[{"data":1,"prerenderedAt":881},["ShallowReactive",2],{"blog-list-/blog":3},[4,226,496,709,864],{"id":5,"title":6,"body":7,"date":216,"description":217,"extension":218,"image":219,"meta":220,"navigation":221,"path":222,"seo":223,"stem":224,"__hash__":225},"blog/blog/rethinking-news-format-for-todays-media-landscape.md","Rethinking News Format for Today's Media Landscape",{"type":8,"value":9,"toc":203},"minimark",[10,14,17,22,25,28,31,35,38,41,45,48,54,57,60,63,70,73,77,80,85,88,93,96,99,104,107,110,113,117,120,124,127,130,133,136,139,143,153,156,161,164,167,171,174,177,180,184,187,190,194,197,200],[11,12,13],"p",{},"Every morning, billions of people reach for their phones to understand the world. The format waiting for them was designed in 1975.",[11,15,16],{},"This is a two-part piece of my thoughts on how news consumption changed, what got broken, and what the publishers missed. Then we talk about what we decided to build because of it.",[18,19,21],"h3",{"id":20},"the-60-word-morning","The 60 Word Morning",[11,23,24],{},"I grew up in India, where my daily relationship with news started with a swipe. Every morning, I'd open Inshorts, a news app that compresses each story into a crisp 60-word brief and skim through national politics, global crises, startup funding rounds, maybe a bit of Bollywood gossip along the way.",[11,26,27],{},"Inshorts' promise was simple: headlines and facts, no fluff, personalised to what you cared about. Over time it became one of the country's most popular news apps, with more than 10 million active users at its peak. For millions of people like me, it quietly reset expectations. News should be quick, snackable, and always within thumb's reach.",[11,29,30],{},"That expectation never left me. But the world it was built for has changed dramatically.",[18,32,34],{"id":33},"when-news-became-a-lifeline","When News Became a Lifeline",[11,36,37],{},"When COVID-19 hit, news stopped being a background habit and became a lifeline. Global lockdowns pushed audiences online, and demand for real-time updates on case counts, restrictions, and economic fallout sent digital news engagement soaring. People weren't just casually checking headlines anymore, they were refreshing apps, following live blogs, and obsessively tracking charts because the news suddenly had direct consequences for their safety and livelihoods.",[11,39,40],{},"At the same time, attention spans were under pressure. Between anxiety, remote work, and doomscrolling, people wanted information that was not just fast, but trustworthy and easy to share or debate with others. Two things happened simultaneously: the appetite for news grew, and the tolerance for friction collapsed.",[18,42,44],{"id":43},"the-multi-format-explosion","The Multi-Format Explosion",[11,46,47],{},"Short-form content platforms like Instagram Reels, TikTok, and Snapchat rewired how the mainstream consumes stories altogether. For a large share of Gen Z and younger audiences, these feeds are not just entertainment — they're where news appears first.\nThe numbers are striking.",[49,50,51],"blockquote",{},[11,52,53],{},"According to Pew Research Center's 2024 Social Media and News Fact Sheet, more than half of US adults (54%) now get news from social media at least sometimes, up from 48% in 2021. Among TikTok users specifically, 52% say they regularly get news from the platform, up from just 22% in 2020, the fastest rise in news consumption of any social platform measured. In total, about 17% of all US adults now say they regularly get their news from TikTok.",[11,55,56],{},"Publishers quickly followed the audience: outlets like The Washington Post, The New York Times, CNN, the Daily Mail, and the New York Post now produce vertical explainers, swipeable image carousels, and short clips tailored to these platforms. These pieces are often designed as top-of-funnel content, 30 seconds of context and a call to action to draw viewers back into their own apps and websites where longer, richer coverage lives.",[11,58,59],{},"But even as short-form video surged, the opposite end of the spectrum was booming. Commuters stuck in traffic, runners on early-morning routes, and people doing chores at home increasingly turned to 30 or 60-minute podcast episodes to go deeper into a story. Podcast listening has grown steadily year over year, with more people making it a daily habit as routines returned post-pandemic.",[11,61,62],{},"In other words, audiences were not choosing one format over another, they were choosing based on context. Sprinting to the subway? A 60-second clip. Long train ride? A 40-minute conversation. This forced publishers to retell the same core story across formats: short clips, long podcasts, articles, and explainers, each calibrated for a different moment, attention span, and level of depth.",[11,64,65,69],{},[66,67,68],"strong",{},"Today, a single news event might exist as a 30-second TikTok, a 5-minute vertical explainer, a 1,200-word article, and a 30-minute podcast episode."," A curious reader who really wants to understand what's going on might encounter the story in their social feed first, then search for a longer video on YouTube, then skim a written explainer, and finally land in a community discussion or fact-check thread.",[11,71,72],{},"What used to be a single touchpoint, reading yesterday's paper, is now a fragmented journey across apps, formats, and communities.",[18,74,76],{"id":75},"the-hidden-problem-format-without-trust","The Hidden Problem: Format Without Trust",[11,78,79],{},"In the current news climate, narratives around major stories are shaped by a messy mix of influencers, creators, publishers, and traditional outlets, each with their own incentives and biases. Some prioritise speed over accuracy, others chase virality, and a few openly push agendas.",[49,81,82],{},[11,83,84],{},"The Reuters Institute Digital News Report 2024 — based on a YouGov survey of over 95,000 online news consumers across 47 markets — found that 39% of respondents say they sometimes or often avoid the news entirely. The primary driver isn't disinterest. It's being overwhelmed. Too many sources, too many formats, too much noise — and no single place that pulls it together in a way you can trust.",[11,86,87],{},"Research on social media shows that people who primarily rely on these platforms for news are often less informed and more exposed to unproven or misleading claims compared with those who lean on dedicated news sites or TV. Validating what is actually true is not an optional extra but has become a critical part of consuming news.",[49,89,90],{},[11,91,92],{},"Pew Research's 2024 study of the four major social platforms found that roughly a quarter or more of news consumers on each platform say they at least fairly often see news that seems inaccurate.",[11,94,95],{},"Validating what's actually true has become a critical — and exhausting — part of consuming news, not an optional extra.",[11,97,98],{},"That's where online communities step in as an informal layer of collective fact-checking. Platforms like Reddit host thousands of posts linking to news stories that attract verification comments. People don’t just scroll Reddit for extra context; they also read the comments to understand how different groups interpret the same story, which claims are being challenged, and what additional sources are being surfaced to support or refute it.",[49,100,101],{},[11,102,103],{},"One large-scale study analysing roughly 29,000 such posts and 9.8 million comments found that threads where information was confirmed as true often drove more engagement than those debunked as false, highlighting how community validation shapes attention.",[11,105,106],{},"On X (formerly Twitter), Community Notes allows users to collaboratively attach context and corrections directly under misleading posts, and AI systems i.e. Grok are being tested to help surface relevant sources and verify claims at scale.",[11,108,109],{},"For the average reader, the reality is still exhausting. To feel confident about just one important story, you might watch a short clip on Instagram, listen to a podcast episode on your commute, skim a couple of articles, and read through Reddit threads to see how others are interpreting and challenging the narrative. Only after this multi-step, multi-platform, multi-format journey do you start to feel like you \"get\" the story — and even then, you might still be unsure what to trust.",[11,111,112],{},"Our information diet has never been richer in formats, but also never more fragmented and fragile in trust.",[18,114,116],{"id":115},"burden-on-the-reader","Burden on the Reader",[11,118,119],{},"For the average reader, though, the reality is still exhausting. To feel confident about just one important story, you might watch a short clip on Instagram or TikTok, listen to a podcast episode on your commute, skim a couple of articles, and then read through Reddit threads or Community Notes on X to see how others are interpreting and challenging the narrative. Only after this multi-step, multi-platform, multi-format journey do you start to feel like you “get” the story and even then, you might still be unsure what to trust. Our information diet has never been richer in formats, but also never more fragmented and fragile in trust.",[18,121,123],{"id":122},"the-format-hasnt-changed","The Format Hasn't Changed",[11,125,126],{},"Here's what I learned from the inside.",[11,128,129],{},"During my time on the constent & monetization team at The Washington Post, I saw product innovation happening constantly, new subscription flows, engagement experiments, format testing at the article level. What I didn't see was anyone questioning the article itself. The inverted pyramid, the headline, the byline, these are conventions from print journalism that have survived the entire internet era untouched. Publishers have innovated everything around the format: distribution (social, push, email), the business model (subscriptions, paywalls, metered access), the discovery surface (personalisation, notifications). But the atomic unit of news, the article, is structurally identical to what ran in a newspaper in 1975.",[11,131,132],{},"The referrer data tells the real story. For most major publishers, the top traffic sources are social platforms, feeds and apps built around short-form, visual, swipeable content. The format that actually acquires a reader is TikTok's or Instagram's. The publisher captures them for 90 seconds on an article page, then loses them back to the feed. The audience belongs to the platform, not the publisher.",[11,134,135],{},"The Reuters Institute's 2024 industry survey confirmed this: publishers said they were planning significant increases in short-form video production for TikTok (+55 net investment score) and YouTube (+44) — not because they're leading the format shift, but because they're chasing an audience that moved without them. Publishers are optimising for the reader they have. Nobody is designing for the reader they're losing.",[11,137,138],{},"That gap is exactly what led to TunedIn.",[18,140,142],{"id":141},"introducing-tunedin-what-were-building","Introducing TunedIn — What We're Building",[11,144,145,152],{},[146,147,151],"a",{"href":148,"rel":149},"https://tunedin.live",[150],"nofollow","TunedIn"," is a multimodal short-form content platform built for how people actually consume information today. How people actually do it, bouncing between apps, formats, and communities, trying to stitch together a complete picture of something that matters to them.",[11,154,155],{},"The core bet: treat any piece of content from any publisher or creator as raw material, and reformat it into whatever shape the moment demands.",[157,158,160],"h4",{"id":159},"the-core-experience","The Core Experience",[11,162,163],{},"A single story in TunedIn is not an article. It's a story object, a swipeable, multimodal narrative that weaves together the article, the relevant video explainer, the podcast clip where an expert went deep on exactly this point, the Reddit thread where the community pushed back, and the tweet from the journalist who broke it. All of it, in one place, in a format you can consume on the subway in 90 seconds or sit with for 20 minutes if you want to go deep.",[11,165,166],{},"The feed learns. TunedIn tracks watch history, scroll depth, interaction signals (likes, comments, saves), and general reading patterns to surface what's relevant to you, not what's trending for everyone, but what's next for you specifically. The same story is presented differently depending on how you read: a skimmer gets the takeaways first; a deep reader gets context and analysis; a visual learner gets led by the images and video.",[157,168,170],{"id":169},"the-story-pipeline","The Story Pipeline",[11,172,173],{},"Long-form articles are automatically converted into 3–10 swipeable story slides — Instagram/Snapchat-style — calibrated for mobile-first consumption. The pipeline uses an LLM to extract key sections, generate slide-sized summaries (50–150 words each), and assemble a fixed narrative arc: an intro slide, content slides that alternate text-heavy and media-rich layouts, and a final takeaways slide with three key bullet points.",[11,175,176],{},"Slide count scales with article length: short pieces become 3 slides, pieces over 2,000 words hit the 10-slide maximum. Every story passes a quality gate before storage, minimum slide count, verified media coverage, a floor on description length relative to the original article — so nothing gets surfaced half-finished.",[11,178,179],{},"The same story reported across multiple outlets is stitched together, with sourcing made visible. You see how the BBC framed it, how the Times framed it, and what the difference tells you. Media bias isn't lectured about — it's made observable.",[157,181,183],{"id":182},"the-community-layer","The Community Layer",[11,185,186],{},"Comments in TunedIn is a credibility layer. Users earn subject-matter expert (SME) badges based on their reading and engagement history within specific topics. A reader who consistently engages with sports coverage, over time, surfaces higher in comment threads on sports stories, lending community-validated credibility to the reporting. This is balanced against upvotes, downvotes, and recency signals to prevent gaming. An in-feed AI assistant lets users call on AI to clarify claims, research a topic, or surface sources without leaving the feed closer to Community Notes than a chatbot.",[11,188,189],{},"Sharing is designed to be contextually rich. When you share a story, you share a screenshot of your current view, not just a bare link, giving the person on the other end immediate visual context. Native to WhatsApp, Messages, Instagram Stories, Snapchat, Discord.",[18,191,193],{"id":192},"why-this-why-now","Why This, Why Now",[11,195,196],{},"The macro conditions for this product are better now than they have ever been.",[11,198,199],{},"The Reuters Institute's 2024 report documents explicitly: \"publishers are exploring different formats as a way of addressing the engagement challenge.\" Two-thirds of the 95,000-person global survey access short news videos weekly. News avoidance is at 39% and rising, not because people don't care, but because the current experience is exhausting. At the same time, the tools to do this — LLMs capable of high-quality summarisation, multimodal embeddings, real-time content APIs — only exist at accessible cost from roughly 2023 onward.",[11,201,202],{},"The format shift is documented. The audience is there. The infrastructure now exists. What's missing is a product that treats format as the core design decision.",{"title":204,"searchDepth":205,"depth":205,"links":206},"",2,[207,209,210,211,212,213,214,215],{"id":20,"depth":208,"text":21},3,{"id":33,"depth":208,"text":34},{"id":43,"depth":208,"text":44},{"id":75,"depth":208,"text":76},{"id":115,"depth":208,"text":116},{"id":122,"depth":208,"text":123},{"id":141,"depth":208,"text":142},{"id":192,"depth":208,"text":193},"2025-02-19","The way we consume news has never been more fragmented, from TikTok explainers to Reddit fact-check threads to 40-minute podcasts. Audiences aren't choosing one format; they're navigating all of them just to understand a single story. This piece traces how that shift happened, and why the burden of stitching it all together has fallen entirely on the reader — and it's exactly what led us to build WaPOW!, a multimodal short-form content platform designed for the way people actually consume information today.","md",null,{},true,"/blog/rethinking-news-format-for-todays-media-landscape",{"title":6,"description":217},"blog/rethinking-news-format-for-todays-media-landscape","Z-a2D9mH1x0sz7BmDGRtaJ7V1F8_tWg9Co0bYCUZYjA",{"id":227,"title":228,"body":229,"date":487,"description":488,"extension":218,"image":489,"meta":490,"navigation":221,"path":491,"seo":492,"stem":494,"__hash__":495},"blog/blog/clipped-short-form-podcast-discovery.md","Clipped - Short form podcast discovery",{"type":8,"value":230,"toc":465},[231,236,239,242,245,249,252,256,259,262,266,279,283,286,289,293,297,308,312,320,324,328,339,343,354,358,378,382,385,389,393,437,441],[232,233,235],"h2",{"id":234},"problem-space-background","Problem space & Background",[11,237,238],{},"Podcasts are one of the leading formats for content consumption, with the market expected to grow from about USD 30.7 billion in 2024 to over USD 131 billion by 2030 at roughly a 27% CAGR.",[11,240,241],{},"One of the primary ways publishers, studios, and creators now drive discovery, impressions, and listenership is by repurposing episodes into short, viral‑ready clips and distributing them across platforms like TikTok, Instagram Reels, YouTube/Shorts, Snapchat, Reddit, X, and Facebook, either directly or via their communities. These clips increasingly function as top‑of‑funnel assets: many listeners first encounter a show through a 30–90 second highlight before moving on to full episodes, subscribing, or entering a broader funnel that can include newsletters, communities, and paid products.",[11,243,244],{},"As a result, clipping has become a significant component of the creator ecosystem value chain, now supported and scaled by AI tools that automatically identify “clip‑worthy” moments and generate social‑native assets. However, the engagement and business outcomes of these clips are hard to predict and even harder to attribute, due to opaque platform algorithms, heterogeneous user demographics and usage patterns, and fragmented cross‑platform analytics, meaning there is still no reliable recipe for virality or clear line of sight from any given clip to downstream listens or revenue.",[232,246,248],{"id":247},"product-overview","Product overview",[11,250,251],{},"A mobile‑first podcast discovery platform where users scroll a feed of short, high‑signal clips (30–120 seconds) with auto‑generated titles, summaries, topics, and viral potential scores, sourced from full podcast episodes via RSS/YouTube and user uploads. Unlike generic short‑form feeds, every clip in the system is tracked end‑to‑end—from impression to full‑episode play, follow, and downstream actions—giving creators a single place to generate, distribute, and measure clips that actually grow their shows.",[232,253,255],{"id":254},"market-analysis","Market analysis",[11,257,258],{},"Global podcasting market ≈ USD 32–40B in 2025–26, projected to exceed USD 360B by 2035 at ~27% CAGR, with YouTube now the top podcast platform by consumption.\nThere are ~8M podcasts globally and hundreds of millions of listeners; discovery is still a pain as supply outpaces listeners’ time, and most discovery is via social clips and platform charts.",[11,260,261],{},"There is room for a listener‑first, clip‑native discovery layer that sits on top of Spotify, Apple Podcasts, and YouTube without trying to replace them as full players.",[232,263,265],{"id":264},"competitive-analysis","Competitive analysis",[267,268,269,273,276],"ul",{},[270,271,272],"li",{},"AI podcast players – e.g., Snipd and Airr focus on note‑taking, AI transcripts, chapters, and personal knowledge management, not a global viral discovery feed.",[270,274,275],{},"AI clipping tools for creators – e.g., Flowjin, OpusClip, Descript, VEED help creators turn long-form content into social clips for TikTok/Shorts/Reels.",[270,277,278],{},"General podcast apps – Spotify, Apple Podcasts, YouTube optimize for full‑episode listening and network lock‑in, not clip‑centric browsing.",[18,280,282],{"id":281},"the-differentiator","The Differentiator",[11,284,285],{},"Listener‑first For You Feed of cross‑show clips, not a note‑taking app or editing SaaS.\nNeutral “clip layer” across platforms that deep‑links out to full episodes, instead of locking people into a single player or only exporting to socials.",[11,287,288],{},"AI‑driven virality and story‑moment detection plus community curation, whereas clipping tools largely stop at generating assets and scheduling posts.",[232,290,292],{"id":291},"goals-and-nongoals","Goals and non‑goals",[18,294,296],{"id":295},"goals","Goals",[267,298,299,302,305],{},[270,300,301],{},"Help listeners quickly discover new shows, topics, and stories via short clips instead of full episodes.",[270,303,304],{},"Enable creators and power users to curate and edit clips, improving AI output over time.",[270,306,307],{},"Build a data flywheel where engagement signals continuously improve clip quality and recommendations.",[18,309,311],{"id":310},"nongoals-v1","Non‑goals (v1)",[267,313,314,317],{},[270,315,316],{},"Full podcast player and library management (we’ll deep-link to Spotify / Apple Podcasts / YouTube).",[270,318,319],{},"Advanced video editing suite; focus is discovery, not full production tooling.",[232,321,323],{"id":322},"target-users-and-use-cases","Target users and use cases",[18,325,327],{"id":326},"primary-users","Primary users",[267,329,330,333,336],{},[270,331,332],{},"Casual listeners: Want entertaining/insightful clips while scrolling, then follow shows or save episodes.",[270,334,335],{},"Curious learners: Want topic-based discovery (e.g., “AI in marketing”, “founder stories”).",[270,337,338],{},"Creators/podcasters: Want more reach and data on which moments resonate.",[18,340,342],{"id":341},"core-use-cases","Core use cases",[267,344,345,348,351],{},[270,346,347],{},"“I have 5 minutes and want to discover something interesting via snackable podcast clips.”",[270,349,350],{},"“I want viral / trending clips in my interests without hunting across platforms.”",[270,352,353],{},"“As a creator, I want automatically suggested clips and an easy way to tweak and publish them.”",[232,355,357],{"id":356},"value-proposition","Value proposition",[267,359,360,366,372],{},[270,361,362,365],{},[66,363,364],{},"Clip-first feed:"," Short, self-contained clips optimized for attention and shareability, not full episodes.",[270,367,368,371],{},[66,369,370],{},"AI + human curation:"," AI proposes clips and scores virality; users upvote, refine boundaries, add tags, and correct summaries.",[270,373,374,377],{},[66,375,376],{},"Topic and goal-oriented discovery:"," Browse by themes and goals (learn X, relax, get inspired), not just shows.",[232,379,381],{"id":380},"key-user-journeys","Key User Journeys",[383,384],"miro-customer-journey",{},[232,386,388],{"id":387},"gotomarket-and-marketing-efforts","Go‑to‑market and marketing efforts",[18,390,392],{"id":391},"ideal-target-segments-sequenced","Ideal target segments (sequenced)",[267,394,395,414,429],{},[270,396,397,398],{},"Early adopter creators and indie podcasters\n",[267,399,400,403],{},[270,401,402],{},"Pain: need viral clips and better discovery but can’t afford expensive video teams.",[270,404,405,406],{},"Tactic:\n",[267,407,408,411],{},[270,409,410],{},"Outreach to 50–100 indie shows in specific verticals (e.g., tech / business /  self‑improvement).",[270,412,413],{},"Offer free clipping + discovery analytics for 3–6 months in exchange for feedback and promotion.",[270,415,416,417],{},"Listeners in specific verticals (focus on “learning” podcasts)\nOverlaps with Snipd / productivity community, but with a discovery focus.\n",[267,418,419],{},[270,420,405,421],{},[267,422,423,426],{},[270,424,425],{},"Launch themed feeds (“AI & Startups”, “Health & Wellness Stories”).",[270,427,428],{},"Run content partnerships with a handful of known shows in each theme.",[270,430,431,432],{},"Agencies and networks\n",[267,433,434],{},[270,435,436],{},"Tactic: show case studies on how “clip discovery” increases full‑episode plays and social engagement vs traditional clipping only.",[18,438,440],{"id":439},"channels-tactics","Channels & tactics",[267,442,443,449,459],{},[270,444,445,448],{},[66,446,447],{},"Content marketing and case studies:"," Write “How we turned 1 hour of podcast into 10K extra listens via clip discovery” breakdowns, benchmarked against existing clipping tools.",[270,450,451,454,455,458],{},[66,452,453],{},"Social proof flywheel:"," Auto‑generate branded videos for TikTok/Reels with “Discover more clips like this on ",[66,456,457],{},"Clipped","” overlays, driving organic installs.",[270,460,461,464],{},[66,462,463],{},"Product‑led growth:"," Free listener app; creators get basic analytics (top clips, topics) for free, with paid tiers for more depth.",{"title":204,"searchDepth":205,"depth":205,"links":466},[467,468,469,470,473,477,481,482,483],{"id":234,"depth":205,"text":235},{"id":247,"depth":205,"text":248},{"id":254,"depth":205,"text":255},{"id":264,"depth":205,"text":265,"children":471},[472],{"id":281,"depth":208,"text":282},{"id":291,"depth":205,"text":292,"children":474},[475,476],{"id":295,"depth":208,"text":296},{"id":310,"depth":208,"text":311},{"id":322,"depth":205,"text":323,"children":478},[479,480],{"id":326,"depth":208,"text":327},{"id":341,"depth":208,"text":342},{"id":356,"depth":205,"text":357},{"id":380,"depth":205,"text":381},{"id":387,"depth":205,"text":388,"children":484},[485,486],{"id":391,"depth":208,"text":392},{"id":439,"depth":208,"text":440},"2025-02-12","A mobile-first podcast discovery platform where users scroll a feed of short, high-signal clips (30–120 seconds) with auto-generated titles, summaries, topics, and “viral potential” scores, sourced from full podcast episodes via RSS/YouTube and user uploads","/clipped.png",{},"/blog/clipped-short-form-podcast-discovery",{"title":493,"description":488},"Clipped - Clipped - Short form podcast discovery file","blog/clipped-short-form-podcast-discovery","M17NmehHKxrfxxFHHwrF2Lm9le06S0XhAo9nLAmzGGY",{"id":497,"title":498,"body":499,"date":702,"description":703,"extension":218,"image":219,"meta":704,"navigation":221,"path":705,"seo":706,"stem":707,"__hash__":708},"blog/blog/lessons-shipping-nabuai.md","Lessons from shipping NabuAI",{"type":8,"value":500,"toc":696},[501,510,517,521,528,539,546,550,553,620,635,639,665,669,672,682,685,688],[11,502,503,504,509],{},"We’ve been building ",[146,505,508],{"href":506,"rel":507},"https://nabuai.me",[150],"NabuAI"," – an AI-powered knowledge management tool for researchers and knowledge workers. Here are a few lessons from the first months.",[11,511,512],{},[513,514],"img",{"alt":515,"src":516},"Laptop and notes","https://images.unsplash.com/photo-1517694712202-14dd9538aa97?w=800&q=80",[232,518,520],{"id":519},"_1-rag-before-custom-models","1. RAG before custom models",[11,522,523,524,527],{},"We tried to get fancy with custom embeddings and fine-tuning early. It slowed us down and didn’t move the needle for our early users. Switching to a simple ",[66,525,526],{},"GraphRAG","-style pipeline (chunk → embed → retrieve → generate) gave us:",[267,529,530,533,536],{},[270,531,532],{},"Faster iteration",[270,534,535],{},"Good enough quality for v1",[270,537,538],{},"Room to improve retrieval and prompts without changing the whole stack",[11,540,541,542,545],{},"So: ",[66,543,544],{},"get RAG right first",", then consider custom models only when you have clear, measurable gaps.",[232,547,549],{"id":548},"_2-where-time-actually-went","2. Where time actually went",[11,551,552],{},"Rough breakdown of where engineering time went in the first 3 months:",[554,555,556,572],"table",{},[557,558,559],"thead",{},[560,561,562,566,569],"tr",{},[563,564,565],"th",{},"Area",[563,567,568],{},"% of time",[563,570,571],{},"Note",[573,574,575,587,598,609],"tbody",{},[560,576,577,581,584],{},[578,579,580],"td",{},"Retrieval & RAG",[578,582,583],{},"~35%",[578,585,586],{},"Chunking, embeddings, ranking",[560,588,589,592,595],{},[578,590,591],{},"UI & workflows",[578,593,594],{},"~30%",[578,596,597],{},"Capture, organize, query",[560,599,600,603,606],{},[578,601,602],{},"Integrations",[578,604,605],{},"~20%",[578,607,608],{},"Chrome extension, APIs",[560,610,611,614,617],{},[578,612,613],{},"Infra & ops",[578,615,616],{},"~15%",[578,618,619],{},"Supabase, Postgres, deploy",[11,621,622,623,628,629,634],{},"Integrations and UI took more than we’d guessed; infra stayed manageable thanks to ",[146,624,627],{"href":625,"rel":626},"https://supabase.com",[150],"Supabase"," and ",[146,630,633],{"href":631,"rel":632},"https://vercel.com",[150],"Vercel",".",[232,636,638],{"id":637},"_3-links-we-leaned-on","3. Links we leaned on",[267,640,641,649,657],{},[270,642,643,648],{},[146,644,647],{"href":645,"rel":646},"https://docs.llamaindex.ai/",[150],"LlamaIndex docs"," – for RAG patterns and graph-based retrieval",[270,650,651,656],{},[146,652,655],{"href":653,"rel":654},"https://vuejs.org/",[150],"Vue 3 + TypeScript"," – for the frontend",[270,658,659,664],{},[146,660,663],{"href":661,"rel":662},"https://tailwindcss.com/",[150],"Tailwind CSS"," – for layout and theming",[232,666,668],{"id":667},"_4-one-code-decision-that-helped","4. One code decision that helped",[11,670,671],{},"We kept “sources” for every answer in a simple table and exposed them in the UI from day one:",[673,674,679],"pre",{"className":675,"code":677,"language":678,"meta":204},[676],"language-text","answer_sources: [ document_id, chunk_id, score ]\n","text",[680,681,677],"code",{"__ignoreMap":204},[11,683,684],{},"That made it easy to show “from which note this came” and to debug bad answers. Small schema choice, big product and debugging win.",[686,687],"hr",{},[11,689,690,691,634],{},"More posts on GraphRAG, the Chrome extension, and our stack choices are planned. If you’re building something similar, ",[146,692,695],{"href":693,"rel":694},"https://twitter.com/coderuth",[150],"say hi",{"title":204,"searchDepth":205,"depth":205,"links":697},[698,699,700,701],{"id":519,"depth":205,"text":520},{"id":548,"depth":205,"text":549},{"id":637,"depth":205,"text":638},{"id":667,"depth":205,"text":668},"2025-01-22","What worked and what didn’t when building an AI-powered knowledge base.",{},"/blog/lessons-shipping-nabuai",{"title":498,"description":703},"blog/lessons-shipping-nabuai","83kaIj0d02ESv3stbEmArb8RMaYjkCCgJ0pyXC9c_uM",{"id":710,"title":711,"body":712,"date":857,"description":858,"extension":218,"image":219,"meta":859,"navigation":221,"path":860,"seo":861,"stem":862,"__hash__":863},"blog/blog/building-products-that-scale.md","Building products that scale",{"type":8,"value":713,"toc":852},[714,722,728,732,752,766,770,773,834,840,844,849],[11,715,716,717,721],{},"When you're building something new, it helps to separate what ",[718,719,720],"em",{},"scales"," from what doesn't. Here’s a short take with concrete examples.",[11,723,724],{},[513,725],{"alt":726,"src":727},"Team and whiteboard","https://images.unsplash.com/photo-1522071820081-009f0129c71c?w=800&q=80",[232,729,731],{"id":730},"what-actually-scales","What actually scales",[267,733,734,740,746],{},[270,735,736,739],{},[66,737,738],{},"Clear problem and audience"," – If you can’t say who it’s for and what pain it solves in one sentence, you’ll keep changing scope.",[270,741,742,745],{},[66,743,744],{},"Simple data model"," – Fancy schemas are tempting; start with the smallest model that supports the core flow.",[270,747,748,751],{},[66,749,750],{},"Automation over manual steps"," – Every manual step will break at 10x. Automate early where it matters.",[11,753,754,755,628,760,765],{},"For more on product thinking, ",[146,756,759],{"href":757,"rel":758},"https://www.lennysnewsletter.com/",[150],"Lenny’s Newsletter",[146,761,764],{"href":762,"rel":763},"https://www.reforge.com/",[150],"Reforge"," are good places to go deeper.",[232,767,769],{"id":768},"picking-the-stack","Picking the stack",[11,771,772],{},"Rough comparison I keep in mind when starting something new:",[554,774,775,788],{},[557,776,777],{},[560,778,779,782,785],{},[563,780,781],{},"Need",[563,783,784],{},"Prefer",[563,786,787],{},"Avoid (until you need it)",[573,789,790,801,812,823],{},[560,791,792,795,798],{},[578,793,794],{},"Backend",[578,796,797],{},"Supabase, Edge functions",[578,799,800],{},"Custom servers, k8s",[560,802,803,806,809],{},[578,804,805],{},"Frontend",[578,807,808],{},"Vue/React + Tailwind",[578,810,811],{},"Heavy design systems",[560,813,814,817,820],{},[578,815,816],{},"AI / search",[578,818,819],{},"GPT, embeddings, RAG",[578,821,822],{},"Custom models, fine-tuning",[560,824,825,828,831],{},[578,826,827],{},"Auth & payments",[578,829,830],{},"Built-in (Supabase, Stripe)",[578,832,833],{},"DIY auth, custom billing",[11,835,836,837,634],{},"Start with the “Prefer” column; move right only when you hit real limits. I used this kind of filter when choosing the stack for ",[146,838,508],{"href":506,"rel":839},[150],[232,841,843],{"id":842},"one-rule","One rule",[49,845,846],{},[11,847,848],{},"Ship the smallest version that delivers the core value. Add scale and polish after people use it.",[11,850,851],{},"Two more posts on shipping and on NabuAI are coming next.",{"title":204,"searchDepth":205,"depth":205,"links":853},[854,855,856],{"id":730,"depth":205,"text":731},{"id":768,"depth":205,"text":769},{"id":842,"depth":205,"text":843},"2025-01-20","A few principles I've learned from shipping across AdTech, EdTech, and startups.",{},"/blog/building-products-that-scale",{"title":711,"description":858},"blog/building-products-that-scale","XbjsQbcdqJOHTYwido2jHJrSib6vfUBsxpE-LNZ4pTE",{"id":865,"title":866,"body":867,"date":874,"description":875,"extension":218,"image":219,"meta":876,"navigation":221,"path":877,"seo":878,"stem":879,"__hash__":880},"blog/blog/welcome.md","Welcome to the blog",{"type":8,"value":868,"toc":872},[869],[11,870,871],{},"This is the first post. More writing on product, engineering, and building things will show up here.",{"title":204,"searchDepth":205,"depth":205,"links":873},[],"2025-01-15","A short note on what to expect from this space.",{},"/blog/welcome",{"title":866,"description":875},"blog/welcome","HcntiTpFMkbbdwu52onT3nCe5tRjCJ3TJLci1vX67A4",1775533848128]