame.</p><p><strong>Step 5: Bring it together</strong></p><p>Remove the lemon slices as the slurry dries.</p><p>Lower the flame and add the Baingan. Mash it completely and add 0.5 cups of water and flame to high. Let the entire slurry dry and become sticky.</p><p>Once the Bharta gets sticky, lower the flame and add coriander and lime juice. If you know how to do Chop Chop, Fold Fold, use that here. Or just keep stirring the thing around. Usually, at this step, the salt can be off sometimes, so add a pinch while you Chop Chop, Fold Fold.</p><h3>What I Learned</h3><p>Well, it came out exactly how it should. I used comparatively less amount of oil this time. Usually, the dish would take at least 2–3 Tbsp of Mustard oil.</p><p>But it had been a while since I COOKED cooked, and that too this early in the morning. I got late to the Gym but totally worth it, or I would have regretted it later.</p><p>But to you, as a reader, if you need a challenge here, this one is no Joke for sure. It starts from selecting the right Baingan to roasting it right, to adding the right amount of water. It’s all about the texture; don’t make it too dry, don’t keep it slurry, like a curry. Just enough to make it all sticky and mushy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dac2eae3d63d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Amritsari Paneer Bhurji: The Date Night Dish That Actually Worked.]]></title>
            <link>https://rishabhbarnwal.medium.com/amritsari-paneer-bhurji-the-date-night-dish-that-actually-worked-f88e3206bc38?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/f88e3206bc38</guid>
            <category><![CDATA[recipes-for-cooking]]></category>
            <category><![CDATA[indian-recipe]]></category>
            <category><![CDATA[healthy-lifestyle]]></category>
            <category><![CDATA[healthy-eating]]></category>
            <category><![CDATA[cooking]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Tue, 14 Oct 2025 07:49:37 GMT</pubDate>
            <atom:updated>2025-10-14T07:49:37.782Z</atom:updated>
            <content:encoded><![CDATA[<p>This is a rich recipe that combines the flavours of North India and, well, paneer. The spices and fats used to create this dish are what make it authentically Amritsari, and when we talk Amritsari, we absolutely cannot back down on the butter or cream. No apologies, no substitutions, no “can we use olive oil instead?” This is romance, not a calorie-counting competition.</p><p>Serves — 2…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SWj2nAwsR5dQZUxpsbxlUw.jpeg" /><figcaption>You will need more Breads 😆</figcaption></figure><h3>Ingredients</h3><ol><li>Paneeeeer (Grated) — 200g</li><li>Onion (chopped) — 2 Nos</li><li>Tomato (Diced) — 4 Nos</li><li>Besan (Gram Flour) — 1 Tbsp</li><li>Coriander — 🤔</li><li>Cashews— 6/8 Nos</li><li>Dahi (Curd) — 1 cup</li><li>Green Chillies — 2 Nos</li><li>Red chilli powder, garam masala, coriander powder — 1 Tbsp each (for base)</li><li>Dry mango powder, black pepper, garam masala — 1 Tsp each (for final curry)</li><li>Turmeric — 1 Tsp</li><li>Cream — 0.5 cup</li></ol><h3>The Journey (aka How to Not Mess This Up)</h3><p><strong>Step 1: Tomato Purée — The Foundation of Everything Good</strong></p><p>Pan-toss your diced tomatoes with <strong>1 slit green chili</strong> in moderate oil and a pinch of salt. Once the tomatoes have turned slightly mushy and soft (much like your hopeless heart), remove them from the pan. Let them cool for 2–3 minutes, then blend into a smooth purée.</p><blockquote>Pro tip: Don’t skip the cooling time. I learned this the hard way when my blender nearly staged a mutiny.</blockquote><p><strong>Step 2: Cashew Paste — Creamy Magic</strong></p><p>Create a paste of cashews with 1–2 teaspoons of water. This is what gives the dish its luxurious, velvety texture. Think of it as the smooth-talking wingman of your ingredients.</p><p><strong>Step 3: The Curd Base — Where Chemistry Happens<br></strong>This is where things get interesting (and slightly nerve-wracking).</p><p>Mixture 1: In 1 cup of curd, whisk together:</p><ul><li>All Spices + Salt</li><li>Your cashew paste</li></ul><p>Mixture 2: Heat butter in a pan and sauté peanuts (if using). Once the butter simmers, lower the heat and add gram flour, stirring <em>constantly</em>. This is crucial; gram flour burns faster than your chances with bad pickup lines. Turn off the heat when the mixture changes color. Cool for 2–3 minutes.</p><p>Combine: Whisk Mixture 1 and Mixture 2 together in a bowl. This is the heart of your dish.</p><p><strong>Step 4: The Grand Finale — Bringing it All Together</strong></p><ul><li>Heat a pan with butter and oil (yes, we’re not holding back tonight).​​</li><li>Add 1 green chili and onions. Sauté until golden and fragrant.​</li><li>Add your tomato purée. Lower the heat to medium as it starts thickening.​</li><li>Add 1 tsp dry mango powder, 1 tsp garam masala, and 1 tsp black pepper powder. Combine and add 1/2 cup of cream.​</li><li>Now for the moment of truth: Add your curd base and keep stirring until everything is combined.</li></ul><blockquote>Important note from someone who’s been there: If the curd base sticks to the side walls of your pan, happy cleaning. Better get your heat low or your hands faster. I managed to avoid this catastrophe by keeping the flame low and stirring like my romantic future depended on it (because it kind of did).</blockquote><ul><li>Once everything is beautifully combined, lower the heat and add your grated paneer. This is also when you’d add coriander if you want, but I skipped it and it turned out great.​</li><li>Toast your bread. Yes, it really only needs one bullet point.</li></ul><h3>What I Learned</h3><p>Cooking for someone you care about is terrifying and wonderful in equal measure. The Amritsari Paneer Bhurji doesn’t just taste amazing, it tells a story. A story about taking a chance, putting in effort, and creating something special for someone special.​</p><p>The butter-rich, spice-forward flavors of North India combined with the creamy texture of paneer and the tangy kick from the curd base create something that’s comfort food and celebration all at once. It’s the kind of dish that says, “I care enough to spend time in the kitchen for you, but I’m also confident enough to serve you something bold and flavorful ” :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f88e3206bc38" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Basil Pesto Pasta: The Remedy I Didn’t Ask For (But Desperately Needed)]]></title>
            <link>https://rishabhbarnwal.medium.com/basil-pesto-pasta-the-remedy-i-didnt-ask-for-but-desperately-needed-a2ac2463016b?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/a2ac2463016b</guid>
            <category><![CDATA[food]]></category>
            <category><![CDATA[italian-food]]></category>
            <category><![CDATA[healthy-foods]]></category>
            <category><![CDATA[healthy-eating]]></category>
            <category><![CDATA[pasta]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Tue, 14 Oct 2025 06:34:46 GMT</pubDate>
            <atom:updated>2025-10-14T06:34:46.601Z</atom:updated>
            <content:encoded><![CDATA[<p>I didn’t plan to become someone who makes pesto from scratch. That’s for people with spice racks that aren’t just oregano and chili flakes bought during a moment of ambition in 2022. But here I am, grinding basil at 11 PM on a Tuesday :)</p><p>This recipe is a flavourful and aromatic dish that brings fresh herbs to your plate — made with pesto sauce of basil, garlic, parmesan, almonds, and olive oil that blends perfectly with penne pasta. Fun fact: pesto doesn’t mean “paste,” and it’s not necessarily associated with basil. Pesto means “to pound” which feels oddly appropriate for someone who’s been emotionally pounded lately.</p><p>Serves — 2 or just you, twice. No judgment!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k2xAydBIlXADmMHUuE7Xog.png" /></figure><h3>Ingredients</h3><ol><li>Penne Pasta — 200g</li><li>Basil— 50g</li><li>Garlic (I personally like it a little extra) — 4/5 Nos</li><li>Milk (optional) — 2 Tbsp</li><li>Cream (optional) — 1 Tbsp</li><li>Cheese (Parmesan) — 1 Cup</li><li>Walnuts — 5/6 Nos</li><li>Olive Oil — 4/5 Tbsp</li><li>Oregano, Chilli Flakes, Black Pepper— 1/2 Tsp</li></ol><h3>The Therapy Session (aka Preparation)</h3><p><strong>Step 1: Boil Water</strong></p><p>There’s no <em>al dente</em> nonsense here. Cook your pasta till it’s easy to bite but not so soft it breaks, kind of like maintaining emotional boundaries. Drain and set aside.</p><p><strong>Step 2: Make the Pesto (The Real Star)</strong></p><p>This is where the magic happens, and by magic, I mean aggressively grinding ingredients because your gym membership expired and you need an outlet.​</p><ul><li>Grind the dry stuff first: nuts and garlic. Pretend they’re your problems.</li><li>Add<strong> 25g basil</strong> with <strong>2 tbsp</strong> olive oil. Blend like your life depends on it.</li><li>Add the next <strong>25g basil</strong> and <strong>2 tbsp</strong> olive oil. Blend again. You’re doing great.</li><li>Toss in your parmesan. Blend.</li><li>Season with salt and black pepper. Blend one last time.</li><li>Keep it chunky. If the leaves are still too big, add another tbsp of olive oil. If it’s not coming together (much like your last relationship), add <strong>1–2 tbsp</strong> milk/cream and blend again.</li></ul><p><strong>Step 3: Combine Everything (The Grand Finale)</strong></p><p>1 Tbsp olive oil in a pan. Keep the heat low; you don’t want to cook the pesto, just warm it. This isn’t a metaphor; it’ll actually ruin the flavor. Mix your pasta into the sauce. Adjust with salt, oregano, or chili flakes depending on how spicy you’re feeling about life today.</p><h3>Final Thoughts From Your Friendly Neighborhood Engineer-Turned-Reluctant-Chef</h3><p>This recipe is proof that sometimes the best things come from moments of mild desperation and a well-stocked pantry. It’s fresh, it’s inviting, it’s simple enough for weeknights yet impressive enough that you can pretend you’re a functioning adult.​</p><p>The beauty of cooking is that, unlike engineering problems or relationships, you can always try again tomorrow. Plus, carbs are scientifically proven to improve mood.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a2ac2463016b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Broccoli Cheddar Stew with Cherry Tomato Dressing: An Indo-Italian Fusion for Comfort]]></title>
            <link>https://rishabhbarnwal.medium.com/broccoli-cheddar-stew-with-cherry-tomato-dressing-an-indo-italian-fusion-for-comfort-263e8eecfd45?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/263e8eecfd45</guid>
            <category><![CDATA[healthy-foods]]></category>
            <category><![CDATA[food]]></category>
            <category><![CDATA[recipes-for-cooking]]></category>
            <category><![CDATA[cooking]]></category>
            <category><![CDATA[healthy-eating]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Tue, 14 Oct 2025 05:40:39 GMT</pubDate>
            <atom:updated>2025-10-14T05:42:55.276Z</atom:updated>
            <content:encoded><![CDATA[<p>There’s something magical about bringing together classic comfort foods and bold new flavors. This recipe is a fusion of American broccoli cheddar stew, Italian-seasoned cherry tomato dressing, and an unmistakable Indian twist. It’s inspired by experimentation and a love for breaking culinary boundaries.</p><p>Serves — 2 (and tastes romantic)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zHmhi0F1BqI9HxGPK9ie4A.jpeg" /><figcaption>I swear the hunger was real</figcaption></figure><pre><br>.-------------------------------------.----------------.<br>| Ingredient                          |   Amount       |<br>:-------------------------------------+----------------:<br>| Broccoli                            |   100 gm       |<br>:-------------------------------------+----------------:<br>| Cheddar Cheese                      |   1/2 cup      |<br>:-------------------------------------+----------------:<br>| Cherry Tomatoes                     |   10           |<br>:-------------------------------------+----------------:<br>| Cashew                              |   6–8          |<br>:-------------------------------------+----------------:<br>| Red Chilli                          |   2            |<br>:-------------------------------------+----------------:<br>| Peas                                |   1/3 cup      |<br>:-------------------------------------+----------------:<br>| Paneer                              |   200 gm       |<br>:-------------------------------------+----------------:<br>| Scallions                           |   3            |<br>:-------------------------------------+----------------:<br>| Oregano, Chilli Flakes              |   3 tbsp each  |<br>:-------------------------------------+----------------:<br>| Curd                                |   1/2 cup      |<br>:-------------------------------------+----------------:<br>| Brown Bread                         |   8 slices     |<br>:-------------------------------------+----------------:<br>| Salt, Black Pepper, Garlic Powder   |   1/2 tsp each |<br>:-------------------------------------+----------------:<br>| Garlic Butter                       |   2 tbsp       |<br>&#39;-------------------------------------&#39;----------------&#39;</pre><h3>Step-by-Step Preparation</h3><h3>Preparing the Base</h3><p>Add paneer cubes, curd, red chillies, black pepper, garlic powder, cashews, cheddar cheese, and oregano into a mixer. Blend until you achieve a smooth, stew-like consistency.​</p><h3>Preparing Broccoli and Greens</h3><p>Cut broccoli into pieces and boil with peas for 10–12 minutes. Chop scallions and set aside. Sauté broccoli and peas with scallions, salt, and oregano in 1 tablespoon butter.​</p><h3>Cherry Tomato Dressing</h3><p>Slice all cherry tomatoes. In a pan, cook them with butter, chilli flakes, and oregano (1 tablespoon each). Don’t let the tomatoes fry; gently heat until soft and juicy, then let cool.​</p><h3>Toasted Bread &amp; Final Assembly</h3><p>Toast the brown bread. In serving bowls, pour the blended base, top with cherry tomatoes (including juices!), and finish with sautéed greens. Serve alongside crispy toast for the complete experience.​</p><h3>Personal Tips &amp; Story</h3><p>The real trick is the blend — experimentation is key! It took several tries to get the base just right. Feel free to play with sweetness, and crank up spices to match your preferences. This recipe is all about adaptation, so don’t hesitate to make it your own.​</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=263e8eecfd45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Garlic Herbed Rice with Creamy Pesto Sauce]]></title>
            <link>https://rishabhbarnwal.medium.com/garlic-herbed-rice-with-creamy-pesto-sauce-c3017b4fd235?source=rss-17e30ed92c36------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3017b4fd235</guid>
            <category><![CDATA[healthy-lifestyle]]></category>
            <category><![CDATA[cooking]]></category>
            <category><![CDATA[healthy-foods]]></category>
            <category><![CDATA[food]]></category>
            <category><![CDATA[italian]]></category>
            <dc:creator><![CDATA[Rishabh]]></dc:creator>
            <pubDate>Sun, 25 May 2025 09:56:36 GMT</pubDate>
            <atom:updated>2025-05-25T10:40:20.505Z</atom:updated>
            <content:encoded><![CDATA[<p>Honestly, garlic rice sounded like that kind of thing you make when your fridge is judging you. I almost went the lemon-coriander route and wrapped it up in 20 minutes, but come on, it’s Sunday, not a sad Wednesday.</p><p>So I called in the big guns: pesto, milk, and cream. Suddenly, it’s giving me the impression that I know what I’m doing, even if I do not ;)</p><p>Serves — 3 (the third wheel shouldn’t stay hungry, right?)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YQ9NaYOcq2GEHzqD_2aZsg.jpeg" /></figure><h3><strong>Ingredients</strong></h3><ol><li>Paneeeer (Diced) — 200g</li><li>Rice — 3 Cups</li><li>Garlic (sliced) — 10 Nos</li><li>Coriander — 🤔</li><li>Coriander stem (diced) — 5/6 Nos</li><li>Capsicum (Diced) — 2 Nos (Yellow and Red)</li><li>Walnuts — 3/4 Nos</li><li>Refined Flour — 3/4 Tbsp</li><li>Basil Leaves — 100 gm</li><li>Milk — 500 ml</li><li>Cream — 4/5 Tbsp</li><li>Sweet Corn (boiled) — 1 Cup</li><li>Spices/Herbs/Seasoning— Black pepper, Lemon Zest, Oregano, Chilli flakes, Red chilli powder, Salt</li></ol><h3>Preparation</h3><ol><li><strong>Rice:</strong> 3 Cups of Rice with 4 cups of water. Everything goes into the cooker and cooks for 2 whistles. Take it out, let it rest till the pressure is released.</li><li><strong>Pesto Sauce: <br></strong> — In a blender, combine fresh basil leaves, walnuts, and a pinch of salt. Blend until smooth, gradually adding olive oil to achieve the desired consistency.<br> — For the creamy twist, heat milk in a pan, add flour to make a roux, then mix in your pesto and cream until velvety.<br> — Once the cream starts “cream-ing”, add in the capsicums, corn, salt, and cook them under low heat, covered.</li><li><strong>Paneer: </strong>Heat oil in a pan on high flame and toss in the Paneer cubes. Don’t let them stay there for too long; once they get the desired colour, take the paneer out and dry the oil using tissues.</li><li><strong>Aromatisation of Rice:<br></strong> — Heat oil in a pan on high flame and toss in the garlic and coriander stems. <br> — Just when the garlic reaches the golden to brown stage, add in the rice and lower the flame. <br> — Add black pepper, oregano, chilli flakes, salt, and lemon zest.<br> — Add paneer on low heat, sprinkle some lime water to make the rice moist, and keep it covered on low flame for 5 minutes.</li><li><strong>Final: </strong>Serve it hot, don’t toss it all together. And let me know how your experience was :)</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c3017b4fd235" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>