[{"content":"","date":"24 May 2026","permalink":"/tags/ai/","section":"Tags","summary":"","title":"AI"},{"content":"","date":"24 May 2026","permalink":"/tags/ai-economics/","section":"Tags","summary":"","title":"AI Economics"},{"content":"","date":"24 May 2026","permalink":"/blog/","section":"Blog","summary":"","title":"Blog"},{"content":" I'm a product manager, but I started my career in cybersecurity. I worked as a security engineer on the front lines of detecting and responding to threats, but I found my real passion was in building the products, not just protecting them. That love for problem-solving and creating things people value is what pulled me into product management. My technical roots are still a big part of who I am. I\u0026rsquo;m an avid hobbyist programmer, and I love spending free time building small things just for the fun of seeing an idea come to life.\nCapital markets have fascinated me since school, sparked by reading business newspapers. At 20, I began investing, which has been a fascinating journey that has greatly enriched my perspective. It\u0026rsquo;s a constant lesson in patience and long-term thinking, teaching me how to navigate uncertainty, manage emotions, and separate signal from noise.\nTravel is another passion. For me, exploring new places, cuisines, and cultures is all about collecting stories. I’ve backpacked across India and internationally, trying my best to capture the landscapes and experiences I encounter along the way.\nOn the Bookshelf section, you\u0026rsquo;ll find some of my favorite reads. The blog is primarily my musings on tech, business, and life. ","date":"24 May 2026","permalink":"/","section":"Nishant Singh","summary":"I'm a product manager, but I started my career in cybersecurity.","title":"Nishant Singh"},{"content":"","date":"24 May 2026","permalink":"/tags/saas/","section":"Tags","summary":"","title":"SaaS"},{"content":"","date":"24 May 2026","permalink":"/tags/","section":"Tags","summary":"","title":"Tags"},{"content":"Satya Nadella said something that most SaaS founders probably don\u0026rsquo;t want to hear. Business applications, he argued, are at their core just CRUD databases — Create, Read, Update, Delete — with business logic layered on top. In an agentic world, that logic migrates entirely to the AI tier. Agents won\u0026rsquo;t care about your backend. They\u0026rsquo;ll update multiple databases simultaneously, orchestrate workflows across systems, and execute decisions. When the logic leaves the app, the product proposition leaves with it.\nhttps://youtu.be/9NtsnzRFJ_o?si=hx5Hl5wtiH5UHVsO\u0026t=2806\nThe public markets got there first. The SaaS premium, the elevated multiple investors assigned to software businesses on the logic that they scale to near-zero marginal cost has cratered. Companies that commanded 30x, 40x revenue multiples at the 2021 peak have been repriced dramatically. Some have shed 60–70% of their market cap from peak. This isn\u0026rsquo;t a macro correction. It\u0026rsquo;s the market pricing in exactly what Satya described.\nAnd it isn\u0026rsquo;t just happening to enterprise software companies. The same pattern is playing out at every layer of the stack, all the way down to a hobbyist who spent a weekend building an app and found the frontier lab had already shipped their feature as a footnote in a changelog. Different scales. Same wall.\nWhere does value actually accrue in AI? # Think about the AI stack in three layers: applications at the top, infrastructure and inference in the middle, semiconductors and frontier models at the base.\nI came across an article by Apoorv Agrawal, partner at Altimeter, that I\u0026rsquo;d genuinely recommend reading in full. It\u0026rsquo;s one of the clearest pieces I\u0026rsquo;ve found on where the money actually flows in generative AI, and a lot of what follows draws from his framing. You can find it here.\nSource: https://apoorv03.com/p/the-economics-of-generative-ai-two Source: https://apoorv03.com/p/the-economics-of-generative-ai-two The cloud comparison is useful but mainly because it shows exactly where AI diverges. Cloud followed classic economies of scale. More customers meant lower unit costs, wider moats, compounding margins. The value triangle flipped upward toward software over time. WS made cloud infrastructure cheap. A generation of software companies built on top and barely touched the hardware. A decade of premium SaaS multiples followed.\nAI hasn\u0026rsquo;t done this. Not yet and there\u0026rsquo;s a serious argument it won\u0026rsquo;t, at least not at the app layer.\nApoorv\u0026rsquo;s data makes this concrete. Over the last two years, roughly $350 billion of new AI revenue was added to the ecosystem. Around 75% of it went to semiconductors. The application layer grew more than 10x in that same period and still barely registered a change in the overall shape of the stack.\nThe triangle has not flipped. It has become more concentrated at the base. The margin data confirms it. Nvidia\u0026rsquo;s data centre gross margins are running around 75%. Estimated application-layer margins sit somewhere between 0% and 30%. That gap isn\u0026rsquo;t a timing lag. It reflects where the structural advantages actually live.\nGavin Baker said something on \u0026lsquo;Invest Like the Best\u0026rsquo; that genuinely shifted how I see this. Even accounting for Cursor and Cognition, the most successful AI-native app companies of this cycle, net value at the application layer has been destroyed by AI. The companies seeing the highest value creation share one trait: the highest ratio of utilized GPUs per human. That\u0026rsquo;s not a software metric. Profits are accruing to energy, data centres, chips, and frontier models. Not to the layer where most founders and product teams are spending their time.\nGavin also made an observation that stuck with me: in venture, the companies that survive are the ones building something not obvious to the world before they could build real scale, something genuinely different, genuinely hard to replicate. A lot of app-layer AI companies fail this test. They\u0026rsquo;re building obvious things in an era where obvious things get commoditized before they get traction.\nCursor has operated for years as a product company running on Anthropic\u0026rsquo;s and OpenAI\u0026rsquo;s models, paying frontier API rates while those same labs marketed directly competing tools to its customers. Claude Code crossed $2.5 billion in annualized revenue and over 300,000 business customers by early 2026. That\u0026rsquo;s the app layer trap in its most uncomfortable form: your supplier is also your competitor, and they have structurally lower costs than you. Cursor\u0026rsquo;s response was Composer, their own in-house coding model built on open-source base weights but with 85% of total compute spent on their own reinforcement learning pipeline and post-training stack. Composer 2.5, released this week, matches Claude Opus 4.7 on Cursor\u0026rsquo;s own benchmark at roughly one-tenth the per-token cost. There is no public API. The model runs inside Cursor only: the IDE, the CLI, the web product. That\u0026rsquo;s a deliberate choice. Cursor isn\u0026rsquo;t trying to become a model company. They\u0026rsquo;re trying to stop being a harness. The move down the stack, owning the model, owning the unit economics, locking the experience inside their own environment, is what fighting the app layer gravity actually looks like in practice.\nWhat this doesn\u0026rsquo;t mean # This isn\u0026rsquo;t an argument that the app layer is dead. It\u0026rsquo;s an argument about what survival requires. Cursor is real. Sixty-seven percent of Fortune 500 companies are customers.(Link to article). They were processing over a billion lines of code daily in 2025 (Aman Sanger - X). But notice what Cursor is doing to stay there: building proprietary models, internalising inference costs, locking the experience into a closed environment. They\u0026rsquo;re not resting on distribution. They\u0026rsquo;re actively moving down the stack.\nThat\u0026rsquo;s the distinction worth sitting with. The app layer companies that will survive aren\u0026rsquo;t the ones who built clean interfaces on top of frontier APIs and called it a product. They\u0026rsquo;re the ones who accumulated something genuinely hard to replicate: proprietary data, deep workflow lock-in, or like Cursor, the willingness to get into the infrastructure business themselves. Gavin Baker\u0026rsquo;s framing holds: the question isn\u0026rsquo;t whether your idea is good. It\u0026rsquo;s whether it\u0026rsquo;s different and hard enough that the model companies won\u0026rsquo;t get there before you\u0026rsquo;ve built real scale.\nMost won\u0026rsquo;t clear that bar. A few will. Knowing which side you\u0026rsquo;re on is the whole game.\n","date":"24 May 2026","permalink":"/blog/the-app-layer-is-a-trap/","section":"Blog","summary":"Satya Nadella said something that most SaaS founders probably don\u0026rsquo;t want to hear.","title":"The App layer is a trap"},{"content":"","date":"10 May 2026","permalink":"/tags/product/","section":"Tags","summary":"","title":"Product"},{"content":"","date":"10 May 2026","permalink":"/tags/vibe-coding/","section":"Tags","summary":"","title":"Vibe coding"},{"content":"There is a specific kind of weekend that a lot of people had in the last two years.\nYou open a conversation with a frontier model on a Friday night with a half-formed idea. By Sunday evening you have a working product. A Chrome extension. A finance tracker. A tool that solves something you personally found annoying. The code is cleaner than you expected. The UI is sharper than it has any right to be. You feel like you have crossed some threshold.\nI had several of those weekends. I built things I had no business shipping that fast. And for a while, that feeling of building felt like the point.\nI have had a fair amount of coding experience over the years, but I had never once built a Mac application. Then one Sunday, messing around with Codex, I built a portfolio tracker from scratch. It works across currencies. It pulls from stock exchanges globally. It handles corporate actions. It is private, runs locally, and does exactly what I needed it to do. I was genuinely amazed. That Sunday changed something in how I think about what is possible. Portfolio Tracker built with Codex. Sunday-ing with Codex Another weekend, I built a Firefox extension. It tracks which websites you visit, blocks the ones that pull your attention away, and keeps everything local — no API calls, no telemetry, nothing leaves the browser. The analytics surfaces which domains eat most of your time, broken down by category, across whatever window you want to look at: a day, a week, a month, up to a year. The blocking side has nuance to it. You can block any domain for a fixed window or indefinitely. Blocked pages show a countdown. If you genuinely need to get through, there is an exception mechanism — five per day per domain, each one granting a short timed window before the block returns. Over eighty domains come pre-categorised out of the box. Everything else falls back to keyword inference. It took a Sunday with Claude Code and it does exactly what no existing extension did in quite the way I wanted. Clauding Sundays hit different.\nA Firefox browser extension for tracking your browsing habits and blocking distracting websites. All data stays local; No external API calls, no telemetry. Built with Claude Code. A Firefox browser extension for tracking your browsing habits and blocking distracting websites. All data stays local; No external API calls, no telemetry. Built with Claude Code. And that is exactly the use case where vibe coding is extraordinary. Building for yourself. Scratching your own itch. Creating personal software that fits your life precisely, without compromise, without a subscription, without trusting some SaaS company with your financial data. For this, the tools are nothing short of remarkable and I am completely all in.\nIt is not the point.\nHere is what nobody tells you clearly enough going in: the bottleneck in software was never writing code. It was finding people who needed the thing, convincing them to try it, and giving them enough reason to come back. That bottleneck did not move. Vibe coding routed around the wrong obstacle entirely.\nThe moment you decide to market your product, something fundamental shifts. You are no longer building a product. You are building a business. And that is a completely different game with completely different rules. One that AI cannot shortcut for you. Distribution, positioning, trust, retention — none of that ships over a weekend.\nAnd the second problem is worse. The frontier labs are building the same things you are, just with more compute, more distribution, and a built-in user base of millions. You spend a weekend on a markdown editor. OpenAI ships it as a feature update. You build a grammar tool. It becomes a toggle inside an existing product someone already has open. By the time you are thinking about your first hundred users, the thing you built has already been absorbed into something bigger.\nTheir feature is your product.\nI am not saying do not build. The tools are genuinely extraordinary and the learning is real. But there is a difference between building to learn and building under the assumption that the product itself is the hard part.\nThe hard part is the same as it always was. Distribution. Trust. Accumulated data that nobody else has. A workflow so embedded that switching feels painful. Those things take longer than a weekend. They always did.\nVibe coding collapsed the cost of starting. It did not change what it takes to matter.\n","date":"10 May 2026","permalink":"/blog/vibecoding-is-the-easy-part/","section":"Blog","summary":"There is a specific kind of weekend that a lot of people had in the last two years.","title":"Vibe coding is the easy part"},{"content":"","date":"23 April 2026","permalink":"/tags/gemma4/","section":"Tags","summary":"","title":"Gemma4"},{"content":"","date":"23 April 2026","permalink":"/tags/llm/","section":"Tags","summary":"","title":"LLM"},{"content":"","date":"23 April 2026","permalink":"/tags/localai/","section":"Tags","summary":"","title":"LocalAI"},{"content":"","date":"23 April 2026","permalink":"/tags/ollama/","section":"Tags","summary":"","title":"Ollama"},{"content":"I\u0026rsquo;ve been using LLMs long enough to have opinions about prompting, to benchmark outputs, to argue about which model to pick for what task. And for a while, \u0026ldquo;free and open source\u0026rdquo; felt like the whole story. Download the model, run it, done.\nThen I actually did it.\nI installed Ollama, and pulled Gemma 4, Google\u0026rsquo;s latest open model released in April 2026, directly from my terminal. No API key. No cloud. Just a model running on my MacBook Pro M1.\nollama pull gemma4:26b ollama run gemma4:26b On Apple Silicon, Ollama automatically runs inference through Metal (the M1\u0026rsquo;s GPU framework) with zero configuration. The M1\u0026rsquo;s unified memory architecture means the CPU and GPU share the same memory pool, which is a big part of why modern MacBooks handle local models better than most people expect.\nThat was the crack in the black box. Everything else followed from there.\nWhat the model actually looks like up close # With a cloud API, the model is a name in a dropdown. Locally, you can inspect what you\u0026rsquo;re actually running. One command changed how I read model spec pages:\nollama show gemma4:26b This is what came back: Model architecture gemma4 parameters 25.8B context length 262144 embedding length 2816 quantization Q4_K_M Capabilities completion vision tools thinking Parameters temperature 1 top_k 64 top_p 0.95 A few things worth pausing on. The context length of 262144 is 256K tokens, enough to fit entire codebases or long documents in a single prompt. The capabilities list tells you what the model can actually do: not just text completion, but vision, tool use, and thinking. And the quantization label : Q4_K_M is the one that explains why a 25.8B parameter model downloaded as a 17GB file.\nAt full 16-bit precision, a 25.8B model would occupy over 50GB of memory. Quantization compresses the weights down to 4-bit, cutting that to 17GB. There\u0026rsquo;s a quality tradeoff, but it\u0026rsquo;s smaller than you\u0026rsquo;d expect and it\u0026rsquo;s what makes running a frontier-class model on a laptop viable at all.\nThe model you call through an API may also be quantized. You often can\u0026rsquo;t tell from the outside. That\u0026rsquo;s worth knowing when you\u0026rsquo;re comparing outputs or debugging unexpected behavior.\n\u0026ldquo;26 billion parameters\u0026rdquo; doesn\u0026rsquo;t mean what you think it means # The model I downloaded is called Gemma 4 26B. That number implies something heavier than what it is to run.\nGemma 4 26B is a Mixture of Experts model. Instead of one large network where every parameter fires on every input, it has 128 specialized sub-networks, experts, and routes each token through only 8 of them at a time. The result: roughly 3.8 billion active parameters per token, not 25.8 billion. Inference cost sits closer to a 4B model than a 26B one.\nThe catch that rarely gets mentioned: even though you only compute on 3.8B parameters at a time, all 25.8B have to be loaded and accessible in memory. Gemma 4 26B is cheaper to run than its name suggests, but not cheaper to fit. Compute cost and memory cost are different things — and conflating them is one of the most common mistakes in infrastructure planning for LLMs.\nThe full model name — gemma-4-26B-A4B — actually tells you this directly. The \u0026ldquo;A4B\u0026rdquo; stands for approximately 4 billion active parameters. The information is in the name. Most people don\u0026rsquo;t know to look for it.\nThis is why parameter count as a standalone benchmark claim is incomplete. Once you know to ask \u0026ldquo;active or total?\u0026rdquo;, you start finding that distinction absent almost everywhere it matters.\nThe model thinks out loud, if you let it # I ran a simple prompt to see what the model would do:\nollama run gemma4:26b \u0026#34;Explain yourself in one sentence\u0026#34; What appeared in the terminal wasn\u0026rsquo;t an immediate answer. It was this: Thinking... Goal: Explain myself (an AI model) in exactly one sentence. Audience: General user. Constraint: One sentence. ... Draft 1: I am a large language model, trained by Google... Draft 2: I am an AI trained by Google that can help you write... ...done thinking. I am a large language model, trained by Google, designed to assist you with a wide range of tasks by processing and generating human-like text. This is Gemma 4\u0026rsquo;s thinking mode — a reasoning capability where the model works through a problem step by step before producing an answer. The thinking tokens are internal scratchpad, not part of the final response. The model drafted multiple versions, evaluated them, and selected one.\nI could feel the two phases of inference # Running a model locally strips away the abstraction that cloud APIs add. You stop seeing a response and start seeing a process. And that process has two distinct phases that behave nothing like each other.\nBefore the model generates a single word of response, it reads your entire prompt in parallel. Every token at once. This is the prefill phase : fast, parallelizable, and relatively cheap. Then it switches modes. The response gets generated one token at a time, each one depending on all the tokens before it. This is the decode phase : sequential by nature, impossible to parallelize, and where almost all of your wait time lives.\nOn a cloud API these two phases collapse into one experience: you send a message, you wait, you get an answer. Locally, they\u0026rsquo;re separable. And Ollama lets you measure them directly.\nStart an interactive session and type /set verbose before sending a prompt:\nbash\nollama run gemma4:26b \u0026gt;\u0026gt;\u0026gt; /set verbose Set \u0026#39;verbose\u0026#39; mode. \u0026gt;\u0026gt;\u0026gt; Explain what a transformer architecture is in two paragraphs After the response, you\u0026rsquo;ll see something like this printed to your terminal:\nprompt eval count: 12 tokens prompt eval duration: 0.11s prompt eval rate: 109.1 tokens/s ← prefill speed eval count: 247 tokens eval duration: 18.2s eval rate: 13.6 tokens/s ← decode speed The gap between those two rates is the entire story. Prefill processes tokens in parallel so it\u0026rsquo;s fast regardless of prompt length. Decode is sequential, the rate you see there is roughly your ceiling for how fast this model can generate text on this machine, no matter how short your prompt is.\nTry it with two different prompts back to back. One that asks for a one-word answer. One that asks for a detailed explanation. Watch what changes: the prefill rate stays roughly the same (you\u0026rsquo;re reading a similar-length prompt both times). The decode rate stays roughly the same too. But the second response takes dramatically longer because you\u0026rsquo;re paying the decode cost per output token, and you asked for more of them.\nThis is why \u0026ldquo;why is it slow?\u0026rdquo; is usually the wrong question. The right questions are: how long is the output, and how many tokens per second can this hardware sustain during decode?\nWhat\u0026rsquo;s sitting in memory the whole time # While the model is running, open a second terminal window and run:\nbash\nollama ps You\u0026rsquo;ll see something like:\nNAME SIZE PROCESSOR UNTIL gemma4:26b 17 GB 100% GPU 4 minutes from now Two things worth noting here.\nThe 100% GPU confirms that Ollama is using Metal (the M1\u0026rsquo;s GPU framework) automatically. No configuration, no flags. The unified memory architecture on Apple Silicon means the CPU and GPU share the same physical memory pool, which is why a 17GB model fits and runs at all on a MacBook. On a machine with a discrete GPU, that 17GB would need to fit in dedicated VRAM, a much harder constraint.\nThe 17GB figure is the quantized model weights sitting in memory. They stay there as long as the session is active. But as your conversation grows, something else starts growing alongside it: the KV cache.\nDuring decode, the model needs to attend to every previous token to generate the next one. Recomputing that attention from scratch on every new token would be brutally slow, so instead it caches the intermediate computation, the key and value matrices, for every token it has already seen. That cache lives in memory, and it grows with every exchange.\nFor Gemma 4 26B with a 262,144-token context window, a full-context conversation could accumulate several additional gigabytes of KV cache on top of the model weights. You\u0026rsquo;re not just paying for the model, you\u0026rsquo;re paying for the memory of the conversation.\nThis is why long-context inference is expensive in ways that per-token pricing doesn\u0026rsquo;t fully capture. A short query on a fresh session and the same query after a 100,000-token conversation have very different memory footprints, even if the output is identical.\nOne laptop, one user and that\u0026rsquo;s already the limit # Everything above assumes something important: one person sending one request at a time.\nRunning a model for yourself on a laptop is a fun experiment. Trying to serve multiple users simultaneously introduces a whole new level of complexity. If you process requests sequentially, User #100 is going to give up long before their request reaches the front of the queue. The memory you\u0026rsquo;ve allocated to User #1\u0026rsquo;s KV cache is memory User #2 can\u0026rsquo;t use. And the decode bottleneck that felt manageable at 1x becomes the ceiling that breaks everything at 10x.\nHow production inference systems solve this — batching, continuous batching, KV cache eviction, inference servers — is a topic for the next piece.\n","date":"23 April 2026","permalink":"/blog/the-model-is-free-the-inference-isnt/","section":"Blog","summary":"I\u0026rsquo;ve been using LLMs long enough to have opinions about prompting, to benchmark outputs, to argue about which model to pick for what task.","title":"The model is free. The inference isn't"},{"content":"Read widely, think critically! # I read non-fiction books, mostly. This isn\u0026rsquo;t a list of every book I\u0026rsquo;ve read, but a curated collection of the ones that have had a lasting impact on my thinking.\nTitle Author Grit: The Power of Passion and Perseverance Angela Duckworth Linchpin Seth Godin No Rules Rules Reed Hastings \u0026amp; Erin Meyer One Up Wall on Wall Street Peter Lynch Poor Charlie’s Almanack Charlie Munger Sapiens Yuval Noah Harari Six Easy Pieces Richard Feynman That will never work Marc Randolph The Courage to be Disliked Ichiro Kishimi The Hard Thing about Hard Thing Ben Horowitz The Lean Product Playbook Dan Olsen The Mom Test Rob Fitzpatrick The Psychology of Money Morgan Housel The Wealth of Nations Adam Smith Thinking, Fast and Slow Daniel Kahneman This is Marketing Seth Godin Tuesday with Morie Mitch Albom Zen \u0026amp; The Art Of Motorcycle Maintenance Robert Pirsig ","date":"31 October 2024","permalink":"/bookshelf/","section":"Nishant Singh","summary":"Read widely, think critically!","title":"Bookshelf"},{"content":"","date":"2 February 2024","permalink":"/tags/culture/","section":"Tags","summary":"","title":"culture"},{"content":"It was a cold, rainy October morning at Haneda Airport. I was freezing, pulling my hoodie tight against the damp. I’d just landed and, being a total aviation geek, my first stop had to be the observation deck before I even thought about my hostel. I was watching the planes cutting through the rain on the tarmac, but that’s not what grabbed me. It was the Costa Coffee. There was a line, a steady stream of people just filing in, wrapping their cold hands around steaming cups.\nJAL aircraft as seen from observation deck at Tokyo Haneda Airport Coffee at every corner # My coffee revelation at Haneda was just the beginning. As I walked around Tokyo, I was stopping every few minutes, just stunned. There was coffee everywhere. Vending machines are on basically every street, lit up like silent robot baristas, offering every kind of coffee you can imagine—hot, cold, black, latte, sweet, not sweet. But the real shock? My first time in a Japanese convenience store. The konbini. These places (7-Eleven, Lawson, Family Mart) aren\u0026rsquo;t just for quick snacks; they\u0026rsquo;re coffee shrines. I’m talking entire aisles. Glass-door fridges packed with rows and rows of bottled and canned coffees, from simple black brews to seasonal flavours.\nIced black coffee from 7-eleven Like many, I had long imagined Japan as the quintessential tea-drinking nation, drawing comparisons to my home country India, where tea is ubiquitous. However, 14 days in, it was obvious I was wrong. Japan is just as obsessed with coffee, and it’s not a small thing, it’s deep and it’s everywhere. That was the first big shock of my trip.\nThe coffee scene in Japan runs far deeper than vending machines and konbini cups. You can\u0026rsquo;t walk two blocks without seeing a big chain: Starbucks, Tully’s, Duotor. These weren’t just occasional outposts, they were everywhere, and always busy. But the real magic happened when I started wandering off the main roads and found the smaller, artsy cafés.\nBefore visiting Japan, I had a completely misguided notion. I\u0026rsquo;d put Japan in the \u0026ldquo;tea-only\u0026rdquo; box. While the country\u0026rsquo;s tea traditions are indeed rich and enduring, my assumption about coffee couldn\u0026rsquo;t have been more wrong. This wasn\u0026rsquo;t just a recent trend or Western influence, Japan\u0026rsquo;s relationship with coffee runs surprisingly deep.\nTokyo cafe hopping: A celebration of aesthetics \u0026amp; coffee # Tokyo\u0026rsquo;s café scene is meticulous but feels effortless. I only had four days in the city and a massive list of places to visit. I embarked on a whirlwind coffee tour that took me to some of the most unique and captivating cafes in the city.\nFrom the minimalist charm of Little Nap to the industrial-chic vibes of Glitch, each space told its own story. Glitch is housed in a modern minimalist space with concrete floors, simple wooden furniture, and a prominent Probat roaster where they roast their beans in-house. The interior design reflects Japanese minimalism while maintaining an industrial coffee-roasting atmosphere. I was mesmerized watching the barista. His pour-over was perfect, all precise concentric circles. Pure craft.\nGlitch Coffee \u0026amp; Roasters Glitch Coffee \u0026amp; Roasters I visited Onibus, a converted traditional Japanese house turned café, where old architecture meets modern coffee culture. What makes this spot special is the second-floor seating area where customers can enjoy their coffee while watching trains glide past at eye level through the windows. This creates a uniquely Tokyo experience - the combination of traditional architecture, specialty coffee, and urban railway infrastructure.\nOnibus Fuglen brought a touch of Norwegian coffee culture to a quiet Shibuya backstreet, with its mid-century modern furniture and Nordic approach to light roasts. What makes Fuglen Tokyo particularly special is its dual identity: During the day, it operates as a serious specialty coffee shop; At night, it transforms into a cocktail bar, making it a rare \u0026ldquo;day-to-night\u0026rdquo; establishment in Tokyo\u0026rsquo;s coffee scene.\nAll Seasons Coffee proved to be a hidden gem in Shinjuku, where the seasonal menu changes four times a year to reflect nature’s rhythm. They carefully select and roast beans to complement different weather conditions and seasonal changes. The café\u0026rsquo;s minimalist interior design follows classic Japanese aesthetics, while their in-house roasting operation maintains a focus on medium roast profiles.\nHeart’s Light Coffee sits in a peaceful corner, tucked away from the area\u0026rsquo;s high-end boutiques and busy streets. They\u0026rsquo;re particularly known for their light roast approach, which aims to highlight the delicate, fruit-forward notes in their carefully sourced single-origin beans. As soon as you enter you get hit by an amazing smell of fresh coffee and cool jazzy tunes being played by an amazing sound system from the owner. What sets Heart\u0026rsquo;s Light apart is their distinctive menu that bridges Australian and Japanese coffee cultures. Their signature drink, the LSD (their take on Melbourne\u0026rsquo;s \u0026ldquo;magic\u0026rdquo; coffee), showcases this fusion perfectly.\nPaddlers is a significant player in Tokyo\u0026rsquo;s specialty coffee scene with its strong connection to Portland, Oregon\u0026rsquo;s coffee culture. While the shop has made its name serving exclusively Stumptown Coffee Roasters beans - a nod to its Portland connections - it\u0026rsquo;s the vintage analog sound system that truly sets it apart. The owners are passionate about both coffee and music, specifically jazz and soul, and have created an environment where these two cultures blend seamlessly.\nAfter hopping through Tokyo\u0026rsquo;s intimate coffee spots I ended up at an unexpected final stop: the massive Starbucks Reserve Roastery in Nakameguro. Sure, it may seem ironic to end a specialty coffee tour at a Starbucks, but this isn\u0026rsquo;t the typical chain outlet. The Starbucks Reserve Roastery in Tokyo stands as the coffee giant\u0026rsquo;s largest roastery in the world, spanning four floors across 32,000 square feet. It was designed by Japanese architect Kengo Kuma, featuring a stunning origami-inspired ceiling with over 2,100 wooden blocks in a twisting pattern that represents the unfolding of coffee\u0026rsquo;s aroma. Beyond coffee, the Roastery houses Japan\u0026rsquo;s first Princi Italian bakery, a Teavana tea room offering innovative tea cocktails, and the exclusive Arriviamo cocktail bar where coffee and spirits merge. Each floor offers distinct experiences, from watching the coffee roasting process to enjoying coffee-infused cocktails, making it more than just a café.\nStarbucks Reserve Roastery in Tokyo What struck me most was the way each café carved out its own distinct personality while maintaining that unmistakably Japanese attention to detail. The real magic wasn’t in any single element, but in the way everything came together seamlessly. Looking back, my café-hopping adventure barely scratched the surface of what Tokyo\u0026rsquo;s coffee scene has to offer, but it left me with a deep appreciation for how this city manages to make every cup feel like a unique artwork.\nHistorical roots # How did coffee even get this big in Japan? It turns out it goes way back. The story starts in the 17th century with the Dutch, who first brought it to Nagasaki during the Edo period. Back then, Japan was totally closed off to the world (the Sakoku policy), so coffee was a rare commodity reserved for governors, translators, and few businessmen.\nWhen Japan finally opened up in the Meiji period (late 1800s), more people could get it, but it was still expensive, mostly for the upper class. The first real coffeehouse, Kahiichakan, opened in Tokyo in 1888. It got people curious, but it failed pretty quickly. Coffee finally got a real foothold in the 1910s and 20s, thanks to a place called Café Paulista that served affordable Brazilian coffee.\nIn 1933, Tadao Ueshima, the “Father of Coffee in Japan,” launched his company, Ueshima Tadao Shoten, which eventually revolutionized Japan\u0026rsquo;s coffee industry by developing canned coffee in the 1960s. World War II had temporarily halted Japan’s coffee imports, but by the 1960s, coffee culture surged back, fuelled by an increasing fascination with Western lifestyle and products.\nGreen Tea \u0026amp; Coffee: A harmonious coexistence # Interestingly, the rise of coffee hasn\u0026rsquo;t diminished Japan\u0026rsquo;s love for green tea. If anything, it has created a beautiful coexistence. In Japan\u0026rsquo;s diverse beverage market, coffee has carved out a remarkable position, standing second only to green tea among non-alcoholic drinks.\nEverywhere I turned during my stay, I encountered the familiar sight of green tea bottles. Every convenience store\u0026rsquo;s refrigerated section was dominated by an impressive array of tea brands – Ito En, Suntory, and Kirin leading the charge with their signature products like Oi Ocha and Ayataka. RTD green tea seamlessly bridges traditional tea culture with contemporary convenience. The Japanese have managed to preserve the essence of their tea heritage while making it instantly accessible. Whether it\u0026rsquo;s businesspeople grabbing a bottle for their commute or tourists like me needing refreshment, RTD green tea serves as the go-to thirst quencher.\nOi Ocha Green Tea Matcha is still culturally vital, but it\u0026rsquo;s also modern. The matcha latte is a staple in every café.\nMatcha Latte Japan’s coffee culture is a perfect mix of tradition and innovation. From vending machine cans to meticulously crafted pour-overs, Japan\u0026rsquo;s coffee scene perfectly embodies the nation\u0026rsquo;s pursuit of perfection. Even the subtle addition of green tea to coffee showcases the nation\u0026rsquo;s flair for experimentation. It\u0026rsquo;s a culture that has taken a foreign drink and transformed it into something uniquely Japanese.\n","date":"2 February 2024","permalink":"/blog/japan-vibrant-coffee-culture/","section":"Blog","summary":"It was a cold, rainy October morning at Haneda Airport.","title":"Discovering Japan's vibrant coffee culture"},{"content":"","date":"2 February 2024","permalink":"/tags/travel/","section":"Tags","summary":"","title":"travel"},{"content":"","date":"7 February 2023","permalink":"/tags/anecdote/","section":"Tags","summary":"","title":"anecdote"},{"content":"Anecdote # February 2016\n“Paytm?” I hesitantly asked the security personnel at the entrance of Paytm’s old headquarters, the logo emblazoned on the building. For readers outside India, Paytm is one of the largest fintech companies in the country. In 2015-16, Paytm was the front-runner among all payment apps and was gaining incredible traction.\nPaytm\u0026rsquo;s Old HQ at Noida That day, my classmate Utkarsh and I were on the ground, seeking sponsors for our college’s annual cultural festival. It was a challenging time, as many top institutions, including IITs, were holding their festivals around the same time. We faced stiff competition, making it unlikely for companies to invest in multiple events in the same geography simultaneously.\nAfter some initial missteps, the security staff directed us to a nearby office that appeared bustling with activity. Unfortunately, our unprofessional approach backfired; we hadn’t made any appointments or contacted anyone in advance, assuming we could simply walk in, as we had done with Times Internet and Kent Systems in Noida. Our lack of preparation led to disappointment, as the security staff denied us entry. After traveling 40 kilometers to pitch our sponsorship proposal, we left feeling defeated. Looking back, I now realize we deserved that rejection.\nNearby, we noticed a smoking area frequented by Paytm employees. In a last-ditch effort, Utkarsh and I decided to engage with the employees there, hoping to connect with the right team at Paytm. Our first conversation was with a gentleman who seemed to hold a higher position within the company. Instead of pitching for sponsorship directly, we proposed a brand association—selling tickets for our event through the Paytm app. To our surprise, he handed us his business card and asked us to follow up. Only later did we discover he was the DGM of Brands at Paytm—an unexpected stroke of luck! Although we were denied office access, we had connected with the right person.\nUnfortunately, the joy was short-lived. The following day, our sponsorship team secretary spoke with the same individual, but the deal fell through for various reasons, most of which were our team\u0026rsquo;s fault.\nHaving spent some time in the corporate world since then, I now understand why our proposal didn\u0026rsquo;t materialize. It wasn\u0026rsquo;t just Paytm that eluded us; we missed out on several sponsorship opportunities that year.\nOur pitch primarily focused on visibility for brands—offering title space, category partnerships (like a payments partner), and physical space for marketing collateral (facepalm). But we were not the Aston Martin of a Formula One race that attracts a massive audience. Most brands we targeted were already well-known to our audience, and more than they needed us, we desperately needed them. Lastly, there were too many brands on display already! Brands needed more than just visibility. Even if there were brands that wanted to be promoted during the event, we probably didn’t pitch to them. We were busy chasing the big business that already had a good brand recall among the fest attendees.\nWhy would brands not want publicity or visibility? Let me put it in another way - “Why were brands not excited about the space or visibility we offered?”. The answer boils down to one word: price. We were eager for sponsorship money, but our valuation was unconvincing. Marketing is a costly endeavor, and companies have quarterly budgets to allocate. If given a choice between two marketing avenues, they would select the one that appears more promising and justifies the expense. OOur event statistics were flawed, and we pitched them without proper justification. Even worse, we approached brands just a month before the festival—far too late to allow for any serious consideration.\nDespite our festival\u0026rsquo;s success that year, our team learned hard lessons. Every year, new secretaries and club heads were designated for the student council. In August 2016, I had the opportunity to lead the sponsorship team, with Satyam as my co-head. Our team expanded to over 40 members, and armed with the lessons from our past failures, we knew we needed to change our approach.\nOur festival was rebranded as Unifest (previously known as G-Quasar), giving us nearly six months to plan and execute. We adjusted our strategy in several key ways:\nBuild relations, create value for the brand, and money will follow. # We recognized that creating long-term relationships and providing value to brands would lead to sponsorship. While money was essential, our focus shifted to establishing trust and offering brands more than just title visibility. To manage relations with the brands and vendors, each team member was assigned 2-3 brands to manage, allowing us to cultivate these relationships. The results of the efforts could not be judged in the short term, it was a long-term bet.\nGet the math right. # In our attempt to get sponsors, we had been over-glorifying the stats about the event. Be it the footfall, participation from other colleges, or social media engagement stats. We learned the hard way that exaggerating our statistics only hurt us. This time, we critically reviewed our data and corrected the following aspects → Footfall: 3-year mean and absolute year-on-year increase in footfall, vendor participation: honest split of digital and on-campus participation, participants from other colleges: 3-year mean, number of colleges participated: 3-year mean and year-on-year repeat participation. We presented honest numbers that held up under scrutiny, eliminating the need for overinflation.\nJunk unprofessionalism, define standards. # While in the past, we had been able to get sponsors by going door-to-door, we decided to give up on this approach. We will not walk to any vendor without an appointment. Discuss over email/chat (no spam, strictly) and get an appointment if an in-person meeting is required. Nobody wants to be spammed, we decided not to follow up too many times and give some time to the respondent.\nThe brochure and sponsorship proposal that we sent to brands was exquisite, always. We didn’t make changes to both except adding a comparison view of various sponsorship value brackets. Earlier, when we onboarded a sponsor, a memorandum of understanding (MoU) was established between the sponsor and us. The MoU was drafted as per the agreed terms and deliverables and varied for every sponsor. We decided to make a standard MoU for each value bracket of sponsorship.\nStart Early # Time is the best teacher; We have had an experience the previous year. We didn’t want to get ourselves in the same situation because we had learned that proposal, negotiation, and onboarding take time. Companies have their own bureaucratic processes and diligence checks, which they adhere to, so, any bargain of time won’t help. That year, we started earlier than the usual timeline. Did it make any difference? A lot! Starting early helped both - brands and us.\nLuck # While luck can’t be quantified, it plays a crucial role in success. No matter how much effort our team had put in for each sponsorship deal, luck was a crucial vector of the overall success. Having said that, the team did an excellent job in all spheres. We were students, with no professional exposure, and for many of us, this was our first experience dealing with businesses.\nThrough these experiences, I learned that success in sponsorship isn’t just about selling visibility; it’s about understanding the needs of brands, building genuine relationships, and maintaining professionalism throughout the process. Each interaction provided valuable insights that extended far beyond that single festival.\n","date":"7 February 2023","permalink":"/blog/anecdote-failed-sponsorship/","section":"Blog","summary":"Anecdote # February 2016","title":"Anecdote: Taking lessons from a failed sponsorship deal."},{"content":"","date":"7 February 2023","permalink":"/tags/business/","section":"Tags","summary":"","title":"business"},{"content":"","date":"7 February 2023","permalink":"/tags/college/","section":"Tags","summary":"","title":"college"},{"content":"","date":"3 January 2023","permalink":"/tags/blog/","section":"Tags","summary":"","title":"blog"},{"content":"“Should I really start a blog?”\nHonestly, I’ve been asking myself that for about three years now. It’s been a constant cycle of getting excited about the idea, and then immediately talking myself out of it. The doubts are always the same:\nWho actually cares? Am I just shouting into the void? I\u0026rsquo;m just an ordinary person, after all.\nIs blogging dead? In an era of 60-second videos and podcasts, does anyone still want to read long-form text?\nThe commitment. Do I really need another deadline in my life?\nI eventually realized I was stuck because I was trying too hard to define a \u0026ldquo;niche.\u0026rdquo; Standard advice says pick one topic and stick to it, but that felt limiting. I didn’t want this to be just another tech blog, or purely about my adventures in malware analysis, or strictly generic product management advice.\nWriting just hits differently than video or social media. It’s slower. It’s more deliberate. You don’t usually end up on a blog because an algorithm force-fed it to you while you were doom-scrolling; reading a blog is a choice.\nI’m realizing I don\u0026rsquo;t need massive reach or a perfect posting schedule. I just need substance.\nSo, this is me officially stopping the overthinking. I’m here to share my perspective on technology, business, and just life in general.\nWelcome to the blog :)\n","date":"3 January 2023","permalink":"/blog/first/","section":"Blog","summary":"“Should I really start a blog?","title":"Just Start"},{"content":"","date":"1 January 0001","permalink":"/categories/","section":"Categories","summary":"","title":"Categories"}]