<?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 Ravi Khurana on Medium]]></title>
        <description><![CDATA[Stories by Ravi Khurana on Medium]]></description>
        <link>https://medium.com/@ravikhurana_38440?source=rss-9821e85972b1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*if47CpOd_RQQvDkZ.</url>
            <title>Stories by Ravi Khurana on Medium</title>
            <link>https://medium.com/@ravikhurana_38440?source=rss-9821e85972b1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 16:29:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ravikhurana_38440/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[The Hidden Cost of AI Inference (and How It Finally Became Visible)]]></title>
            <link>https://medium.com/@ravikhurana_38440/the-hidden-cost-of-ai-inference-and-how-it-finally-became-visible-04015dc2b534?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/04015dc2b534</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Wed, 07 Jan 2026 06:16:50 GMT</pubDate>
            <atom:updated>2026-01-07T06:16:50.696Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>A practical guide on how AI inference costs accumulate in Vibe-Coded apps and how to audit them - m</strong>ost savings tend to come from fewer calls, not cheaper tokens.</h4><p>I didn’t rack up AI tech debt and token drain because the models were too power hungry.<br>It got expensive because my cost wasn’t visible or actively being monitored.</p><p>Like many teams building AI-powered products, the assumption is that things are already “cost conscious”:</p><ul><li>use the cheaper models where possible</li><li>place some sort of token limits</li><li>spending “credits” and not “dollars” directly</li></ul><p>Inference costs still felt super unpredictable and I had no few ways of forecasting expenses if usage grew overnight. With little visibility into the AI spend and how it translated to margin, it became difficult to project a legitimate and profitable business model.</p><p>So shipping new features needed to stop and the auditing began.</p><p>This post covers:</p><ul><li>where hidden inference costs were actually hiding</li><li>why inference exposure is consistently underestimated</li><li>a practical methodology to audit your vibe-coded AI app</li><li>how to reduce cost without making the product worse</li></ul><h3>The Mistake us Vibe-coders Make</h3><p>Inference cost is usually thought of as:</p><blockquote>(tokens × price) per request</blockquote><p>That framing only holds in toy examples.</p><p>In real systems, inference behaves more like a <strong>graph traversal</strong>:</p><ul><li><strong>fan-out</strong> <em>(one user action triggering many LLM calls across strategies or models)</em></li><li><strong>retries</strong> <em>(automatic re-invocations after errors, timeouts, or schema failures)</em></li><li><strong>judges</strong> <em>(LLM calls used to score, rank, or evaluate other model outputs)</em></li><li><strong>improvement loops</strong> <em>(follow-up generations to refine or “polish” earlier outputs)</em></li><li><strong>fallback chains</strong> <em>(secondary model calls when a primary model fails or scores poorly)</em></li><li><strong>parallel calls</strong> <em>(multiple LLM calls executed simultaneously for the same request)</em></li><li><strong>unbounded context</strong> <em>(inputs that grow over time — histories, transcripts, scratchpads — without hard limits)</em></li></ul><p>By the time a user sees “one AI action,” the backend may have executed <strong>15–40 LLM calls </strong>where the costs outweigh the value provided.</p><p>That wasn’t obvious until the entire path was mapped end-to-end.</p><h3>Where the Hidden Costs Were Actually Hiding</h3><h3>1. Fan-Out Multiplication</h3><p>One feature → many strategies → many models.</p><p>For example:</p><ul><li>6 strategies</li><li>× 2 models per strategy</li><li>= 12 generations</li></ul><p>Then those outputs are judged.</p><p>Fan-out feels harmless because each call is “small.”<br>But costs grow <strong>exponentially</strong>, not linearly.</p><h3>2. Multi-Judge Evaluation</h3><p>Two models were used to judge the same output to “increase confidence.”</p><p>In practice:</p><ul><li>judges correlated heavily</li><li>disagreements were rare</li><li>costs doubled immediately</li></ul><p>Judging is important throughout the process to assess what works in which scenarios. But good enough is often good enough and less is more as long as results are delivered.</p><h3>3. Improvement Loops</h3><p>After generation and judging, workflows often ran:</p><ul><li>“improve this”</li><li>then re-judge</li></ul><p>These calls are expensive and highly variable.<br>They trigger only sometimes, which makes spend harder to predict and explain.</p><p>This is where volatility enters the system.</p><h3>4. Sequential “Parallel” Detectors</h3><p>Some workflows ran:</p><ul><li>multiple detectors</li><li>sequentially</li><li>with early exit</li></ul><p>This is safer than full parallelism, but still creates <strong>worst-case linear exposure</strong> when early exit doesn’t trigger.</p><p>If the worst case isn’t modeled, spend is underestimated.</p><h3>5. Unbounded Context</h3><p>Meeting transcripts. Scratchpads. Long histories.</p><p>Each felt reasonable on its own, but together they created silent token growth. This growth was more tied to a <strong>users behavior </strong>and not necessarily product intent which made it unpredictable and less effective over time.</p><p>These are some of the hardest costs to spot and consequently the easiest to let explode.</p><h3>Why Credits Didn’t Save Things</h3><p>A credit system was used to abstract costs. Many teams do this.</p><p>Credits help with pricing — but they <strong>hide cost shape</strong>, not cost.</p><p>Issues that surfaced:</p><ul><li>different actions burned wildly different credits</li><li>credit → dollar mapping wasn’t intuitive</li><li>overage pricing created margin asymmetry</li><li>internal intuition for real spend degraded</li></ul><p>Credits are a UI layer.<br>Auditing still has to happen underneath.</p><h3>Step 1: Switch to Read-Only Audit Mode</h3><p>Before optimizing anything, our prompts were forced into <strong>read-only mode</strong>.</p><p>No code refactors.<br>No architecture “shoulds.”<br>No solutions.</p><p>Only:</p><ul><li>mapping calls</li><li>counting tokens</li><li>calculating dollars</li><li>understanding scaling behavior</li></ul><p>Skipping this step almost guarantees optimizing the wrong thing.</p><h3>Step 2: Build an Inference Inventory</h3><p>For every AI-powered feature, write down:</p><ul><li>how many LLM calls occur per user action</li><li>which models are used</li><li>average input/output tokens</li><li>fan-out patterns</li><li>retry behavior</li><li>early exits</li><li>worst case vs typical case</li></ul><p>If the answer to these isn’t known, that’s already a risk signal.</p><p>This alone surfaced many blind spots and opportunities to reduce cost.</p><h3>Step 3: Analyze Scaling, Not Just Current Spend</h3><p>The real question isn’t:</p><blockquote><em>“What does this cost today?”</em></blockquote><p>It’s:</p><blockquote><em>“What happens when usage explodes 5–10x?”</em></blockquote><p>What became clear:</p><ul><li>some features scaled linearly</li><li>others scaled <strong>super-linearly</strong></li><li>one feature dominated ~65% of total inference spend!</li></ul><p>Sutton’s law told me to immediately change where attention should go.</p><h3>Step 4: Separate Financial Risk From Product Quality</h3><p>This is more of a cultural decision.</p><p>The question stopped being:</p><blockquote><em>“Is this AI better than what’s out there?”</em></blockquote><p>And the question became:</p><blockquote><em>“Is this priced appropriately for the value it creates?”</em></blockquote><p>Some expensive calls were justified.<br>Others were just inertia.</p><p>Making AI cheaper doesn’t require making it worse but it does require awareness and the <strong>right kind of intelligence</strong>.</p><h3>Step 5: Optimize in This Order (What Actually Works)</h3><p>Once cost exposure is better understood, the highest-leverage reductions tend to be:</p><ul><li>collapsing redundant judges</li><li>reducing strategy fan-out</li><li>gating improvement loops aggressively</li><li>separating generation models from evaluation models</li><li>adding progressive sampling (stop early when quality is already good)</li><li>capping context growth explicitly</li></ul><p>The key insight:<br><strong>Most savings come from fewer calls, not cheaper tokens.</strong></p><h3>Prompts Help When Scoped Correctly</h3><p>Tools like Lovable can help, but only with some constraints.</p><p>Two modes that were helpful:</p><ul><li><strong>Read-Only Financial Audit Mode</strong></li><li><strong>Optimization Mode (explicitly opt-in)</strong></li></ul><p>I think it’s better if these are not mixed.</p><p>A simplified read-only audit prompt:</p><blockquote>You are operating in READ-ONLY audit mode.<em><br></em>Do not recommend changes.<em><br></em>Do not propose optimizations.<em><br></em>Only describe inference cost, scaling behavior, and financial risk</blockquote><p>That constraint matters more than the wording.</p><h3>The Biggest Lesson</h3><p>AI inference doesn’t get expensive because tokens are pricey or models are power hungry.</p><p>It gets expensive because:</p><ul><li>cost compounds quietly</li><li>architecture hides multiplication</li><li>quality is optimized before exposure is understood</li></ul><p>Once inference spend can be explained to:</p><ul><li>a CFO</li><li>an enterprise buyer</li><li>or a future version of yourself at 10× scale</li></ul><p>You earn back control and can confidently spell out your business model.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04015dc2b534" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Intent, Agency, and the Quiet Erosion of Privacy Online]]></title>
            <link>https://medium.com/@ravikhurana_38440/intent-agency-and-the-quiet-erosion-of-privacy-online-e51d9ff78601?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/e51d9ff78601</guid>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[behavioral-economics]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[adtech]]></category>
            <category><![CDATA[privacy]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Sun, 28 Dec 2025 15:53:26 GMT</pubDate>
            <atom:updated>2025-12-28T15:53:26.283Z</atom:updated>
            <content:encoded><![CDATA[<h4>How inference quietly shapes choice, behavior, and privacy online</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*22pJplGH5XGdyRTfnxyMfg.png" /></figure><p>The internet can often feel like we have infinite choices and agency.</p><p>Infinite content that is curated just for us. Infinite products. Infinite opinions. Infinite people who are somehow already into the exact niche thing you just discovered. With recent developments in AI, that volume and variance of content has grown exponentially to fill many whitespaces that may have previously existed.</p><p>It turns out that abundance doesn’t always translate into more freedom. More often, it leads to faster decisions. When everything is available, we tend not to choose carefully nor intently. We choose efficiently, following whatever feels easiest in the moment.</p><p>At internet scale, that changes what “choice” even means. And it isn’t just a go-to-market problem for startups. It’s a humanity problem. GTM just happens to be one of the clearest surfaces it shows up.</p><h3>Bounded Rationality, Now at Scale</h3><p>Modern society has always operated under the same constraints: limited time, limited attention, and limited mental energy. Nothing new about that.</p><p>What’s new is the volume and scale at which tracking and inference now operate.</p><h4>Bounded Rationality, Explained</h4><p>Bounded rationality doesn’t mean people make bad decisions. It means people make reasonable decisions under constraint.</p><p>Consider a few familiar examples.</p><h4>Pricing and subscriptions.</h4><p>Most users don’t read full terms. They scan, compare quickly, and move on. Systems learn which users abandon, which accept higher fees, which respond to urgency cues. Over time, interfaces adapt — not to maximize understanding, but conversion.<br>Different users don’t just pay different prices. They experience different levels of transparency, pressure, and optionality.</p><h4>Job discovery.</h4><p>When browsing roles, most people skim titles, glance at requirements, and apply where the fit feels obvious. Platforms learn which roles you pause on, which you skip, and how long you hesitate before applying. Over time, fewer unfamiliar or adjacent opportunities appear — not because you wouldn’t qualify, but because the system has learned you’re unlikely to stop.</p><h4>Explicit or highly stimulating content.</h4><p>A click driven by curiosity, boredom, or impulse gets registered as preference. Dwell time, replays, and escalation patterns are treated as intent. Feeds adapt accordingly. What began as a momentary choice hardens into a persistent signal, and eventually, a narrower menu.</p><p>This is the environment modern inference systems operate in — one shaped less by explicit intent than by accumulated shortcuts.</p><p>In each case, no single decision is irrational.<br>But at scale, human shortcuts become predictable inputs and that predictability becomes something systems can optimize around.</p><p>Every scroll puts hundreds or thousands of possible micro-decisions in front of you. Every hover, pause, or half-read headline creates a signal. Every system you interact with is collecting fragments of behavior to make sense of who you are.</p><p>No one can reason through this cleanly, so the systems step in.</p><p>They filter, rank, recommend, and then predict.</p><p>This is where “intent” shows up. Not as something discovered, but as something inferred.</p><p>Platforms don’t actually know what you want. They watch what you do, use proxies, build models, assign probabilities, and then treat those probabilities as good enough to act on.</p><p>That’s not sinister. It’s practical.</p><p>Without inference, the whole thing breaks under uncertainty.</p><h3>When Inference Starts Shaping the Menu</h3><p>At first, these systems feel helpful.</p><p>You see posts that resonate. Ads feel vaguely relevant. Your feed feels like it “gets” you.</p><p>But over time, something subtle happens.</p><p>What you click or hover affects what you’re shown.</p><p>What you’re shown affects what you click.</p><p>And what you never see quietly stops being an option.</p><p>You still have “agency” but it exists inside a corridor someone else designed.</p><p>That’s the part we tend to miss.</p><p>Most platforms don’t just respond to expressed intent. They narrow the set of choices before intent ever shows up. Over time, inference becomes the environment. Prediction becomes constraint.</p><p>No conspiracy required. Just feedback loops doing what feedback loops do.</p><h3>The Privacy Layer We Barely Acknowledge</h3><p>Most people still think privacy means secrecy.</p><ul><li>Is your name attached?</li><li>Are you logged in?</li><li>Is the data “anonymous”?</li></ul><p>That framing is outdated.</p><p>Modern tracking doesn’t need your identity or login to track you. It needs your behavior to be legible.</p><p>With recent advancements in AI, once behavior is legible, it becomes reproducible.</p><p>Across the cookie and tracker ecosystem:</p><p>•	timing patterns</p><p>•	device fingerprints</p><p>•	browsing sequences</p><p>•	dwell time</p><p>•	tiny, almost invisible interactions</p><p>These get stitched together across sites, vendors, and middlemen. Your name might be blurred, but your habits often aren’t.</p><p>That creates a quiet asymmetry.</p><p>You experience the internet as a person making choices.</p><p>The system experiences you as a probability distribution that persists across contexts.</p><p>That gap is where things get uncomfortable.</p><p>Not in a sci-fi, mind-control way. In a much more boring, effective way: nudging, shaping, extracting value from predictability.</p><h3>This Isn’t Doom or Paranoia</h3><p>To be clear, this isn’t an anti-tech rant.</p><p>It’s not anti-personalization.</p><p>It’s not a call to unplug from technology.</p><p>And it’s not primarily about governments watching you.</p><p>The more immediate issue is non-government, incentive-driven surveillance baked into ad tech, analytics, personalization, and monetization systems that most people never see or think about.</p><p>These systems don’t need to control you.</p><p>They just need to predict you a little better than random.</p><p>At scale, that’s plenty.</p><h3>Why Intentional Action Still Matters</h3><p>There’s no such thing as perfect privacy anymore. That ship is long gone.</p><p>What does exist is <strong>intentional action</strong>.</p><p>That looks like:</p><p>•	noticing when a choice is being nudged rather than chosen</p><p>•	slowing down where systems want speed</p><p>•	adding friction where platforms optimize for ease</p><p>•	recognizing that convenience is rarely free</p><p>You don’t need to reject platforms to do this. You just need to decide where agency actually matters to you.</p><p>What you subscribe to.</p><p>Where you spend attention.</p><p>When ease is worth the trade-off — and when it isn’t.</p><p>Intentional action isn’t about control.</p><p>It’s about selective resistance.</p><h3>Privacy as a Strategic Resource</h3><p>Privacy isn’t about hiding everything from everyone.</p><p>It’s about controlling how legible your behavior is to systems that don’t share your incentives.</p><p>That can mean:</p><p>•	reducing unnecessary tracking</p><p>•	limiting cross-site leakage</p><p>•	being honest about which tools trade convenience for visibility</p><p>•	separating contexts, identities, or devices when it actually matters</p><p>Think less secrecy, more strategic opacity.</p><p>Systems extract value by reducing uncertainty.</p><p>Individuals retain agency by not being fully reducible.</p><p>You don’t need to disappear.</p><p>You just need to be less predictable to systems that profit from predictability.</p><p>Privacy becomes less about hiding and more about preserving degrees of freedom.</p><h3>The Deeper Point</h3><p>Intent, whether in b2b or on the internet more broadly, is rarely a fact.</p><p>It’s a probability shaped by:</p><p>•	bounded rationality</p><p>•	feedback loops</p><p>•	incentives</p><p>•	and what’s allowed into view</p><p>Seeing that doesn’t remove agency. It restores it.</p><p>Because once you understand the system, you get to choose when to cooperate and when to push back.</p><h3>A More Honest Relationship With the Internet</h3><p>The goal isn’t to “beat” the algorithms.</p><p>It’s to stop mistaking inference for truth.</p><p>Treat recommendations as suggestions, not destiny.</p><p>Treat personalization as convenience, not insight.</p><p>Treat intent scores, feeds, and rankings as hypotheses but not a source of truth.</p><p>Your attention, behavior, and patterns are valuable. Not just to you, but to systems built to extract value from them.</p><p>Being intentional about where that value flows is one of the few levers we still have.</p><p>Not perfectly.</p><p>Not completely.</p><p>But enough to matter.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e51d9ff78601" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ The Growth Problems Nobody Warns You About at $1M ARR]]></title>
            <link>https://medium.com/@ravikhurana_38440/the-growth-problems-nobody-warns-you-about-at-1m-arr-b1a0b3d59c45?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/b1a0b3d59c45</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[customer-success]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[growth]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Thu, 25 Dec 2025 13:46:07 GMT</pubDate>
            <atom:updated>2025-12-25T13:46:07.221Z</atom:updated>
            <content:encoded><![CDATA[<h4>Reaching your first million in ARR and onboarding your first thousand users can feel like crossing a finish line. In reality, it’s the starting line for a new set of systemic growth problems that creep in under the surface.</h4><h3>1. The “Spreadsheet Olympics” Problem</h3><p>Every team pulls their own version of the truth:</p><ul><li>Finance has one churn number from Stripe exports.</li><li>Success tracks renewals in Google Sheets.</li><li>Sales uses HubSpot/CRM pipeline stages.</li><li>Product points at PostHog dashboards that are half configured.</li></ul><p>Suddenly you’ve got five different dashboards and no one trusts the numbers. Board decks turn into debate club.</p><h3>2. The Identity Crisis</h3><p>At small scale, it’s fine if “Acme Inc.” in Stripe doesn’t perfectly line up with “Acme Co.” in HubSpot. At 1k+ users and dozens of accounts, those mismatches explode:</p><ul><li>One user churns, but the account is still active.</li><li>Product shows growth, but billing shows contraction.</li><li>Leads get routed twice — or worse, to nobody.</li></ul><p>Without a shared identity map (user ↔ account ↔ subscription ↔ CRM record), every metric downstream is fuzzy.</p><h3>3. The Silent Failures</h3><p>Your Stripe sync fails overnight. Or PostHog’s export lags. Or someone renames a column in Supabase. Nobody notices until the CFO asks why ARR dropped 20% on the dashboard.</p><p>At this stage, you need clean data and <em>observability</em>. Freshness checks, reconciliation vs. billing totals, and some alerts when things go sideways.</p><h3>4. The “Growth Ceiling” Effect</h3><p>Without clarity on:</p><ul><li>Who your best-fit ICPs are,</li><li>Which cohorts retain and expand,</li><li>Where churn risk hides,</li></ul><p>… you’re basically driving with a foggy windshield. Sales sequences are wasted on bad-fit accounts, success spends time on users who will churn anyway, and marketing just throws darts and guesses attribution.</p><h3>💡 Enter: FullList Pipelines</h3><p>We’re building <strong>FullList Pipelines</strong> to solve these exact systemic problems for SaaS companies crossing the $1M ARR mark.</p><p>Instead of duct-taping Stripe, Supabase, PostHog, HubSpot, and Zendesk together, FullList Pipelines creates a <strong>golden record of your customers</strong>:</p><ul><li>Centralized <strong>identity resolution</strong> (no more “Acme vs. Acme Co.” debates).</li><li><strong>Event-time revenue modeling</strong> (churn, expansion, contraction, reactivation — all tied to actual billing events).</li><li><strong>Health scoring &amp; cohorts</strong> (usage + support + billing rolled into one view).</li><li><strong>Observability baked in</strong> (freshness checks, MRR reconciliation, Slack alerts).</li><li><strong>Activation loops</strong> (push “at-risk” accounts to HubSpot, expansion signals to Salesforce, ICP cohorts to marketing).</li></ul><p>The result? Finance, success, sales, and product teams finally look at the same truth. Growth feels systematic, not chaotic.</p><h3>🚀 Why This Matters</h3><p>At $1M ARR, growth isn’t about “more leads” or “more experiments.” It’s about <strong>removing friction from scaling</strong>.</p><ul><li>If finance, sales, and success argue over churn, you can’t plan confidently.</li><li>If you can’t tell which cohort expands, you waste quarters chasing the wrong ICP.</li><li>If your data breaks silently, you lose trust internally before you lose customers externally.</li></ul><p>FullList Pipelines cracks that. It gives you the <strong>operating system for growth </strong>so that the next $5M ARR comes from clarity, not chaos.</p><p>✨ <strong>Closing Thought:</strong><br>The first million ARR proves you have a product and some market.<br>The next millions prove you have a system.<br>FullList Pipelines is a way to help you build that system.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b1a0b3d59c45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building Trustworthy ICP Scoring: Why We’re Using NDCG to Validate AI-Powered Rankings]]></title>
            <link>https://medium.com/@ravikhurana_38440/building-trustworthy-icp-scoring-why-were-using-ndcg-to-validate-ai-powered-rankings-6eddf253e8bc?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/6eddf253e8bc</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[go-to-market]]></category>
            <category><![CDATA[ideal-customer-profile]]></category>
            <category><![CDATA[ndcg]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 19:40:47 GMT</pubDate>
            <atom:updated>2025-08-21T19:40:47.219Z</atom:updated>
            <content:encoded><![CDATA[<h4>How a search-engine metric helps us prove our AI isn’t just scoring accounts — it’s ranking them right.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rAvPvwVaebLz-8xUeBcyDQ.png" /></figure><p>Over the last year, there’s been a surge in AI-driven ICP scoring tools. With LLMs, enrichment APIs, and increasingly clever prompt engineering, it’s never been easier to build a scoring system that <em>looks</em> smart. But building one that reliably drives GTM outcomes? That’s a much harder problem.</p><p>At FullList, we’ve learned the hard way that if you don’t rigorously validate how your scoring model ranks accounts, you’re just shipping vibes — not pipeline. That’s why we’ve started using <strong>NDCG (Normalized Discounted Cumulative Gain)</strong> to measure whether our scoring model is doing what we <em>think</em> it’s doing: surfacing the most ICP-relevant, ready-to-engage accounts at the top.</p><h3>Why Traditional Scoring Metrics Fall Short</h3><p>Many teams rely on metrics like:</p><ul><li>Average precision</li><li>Accuracy thresholds (e.g. “was this account ICP or not?”)</li><li>Model confidence scores</li></ul><p>But these methods often treat ICP scoring as a <strong>classification problem</strong> — which ignores a critical truth:</p><p><em>Scoring is a </em><strong><em>ranking problem</em></strong><em>.</em></p><p>Your sales team doesn’t just want to know which accounts <em>might</em> be a fit. They want to know which ones to act on <strong>first</strong>.</p><h3>Why NDCG Works for Scoring ICP Fit</h3><p><strong>NDCG</strong> evaluates how well a model ranks items (accounts, in our case) relative to their true relevance, with more weight given to items near the top of the list. That matters because:</p><ul><li>Being in the <strong>top 5 accounts</strong> can make or break a rep’s week</li><li><strong>Small mis-rankings</strong> at the top cost real opportunity</li><li>Most scoring models are calibrated to optimize for classification — not rank position</li></ul><p><strong>How it works (simplified):</strong></p><ul><li>You assign ground truth relevance scores to a set of accounts (e.g. based on win/loss data, MQL conversions, or manual labels)</li><li>You run your scoring model and get predicted rankings</li><li>You compute how well the predicted rank order matches the ideal rank using the NDCG formula</li></ul><p>Scores closer to <strong>1.0</strong> mean your model is placing the right accounts at the top. Lower scores mean your model might be picking up noise — or hallucinating patterns.</p><h3>How We’re Implementing It at FullList</h3><p>We’re testing scoring models that include:</p><ul><li>LLM-based ICP description matching</li><li>Enrichment from LLM/scrapes, G2, job boards, and other external signals</li><li>Time-based triggers like funding, hiring, or reviews</li><li>Decay weighting on fresh signals</li><li>Chain-of-thought explanations for transparency</li></ul><p>To validate each version of the scoring logic, we:</p><ol><li>Assemble a labeled dataset of accounts with known outcomes (won/lost, responded/ignored)</li><li>Run scoring experiments with different weighting, decay, and enrichment settings</li><li>Measure <strong>NDCG@5</strong>, <strong>NDCG@10</strong>, and <strong>mean reciprocal rank</strong> across models</li><li>Compare scoring lift on pipeline conversion</li></ol><p>This lets us test not just whether the model is accurate — but whether it’s actually helping teams <strong>prioritize better</strong>.</p><h3>Alternatives to NDCG (And Why We Didn’t Choose Them)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6Pu9YyGWU3tpfVVUw2cOQQ.png" /></figure><p>NDCG is purpose-built for problems where:</p><ul><li>Ground truth has degrees of relevance (not just 0 or 1)</li><li>You care most about the top N items being correct</li><li>Mistakes at the top are worse than mistakes at the bottom</li></ul><p>Exactly what sales teams need.</p><h3>What’s Next</h3><p>We’re continuing to improve our scoring stack to:</p><ul><li>Pull fresh signals automatically and re-rank accordingly</li><li>Train feedback loops based on downstream outcomes</li><li>Let users compare scoring explanations across models</li></ul><p>Ultimately, our goal is to help GTM teams <strong>act on timing</strong> — not just fit. If we get the top 5 accounts right, everything downstream works better.</p><p>If you’re working on something similar — or struggling to validate your own scoring — we’d love to chat. This space is just getting started, and the bar for what counts as “smart” is rising fast.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6eddf253e8bc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[️ The Clay Ceiling — And the Future Beyond It]]></title>
            <link>https://medium.com/@ravikhurana_38440/%EF%B8%8F-the-clay-ceiling-and-the-future-beyond-it-01b46e99a844?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/01b46e99a844</guid>
            <category><![CDATA[gtm]]></category>
            <category><![CDATA[go-to-market]]></category>
            <category><![CDATA[growth-hacking]]></category>
            <category><![CDATA[product-led-growth]]></category>
            <category><![CDATA[clay]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Wed, 30 Jul 2025 23:53:28 GMT</pubDate>
            <atom:updated>2025-07-30T23:54:47.380Z</atom:updated>
            <content:encoded><![CDATA[<h3>🛠️ The Clay Ceiling — And the Future Beyond It</h3><h4>A Manifesto for AI-Native GTM Builders</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B12bDLdl8dkKn_HbmP8aaQ.png" /></figure><p>For the last year, I’ve lived deep inside the Clay ecosystem-building workflows, hitting credit limits, coaching teams, and iterating on AI-powered lead generation and enrichment pipelines. I’ve also been quietly building something of my own: a new kind of product for go-to-market teams-one that moves faster, costs less, and thinks more like a seller than a spreadsheet.<br>This is a manifesto for those of us who feel like we’re pushing against the Clay ceiling. And more importantly, for those building what comes after.</p><h3>I. Clay Isn’t Broken — But It’s Not the Answer Either</h3><p>Clay is a phenomenal sandbox. It gives you endless knobs to tweak, workflows to chain, and enrichments to run. But here’s the truth:</p><p>Clay is over-engineered — and most teams are under-qualified.</p><p>What begins as a way to save time becomes a new kind of technical debt. Every waterfall enrich, every nested formula, every workflow built on a gut feeling adds drag. And worse: it wastes money, erodes trust, and buries the signal you were trying to find in the first place.<br>It’s a shovel. What we need is a bulldozer.</p><h3>II. Every Credit Is a Contract</h3><p>Most people don’t treat Clay or OpenAI credits like currency. But they are.</p><p>Every credit you spend without intention is a step away from ROI.</p><p>I’ve seen teams burn through thousands of API credits before the first outbound email gets sent. I’ve also had to face down clients asking why their $5,000 bill didn’t drive $5,000 in value.<br>So I built systems: credit governance, approval layers, sandbox testing, incremental runs, error flagging. Because the difference between a good workflow and a great one is often not what it produces-but what it avoids.</p><h3>III. The AI Engineer Is the New RevOps</h3><p>What’s emerging in this moment is a new role: the AI/GTM Engineer.<br>It’s not dev. It’s not ops. It’s a strategic operator who knows:</p><ul><li>how to map data across tools</li><li>how to prompt with precision</li><li>how to evaluate models using frameworks like NDCG and Chain of Thought</li><li>how to string together AI systems that actually drive outcomes</li></ul><p><strong>Most importantly:</strong> this person is not waiting for a sales rep to give them a list. They’re building systems that generate the list, the message, the timing, and the insight-all at once.</p><h3>IV. Signals Decay. Workflows Shouldn’t.</h3><p>Clay gives you data. But it doesn’t tell you what to do with it. More often than not, it’s too late by the time the signal is acted on.</p><ul><li>Job change signals? Stale in 30 days</li><li>Funding rounds? A race to zero</li><li>Web intent? Fades by the hour</li></ul><p>We need systems that treat signals like perishable goods-with decay logic, real-time scoring, and adaptive playbooks. Because in modern GTM, the value of insight has a half-life. And if you’re not ready to move on it, someone else is.</p><h3>V. Dynamic ICPs &gt; Static Lists</h3><p>The future isn’t about building lists. It’s about recognizing patterns in real time.</p><p>Most reps still work off CSVs, waiting for RevOps to deliver the “next best account.” But in a world of behavioral data, CRM activity, product usage, and LLM-powered agents, we can do better.</p><p><strong>I’ve built tools that:</strong></p><ul><li>learn from CRM + Gong data to reverse engineer your best customers</li><li>convert that insight into dynamic prompts and active segments</li><li>surface only the accounts that are showing fresh, composite intent</li></ul><p>This is how we escape spreadsheet hell. This is how we win the signal game.</p><h3>VI. The Sales Process Is About to Flip</h3><p>In 5 years, outbound will look unrecognizable.</p><ul><li>Buyers will navigate LLM-powered discovery journeys.</li><li>Vendors will be recommended, not hunted.</li><li>Salespeople will be connectors, not cold-callers.</li></ul><p>The self-directed buying journey already exists. We’re just not optimizing for it yet. The winners will be the companies that meet buyers in the moment of curiosity-not with a spammy email, but with a signal-backed conversation and a trusted face.</p><h3>VII. Product Analytics Is the Next Frontier for GTM</h3><p>Want to know who’s actually in-market?</p><p>Look at your product, not your CRM.</p><p>Clicks, scrolls, revisits, session length, bounce patterns-these are the new intent signals. And the companies that plug this feedback loop into sales outreach will have the edge.</p><p>I’m already integrating these data points into FullList. Because your next best customer has probably already told you they’re ready. You just haven’t been listening closely enough.</p><h3>VIII. Clay Was a Start. What Comes Next Is the System.</h3><p>I’m not here to bash Clay. It taught me a lot. But it also inspired me to build something better-because I believe the future of go-to-market belongs to systems, not templates.</p><p><strong>Systems that:</strong></p><ul><li>respect your budget</li><li>learn from your best customers</li><li>adapt in real time</li><li>push your team toward action, not analysis</li></ul><p>That’s what I’m building. That’s what I want feedback on. And that’s the community I’m trying to co-create-with you.</p><p>If you’ve felt any of this friction, seen these cracks, or started to build your own escape hatch — let’s talk.</p><p>📍FullList - <a href="https://fulllist.ai/">fulllist.ai</a><br> 🗓 <a href="https://meetings.hubspot.com/ravi-khurana/demo?uuid=9aed5f54-43a6-4b6b-ad4f-d344943473ed"><strong>Book a convo with me</strong><br></a> 📬 Or just <a href="https://www.linkedin.com/in/ravi-khurana-b4891414b/"><strong>DM me on LinkedIn</strong></a><strong>.</strong></p><p>I’m always down to jam. Let’s build what comes after Clay.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=01b46e99a844" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How We Secured A Vibe-Coded App: Real Fixes for Real Vulnerabilities]]></title>
            <link>https://medium.com/@ravikhurana_38440/how-we-secured-a-vibe-coded-app-real-fixes-for-real-vulnerabilities-104a7d9b4284?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/104a7d9b4284</guid>
            <category><![CDATA[information-security]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[database-security]]></category>
            <category><![CDATA[lean-startup]]></category>
            <category><![CDATA[startup-lessons]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Fri, 25 Jul 2025 20:01:48 GMT</pubDate>
            <atom:updated>2025-08-10T17:45:57.840Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>This is a quick guide on how I patched a few critical security issues in my Next.js/Supabase-based app</strong> — <strong>including search_path exploits, API key leaks, and input validation gaps.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vv-RGn9kj4s4jOlduZXdqg.png" /></figure><p>Just going to start off by saying security isn’t a vibe. But it <em>is</em> a necessity.</p><p>If you’ve been building fast with tools like Supabase, Lovable, Next.js, or Vercel — especially in a solo or small team environment — you’ve probably focused on features, design, and getting things live. We did too. But as traction grew, so did our responsibility to protect users, data, and our platform.</p><p>Here’s how we identified common vulnerabilities in modern full-stack “vibe-coded” apps — and what we did to fix them.</p><h3>🔐 Phase 1: Fortifying the Database</h3><h3>🚨 The Risk:</h3><p>By default, PostgreSQL’s search_path setting can be manipulated in a way that malicious users could escalate access or interfere with function execution. Combine this with improperly scoped SECURITY DEFINER functions, and you’ve got a potential privilege escalation vector.</p><h3>✅ What We Did:</h3><ul><li><strong>Set </strong><strong>search_path explicitly</strong> in every function to public, pg_temp</li><li>Used <strong>SECURITY DEFINER</strong> with the correct owner to isolate privileges</li><li>Reviewed and locked down function ownership and roles</li></ul><p><strong>💡 Result:</strong> All our database functions are now immune to search_path manipulation attacks.</p><h3>🔑 Phase 2: Securing API Keys</h3><h3>🚨 The Risk:</h3><p>Storing API keys in localStorage is a huge red flag. They’re accessible via JavaScript and vulnerable to XSS attacks. Many devs do this out of convenience—but the tradeoff isn&#39;t worth it.</p><h3>✅ What We Did:</h3><ul><li><strong>Removed all API key storage from localStorage</strong></li><li><strong>Encrypted and stored API keys server-side</strong> in the database</li><li><strong>Validated key formats</strong> based on the provider (OpenAI, LeadMagic, CrustData, etc.)</li><li><strong>Base64-encoded stored keys</strong> to add an extra layer (yes, encoding isn’t encryption, but it helps reduce accidental exposure)</li></ul><p><strong>💡 Result:</strong> No secrets live on the frontend. We treat keys like passwords now.</p><h3>🧼 Phase 3: Input Validation Everywhere</h3><h3>🚨 The Risk:</h3><p>Unchecked input — especially uploads — is a goldmine for attackers. From XSS-laden CSVs to disguised executables, we found too many gaps.</p><h3>✅ What We Did:</h3><ul><li>Built a <strong>central validation utility</strong> for: <br>- Email formats<br>- URLs <br>- Freeform text <br>- File uploads (type, size, suspicious names)</li><li>For CSVs, we:<br>- <strong>Sanitized content<br>- </strong>Escaped potentially dangerous formulas (e.g., =cmd|&#39; /C calc&#39;!A0)<br>- Blocked HTML/JS injections</li></ul><p><strong>💡 Result:</strong> Uploaded files and user inputs are no longer a liability. Hackers can’t inject malicious code into CSVs that would run on a user machine when downloaded.</p><h3>🚧 Phase 4: Secure Error Handling</h3><h3>🚨 The Risk:</h3><p>Verbose error logs in production often leak secrets — like database errors, failed auth tokens, or stack traces. Perfect for a bug report. Also perfect for an attacker.</p><h3>✅ What We Did:</h3><ul><li>Cleaned up <strong>all production error responses</strong> to remove stack traces and sensitive info</li><li>Added <strong>structured logging</strong> (with request metadata and trace IDs)</li><li>Ensured no secrets are ever included in error messages</li></ul><p><strong>💡 Result:</strong> Our users see helpful errors. We see useful logs. Hackers see nothing.</p><h3>🎯 Phase 5: Defending Against Prompt Injection</h3><h3>🚨 The Risk:</h3><p>If your app pulls in untrusted text (from LinkedIn profiles, web pages, uploaded files, etc.) and passes it straight into an LLM with connected tools, that text can hide malicious instructions.<br>This is called <strong>prompt injection</strong> — and it can trick the model into:</p><ul><li>Querying sensitive data it shouldn’t</li><li>Sending API keys or secrets to an attacker</li><li>Triggering expensive API calls or data deletions</li></ul><p>Unlike traditional exploits, there’s no “vulnerability” in the code — the model is just following instructions you didn’t mean to give.</p><h3>✅ What We Did:</h3><ul><li><strong>Prompt Isolation:</strong> We wrap all untrusted content in a structured “data” section, with system prompts telling the model not to execute instructions found in that section.</li><li><strong>Tool Whitelisting:</strong> The model can only call specific functions with tightly validated input schemas.</li><li><strong>No Direct Env Access:</strong> The LLM never sees raw API keys or env vars. All keys are injected server-side into tool code.</li><li><strong>Sandboxed Execution:</strong> Any tool that could run code or access data runs in a secure, ephemeral container with no secrets.</li></ul><p>💡 <strong>Result:</strong> Even if an attacker hides instructions in scraped content, our LLM can’t touch sensitive data, run arbitrary commands, or leak secrets.</p><h3>🛠️ What’s Left</h3><p>Most critical issues are resolved — but a few <strong>Supabase platform settings</strong> still need tightening:</p><ul><li>OTP expiry settings (reduce from default to 600 seconds)</li><li>Leaked password protection (if we go past Google and MagicLink Auth)</li></ul><p>These are on our radar and will be handled next.</p><h3>🎯 Final Thoughts</h3><p>Security doesn’t have to be overwhelming, but it <strong>does</strong> require intentionality. If you’re building with modern tools and shipping fast, do yourself (and your users) a favor: pause and review your threat surface.</p><p>Even if your app <em>feels</em> secure, vibe-coded vulnerabilities are real — and we hope this checklist helps you spot them early.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=104a7d9b4284" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Art of LLM Context Management: Optimizing AI Agents for App Development]]></title>
            <link>https://medium.com/@ravikhurana_38440/the-art-of-llm-context-management-optimizing-ai-agents-for-app-development-e5ef9fcf8f75?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5ef9fcf8f75</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[ai-development]]></category>
            <category><![CDATA[ai-development-company]]></category>
            <category><![CDATA[vercel]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Thu, 06 Feb 2025 06:04:54 GMT</pubDate>
            <atom:updated>2025-02-06T06:04:54.480Z</atom:updated>
            <content:encoded><![CDATA[<h4>A detailed guide on how to optimize memory and context infrastructure if you’re building AI developer tools.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/870/1*Os5OkTJvSpWUq7ZfuP0_CQ.png" /></figure><p>How can we build an AI assistant that’s not only smart enough to understand your coding needs but also agile enough to handle long-winded chat histories without breaking a sweat. Whether you’re developing with tools like v0, Lovable, or Cursor, one challenge persists: how to juggle extensive context windows without overwhelming the system. Today, we’re diving into a few innovative approaches to tackle this issue — optimizing token usage, slashing latency, and preserving the rich tapestry of conversation history that makes your AI truly intelligent.</p><h3>The Context Conundrum</h3><p>AI models, powerful as they are, have inherent limits on how many tokens they can process in a single request. As your conversation with the AI grows — encompassing evolving code, debugging chats, and brainstorming sessions — feeding everything into the model becomes inefficient. You might end up with high latency, wasted tokens, or even context loss. So, how do we keep our AI agent both responsive and smart?</p><h3>Method 1: Vectorizing Prior Chat Context</h3><h4>What’s the Big Idea?</h4><p>Instead of feeding every snippet of chat history into your AI, convert key parts of the conversation into embeddings — compact numerical representations of text. These vectors are stored in a database, and only the most relevant pieces are retrieved when needed.</p><h4>Why It Rocks</h4><ul><li><strong>Token Efficiency:</strong> Only the golden nuggets of conversation are sent to the AI, saving tokens.</li><li><strong>Dynamic Retrieval:</strong> A similarity search brings back context that matters most to the current query.</li><li><strong>Scalability:</strong> No matter how long the conversation, you’re never bogged down by excessive tokens.</li></ul><h4>Points to Ponder</h4><p>Setting up a vector database (think FAISS, Pinecone, or Weaviate) adds some complexity. There’s also the challenge of ensuring your retrieval algorithm picks the right pieces of context. But with fine-tuning, this method is a win for long-term memory.</p><h3>Method 2: The Running Code File</h3><h4>What’s the Big Idea?</h4><p>Separate your evolving code from the chat by maintaining a single “living” code file. Every new change gets appended. When code is removed, it’s simply commented out. The AI or user sees only the active code, with line numbers keeping everything in context.</p><h4>Why It Rocks</h4><ul><li><strong>Decoupled State Management:</strong> The code file serves as the ground truth, independent of the conversation.</li><li><strong>Versioning-Friendly:</strong> It’s like having a mini version control system built right into your AI’s brain.</li><li><strong>Clarity:</strong> Users get a clean view of what’s active without digging through chat logs.</li></ul><h4>Points to Ponder</h4><p>Relying on line numbers can become fragile if the code is refactored extensively. And while this method handles code evolution beautifully, it doesn’t solve the broader conversation context challenge.</p><h3>Method 3: The Sliding Window Hybrid</h3><h4>What’s the Big Idea?</h4><p>Marry the strengths of direct context with vectorized storage. Maintain a sliding window of the most recent chat history for immediate context, while older parts are safely stored as vectors for retrieval when necessary.</p><h4>Why It Rocks</h4><ul><li><strong>Balanced Approach:</strong> Immediate context is always available, and you can summon older memories on demand.</li><li><strong>Optimized Token Usage:</strong> The sliding window keeps the conversation fresh without overwhelming the model.</li><li><strong>Flexibility:</strong> Dynamic retrieval means nothing gets lost in the shuffle.</li></ul><h4>Points to Ponder</h4><p>This hybrid approach introduces synchronization challenges. If the boundaries between recent and vectorized context aren’t well defined, you might face subtle discontinuities. But with careful design, the benefits far outweigh the risks.</p><h3>Method 4: Prompt Caching — Speed Meets Efficiency</h3><h4>What’s the Big Idea?</h4><p>Leverage prompt caching to store responses or prompt fragments so you can reuse them instead of regenerating them every time. Some platforms, like Anthropic’s Claude via Bedrock, offer this out-of-the-box.</p><h4>Why It Rocks</h4><ul><li><strong>Lightning Fast:</strong> Cached responses reduce latency dramatically.</li><li><strong>Token Savings:</strong> Reuse familiar prompt segments and avoid wasting tokens.</li><li><strong>Engineering Ease:</strong> Built-in caching features reduce the complexity of your system.</li></ul><h4>Points to Ponder</h4><p>Caching is great for static or infrequently changing content, but it might fall short when the conversation evolves rapidly. There’s also the risk of vendor lock-in if you depend too heavily on a specific provider’s caching system.</p><h3>Bringing It All Together</h3><p>No single method is a silver bullet. Instead, the key lies in combining approaches to create a robust, scalable system. Here’s a potential roadmap for your AI development:</p><ol><li><strong>Immediate Context:</strong> Use a sliding window for the most recent and relevant conversation bits.</li><li><strong>Deep Memory:</strong> Store older context as vectors to retrieve only what’s necessary.</li><li><strong>Code Evolution:</strong> Maintain a running code file to manage the state of your project independently.</li><li><strong>Optimize Performance:</strong> Leverage prompt caching to streamline repetitive tasks and reduce latency.</li></ol><p>By thoughtfully integrating these methods, you can build an AI agent that’s not just smart, but also efficient, agile, and capable of evolving alongside your projects.</p><h3>Final Thoughts</h3><p>Building a truly intelligent AI assistant for app development is as much about managing context as it is about the UI. With strategies like vectorizing chat history, maintaining a running code file, using a sliding window approach, and employing prompt caching, you’re well on your way to overcoming token limitations and latency issues. The result? A seamless, responsive, and context-aware AI that makes coding feel like a breeze.</p><p>So, the next time you’re wrestling with long chat threads and sprawling code histories, remember: a little context management goes a long way.</p><p>Happy coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5ef9fcf8f75" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Leveraging Insider & Institutional Trading Data with Forecast Models to Identify Quality Stock…]]></title>
            <link>https://medium.com/@ravikhurana_38440/leveraging-insider-institutional-trading-data-with-forecast-models-to-identify-quality-stock-93c33cfe21a4?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/93c33cfe21a4</guid>
            <category><![CDATA[copy-trading]]></category>
            <category><![CDATA[hedge-funds]]></category>
            <category><![CDATA[wealth-management]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[stock-trading]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Wed, 05 Feb 2025 21:46:40 GMT</pubDate>
            <atom:updated>2025-02-06T17:02:28.674Z</atom:updated>
            <content:encoded><![CDATA[<h3>Leveraging Insider &amp; Institutional Trading Data with Forecast Models to Identify Quality Stock Investments</h3><h4>How I utilize PyApps to pull publicly available 13F Filings directly from the SEC API and a variety of Facebook’s Prophet forecasting models to de-risk and validate some of my stock investments.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CYeLQXS4a1-NtlnHsWb8fg.jpeg" /></figure><h3>Introduction</h3><p>Investing is a game of information and timing. Institutional trading activity, insider buys and sells, and major news events can offer strong signals about a stock’s future performance. But how can investors systematically analyze these signals and use them to validate an entry point? This is where PyApps has helped me out. These PyApps dashboards help me track insider transactions and forecast price movements to make more informed decisions than trading naked.</p><h3>The Power of Insider and Institutional Trading Activity</h3><p>Insider trading, when done legally, provides crucial insights into how executives, politicians, and major shareholders feel about a company’s future. Large insider purchases can indicate confidence, while insider sales may raise red flags. Institutional trading — actions taken by hedge funds, asset managers, and big financial players — can also provide critical buy and sell signals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*ueQXKNWjT-Kb0YnXtAeSxA.png" /><figcaption>Visual depiction of the large insider and institutional stock purchases and sales of OXY</figcaption></figure><p>For example, a recent analysis of Occidental Petroleum (OXY) stock revealed large purchases by Berkshire Hathaway. The insider buy transactions on December 17, 2024, included multiple high-value purchases:</p><ul><li>2,477,362 shares at $46.25 (<strong>$114.58M total</strong>)</li><li>2,430,924 shares at $46.16 (<strong>$112.20M total</strong>)</li><li>3,372,309 shares at $45.55 (<strong>$153.59M total</strong>)</li></ul><p>Tracking such activity in real-time can provide investors with an opportunity to ride the wave of institutional accumulation before prices move significantly.</p><h3>Significant Company Announcements and Research Insights</h3><p>In addition to analyzing insider and institutional trading activities, staying informed about significant company announcements is crucial for making informed investment decisions.</p><p>For instance, GlaxoSmithKline (GSK) recently announced promising results from a study involving their investigational drug, <strong>Dostarlimab</strong>, for rectal cancer treatment. In a groundbreaking clinical trial, all participating patients achieved complete remission, demonstrating the potential efficacy of this therapy. Such unprecedented clinical success could serve as a major catalyst for stock appreciation, boosting investor confidence and paving the way for market expansion.</p><p>By integrating news of these developments with insider and institutional trading data, investors can gain a more comprehensive understanding of a company’s prospects.</p><p>Furthermore, hedge funds and institutional investors have been accumulating GSK shares throughout 2024, leading up to the FDA approval of their drug. Notable funds such as <strong>Dodge &amp; Cox, Fisher Asset Management, Primecap Management, and Bridgewater Associates (led by Ray Dalio)</strong> have increased their holdings, signaling confidence in the company’s future performance. <strong>Bridgewater Associates</strong>, in particular, increased its GSK position by <strong>107.34%</strong>, bringing its total holdings to <strong>965.84k shares valued at $39.48M</strong>. With the drug now set to enter the market, these funds may benefit from the profits, potentially driving further stock price appreciation. Monitoring this accumulation activity provides an added layer of insight when considering investment opportunities.</p><h3>Forecasting Analysis to Validate Entry Points</h3><p>While insider and institutional activity can indicate sentiment, it’s equally important to validate investment decisions with forecasting models. PyApps leverages machine learning models to predict stock price movements and assess risk. The forecasting models used in these charts are based on <strong>Facebook Prophet</strong>, ensuring a well-rounded predictive approach.</p><p>For example, a recent forecast for <strong>GSK</strong> (GlaxoSmithKline) provided the following insights:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*w45uGvgzfciKjhdAKls_iw.png" /><figcaption>Simple forecast model for GSK Stock</figcaption></figure><ul><li><strong>Current Price:</strong> $34.84</li><li><strong>Forecasted Price:</strong> $37.87</li><li><strong>Expected Change:</strong> +8.69%</li><li><strong>Price Range (Next 90 Days):</strong> $33.38 — $38.86</li><li><strong>Trend Direction:</strong> Upward</li><li><strong>Trend Strength:</strong> Moderate</li><li><strong>Model Accuracy (MAPE):</strong> 1.05%</li></ul><p>Similarly, for <strong>OXY</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*JHftUyU6HeQfyR5iRiuLMQ.png" /><figcaption>Simple forecast model for OXY Stock</figcaption></figure><ul><li><strong>Current Price:</strong> $47.59</li><li><strong>Forecasted Price:</strong> $55.76</li><li><strong>Expected Change:</strong> +17.17%</li><li><strong>Price Range (Next 60 Days):</strong> $44.27 — $57.96</li><li><strong>Trend Direction:</strong> Upward</li><li><strong>Trend Strength:</strong> Strong</li><li><strong>Model Accuracy (MAPE):</strong> 1.17%</li></ul><h3>The Risk of 13F Latency and Fine-Tuning Forecasts</h3><p>Institutional trading data, such as <strong>13F filings from the SEC</strong>, provides valuable insights into fund positions. However, a key limitation of 13F data is its <strong>latency</strong> — since filings are submitted quarterly, the reported positions may already be outdated by the time they are published.</p><p>Additionally, 13F filings may not encompass <strong>options and other derivative positions</strong> that large investors use to hedge their risk. Hedge funds and institutional investors often employ complex strategies that involve calls, puts, and synthetic positions that may not be fully reflected in the equity holdings reported in 13Fs. This means that while a firm may appear to be accumulating a position, they could simultaneously be hedging against downside risk using options, making it crucial to analyze additional data points beyond just 13F filings.</p><p>To address this risk, investors should supplement 13F data with more <strong>real-time indicators</strong>, such as option positions, earnings reports, and other AI-driven forecasting models. By integrating <strong>ARIMA, SARIMA, and SARIMAX</strong> models or custom neural-network models alongside tools like Facebook Prophet, investors can fine-tune their predictions and reduce the reliance on stale institutional data. This approach enhances the accuracy of market timing decisions and improves overall risk management.</p><h3>Combining Data for High-Quality Investments</h3><p>The true power of PyApps lies in its ability to merge multiple sources of data. By analyzing insider/institutional trades, recent news, and AI-powered price forecasts together, investors can:</p><ol><li><strong>Spot Opportunities Early</strong> — Insider and institutional buying often precedes major price movements.</li><li><strong>Confirm the Strength of a Trend</strong> — Forecasting models help validate whether an uptrend is sustainable.</li><li><strong>Minimize Risk</strong> — Understanding price range projections ensures smarter entries and exits.</li></ol><p>For example, in the OXY case, Buffett’s Berkshire Hathaway making significant purchases, combined with a strong upward trend in forecasting analysis, provided a compelling investment opportunity.</p><h3>Conclusion</h3><p>In an era of overwhelming financial data, the ability to synthesize multiple signals has been key to improving my investment strategies. PyApps helped me track critical insider and institutional activity while leveraging easy to use forecast models to confirm trade setups. Try it at your own risk or paper trade your findings — this is not financial advice!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=93c33cfe21a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scouting Early-Stage Opportunities: A Research Approach For Any Niche]]></title>
            <link>https://medium.com/@ravikhurana_38440/scouting-early-stage-opportunities-a-research-approach-for-any-niche-4a7daf32d128?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a7daf32d128</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[early-stage-investing]]></category>
            <category><![CDATA[startup-research]]></category>
            <category><![CDATA[product-market-fit]]></category>
            <category><![CDATA[market-research]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Mon, 06 Jan 2025 23:09:23 GMT</pubDate>
            <atom:updated>2025-01-06T23:09:23.951Z</atom:updated>
            <content:encoded><![CDATA[<h4>A guide for early-stage VC Analysts to filter out the noise and find stand-out startups in any industry/niche.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pCuSMOMeIxbuODOY4E4tag.png" /></figure><h3>Introduction</h3><p>Early stage venture building is more competitive than it has ever been. Whether it’s AI session replay tools, fintech platforms, or healthcare analytics — identifying early-stage opportunities requires a structured approach. This article breaks down a repeatable research framework that investors can apply across niches to <strong>uncover emerging players who are solving real market pain points</strong>.</p><h3>Steps to finding the stand-out startups</h3><p>To demonstrate, we’ll use <strong>“AI-powered session replay tools”</strong> as an example. These tools help businesses understand user behavior, optimize performance, and improve customer experiences. While the examples here are specific to this market, the methodology can be applied universally.</p><h4><strong>Step 1: Define the Niche</strong></h4><p>For this example, we focused on AI-powered session replay tools — a category bridging user behavior analytics and predictive troubleshooting. These tools are widely used by product managers, developers, and marketers to understand how users interact with digital platforms.</p><h4><strong>Step 2: Identify Established Players Using Semantic Search</strong></h4><p>Using advanced semantic search APIs such as Exa, we compiled a list of established and emerging players in the AI session replay space. Here are a few that we wanted to look further into:</p><ul><li><strong>LogRocket</strong></li><li><strong>OpenReplay</strong></li><li><strong>Replay.io</strong></li><li><strong>Highlight.io</strong></li><li><strong>CtrlStack (QueryPal)</strong></li><li><strong>Intellize.ai</strong></li></ul><p>Each of these players showcase unique strengths and gaps worth exploring.</p><h4><strong>Step 3: Analyze Customer Feedback for Patterns</strong></h4><p>We turned to platforms like G2 and TrustRadius to identify recurring customer pain points. Across industries, aggregating these reviews helped us highlight universal themes. In our AI session replay example, these patterns emerged:</p><ul><li>Limited customization or flexibility</li><li>Pricing that doesn’t fit smaller teams</li><li>Steep onboarding curves for less-technical users</li><li>Integration challenges with other tools</li></ul><h4><strong>Step 4: Identify Opportunities from Pain Points</strong></h4><p>Customer feedback often reveals specific opportunities for innovation. Common themes include:</p><ul><li>Customizable feature sets</li><li>Granular real-time alerts</li><li>Better integration with backend systems</li><li>Easier onboarding experiences</li></ul><p>These findings act as a roadmap for identifying startups that are poised to address these gaps in the market.</p><h4><strong>Step 5: Scout Emerging Players</strong></h4><p>Using another layer of semantic search, we looked for super early-stage startups addressing these specific gaps. One standout example is <strong>Decipher (getdecipher.com)</strong>, which tackles many of the identified pain points:</p><ul><li><strong>Customizable Session Management:</strong> Tailoring session lengths for better flexibility.</li><li><strong>AI-Powered Insights:</strong> Proactive issue detection and resolution.</li><li><strong>Enhanced Integration:</strong> Strong backend and frontend system connections.</li><li><strong>Simplified Onboarding:</strong> User-friendly setup processes.</li></ul><h3><strong>Why Decipher Stood Out (and Why This Approach Works)</strong></h3><p>Decipher’s strengths align directly with the recurring gaps identified from customer reviews. This structured approach — identifying key players, analyzing feedback, and scouting for specific early-stage solutions — focuses on identifying founders that have a unique understanding of the challenges that target users face.</p><h3><strong>Conclusion:</strong></h3><p>Whether you’re exploring AI tools, B2B SaaS platforms, or emerging tech in healthcare, this framework offers a repeatable method to scout investment opportunities. The key lies in systematically identifying market gaps and discovering startups that are uniquely positioned to address them.</p><p>In our example, <strong>Decipher</strong> stood out in the AI session replay space — but the approach itself is universal.</p><p>If you’re an investor exploring emerging opportunities, try out this research framework and let me know if you find anything interesting. And if AI session replay tools are on your radar, check out <a href="https://getdecipher.com/">getdecipher.com</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a7daf32d128" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Job Boards: How Real Connections Unlock Opportunities That Don’t Exist (Yet)]]></title>
            <link>https://medium.com/@ravikhurana_38440/beyond-job-boards-how-real-connections-unlock-opportunities-that-dont-exist-yet-fbd5951642c4?source=rss-9821e85972b1------2</link>
            <guid isPermaLink="false">https://medium.com/p/fbd5951642c4</guid>
            <category><![CDATA[recruiting]]></category>
            <category><![CDATA[tech-jobs]]></category>
            <category><![CDATA[networking]]></category>
            <category><![CDATA[job-search]]></category>
            <category><![CDATA[job-hunting]]></category>
            <dc:creator><![CDATA[Ravi Khurana]]></dc:creator>
            <pubDate>Wed, 01 Jan 2025 03:05:03 GMT</pubDate>
            <atom:updated>2025-01-01T18:42:20.732Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why the best job opportunities aren’t on the job boards — and how building meaningful connections can unlock roles that don’t even exist yet.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*48a-4mz96ZxMYel9slbxyg.png" /></figure><h3>The Hidden Truth About Job Searching</h3><p>For most job seekers, the search starts in the same place: job boards. You scroll endlessly through listings, tailor your resume for every role, hit “apply,” and hope for the best. But despite your diligence, responses are rare, and meaningful opportunities feel few and far between.</p><p>The reality is harsh but simple: <strong>the best opportunities are often not listed publicly, and when they are, they’re already halfway filled before you ever see them.</strong></p><h3>🚨 The Pitfalls of Traditional Job Searching</h3><h3>1. The Hidden Job Market</h3><ul><li>Research consistently shows that <strong>70–80% of jobs are never advertised publicly</strong>.</li><li>These roles are filled through referrals, internal promotions, or direct outreach from recruiters.</li></ul><h3>2. The Resume Black Hole</h3><ul><li>For every job posting, hundreds — if not thousands — of applications are submitted.</li><li>Applicant Tracking Systems (ATS) filter out resumes based on arbitrary keyword matches.</li></ul><h3>3. Late to the Conversation</h3><ul><li>By the time a job hits a public job board, hiring managers often already have internal referrals, preferred candidates, or even someone they’re in the final stages of negotiating with.</li><li>Public job posts are frequently a procedural step rather than a genuine call for talent.</li></ul><h3>4. The Algorithmic Wall</h3><ul><li>Many candidates are rejected before their application ever reaches human eyes.</li><li>Minor formatting errors, a lack of specific keywords, or even the timing of submission can mean instant disqualification.</li></ul><h3>🤝 Why Connections Outperform Applications</h3><p>If the traditional system is flawed, what works better? The answer lies in <strong>strategic human connections</strong>.</p><h3>1. Advocates Get You Past the Gatekeepers</h3><ul><li>An internal advocate — someone who already works at the company — can bypass the ATS entirely.</li><li>Their recommendation puts your profile directly in front of hiring managers.</li></ul><h3>2. Conversations Precede Job Descriptions</h3><ul><li>Many roles begin as problems, not job postings.</li><li>If you’re in the right conversation at the right time, roles can be <em>shaped</em> around your skills and expertise.</li></ul><h3>3. Advocates Provide Insider Insights</h3><ul><li>What is the company really looking for in this role?</li><li>How can you position yourself in alignment with their priorities?</li><li>These insights rarely exist in public job descriptions but flow freely in conversations with insiders.</li></ul><h3>4. Timing is Everything</h3><ul><li>Connections often lead to hearing about opportunities <em>before</em> they become public.</li><li>You’re no longer reacting to postings — you’re part of the process as roles are created.</li></ul><h3>🛠️ The Advocate-First Approach: A Smarter Way to Job Search</h3><p>Shifting from passive applications to an <strong>advocate-first approach</strong> isn’t about “networking” in the traditional, transactional sense. It’s about building authentic, purposeful relationships that align with your career goals.</p><h3>Step 1: Identify Target Companies</h3><ul><li>Start with a list of companies that align with your values, career aspirations, and preferred industry.</li><li>Don’t limit yourself to those with open job postings — focus on <em>fit</em> first.</li></ul><h3>Step 2: Map Your Network</h3><ul><li>Look for existing connections: alumni, former colleagues, or mutual LinkedIn contacts.</li><li>Even second-degree connections can lead to powerful introductions.</li></ul><h3>Step 3: Initiate Meaningful Outreach</h3><ul><li>Don’t ask for a job — ask for insights/feedback/advice.</li><li>Questions like, <em>“What’s the company’s biggest focus right now?”</em> or <em>“What do you wish you knew when you started there?”. </em>These questions shift the conversation from <em>“What can you do for me?”</em> to <em>“How can I learn from you and add value?”</em> — a mindset that builds more authentic connections.</li></ul><h3>Step 4: Listen and Learn</h3><ul><li>Pay attention to challenges, opportunities, and themes in your conversations.</li><li>Often, advocates will surface opportunities that aren’t public yet — or suggest ways you could fill a gap they hadn’t considered before.</li></ul><h3>Step 5: Position Yourself as a Problem-Solver</h3><ul><li>Frame your skills and experience in the context of the company’s goals and challenges.</li><li>Be proactive: suggest ways you could contribute to their team, even if a role doesn’t formally exist yet.</li></ul><h3>Step 6: Follow Up and Stay Visible</h3><ul><li>Building relationships takes time.</li><li>Keep advocates updated on your job search progress and offer value where you can (e.g., share relevant articles, insights, or introductions).</li></ul><h3>📊 Why This Approach Works Better</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t8U94ceBZGfNK0Mu3p0N8w.png" /></figure><h3>🔑 Start Thinking Like a Value Creator, Not Just an Applicant</h3><p>Companies hire to solve problems, not to tick boxes on a job description. When you approach your job search as a problem-solver rather than just an applicant, you unlock opportunities that simply wouldn’t exist otherwise.</p><ul><li><strong>Start conversations, not just applications.</strong></li><li><strong>Focus on relationships, not transactions.</strong></li><li><strong>Look beyond job boards to where real opportunities are shaped: in conversations, meetings, and moments of connection.</strong></li></ul><h3>🧠 The Takeaway</h3><p>Your next opportunity might not be listed on a job board. It might not even <em>exist yet</em>. But with the right approach — one centered on meaningful connections, strategic outreach, and a problem-solving mindset — you can uncover or even <em>create </em>your dream role.</p><p>The future of job searching isn’t about <em>applying more</em>. It’s about <em>connecting better</em>. I hope this approach helps you out and feel free to reach out to me if you want to try our free JobSwiper that suggests the best people to reach out to from your target companies!</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.loom.com%2Fembed%2Feb4592b0d2ff4ccc9c0a69e1928cdaf6&amp;display_name=Loom&amp;url=https%3A%2F%2Fwww.loom.com%2Fshare%2Feb4592b0d2ff4ccc9c0a69e1928cdaf6%3Fsid%3Dbae1f29d-7ad9-49da-aae8-cb1b326130f1&amp;image=https%3A%2F%2Fcdn.loom.com%2Fsessions%2Fthumbnails%2Feb4592b0d2ff4ccc9c0a69e1928cdaf6-3133b185d3eac469.gif&amp;type=text%2Fhtml&amp;schema=loom" width="1659" height="1244" frameborder="0" scrolling="no"><a href="https://medium.com/media/34378c43cd028ae2e578ef17e35525ae/href">https://medium.com/media/34378c43cd028ae2e578ef17e35525ae/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fbd5951642c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>