<?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 Zhong Hong on Medium]]></title>
        <description><![CDATA[Stories by Zhong Hong on Medium]]></description>
        <link>https://medium.com/@zhonghong9998?source=rss-c99fe335779c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dvsLchi2btzF0EWy83dvpA.jpeg</url>
            <title>Stories by Zhong Hong on Medium</title>
            <link>https://medium.com/@zhonghong9998?source=rss-c99fe335779c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:00:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zhonghong9998/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[AI Didn’t Make You Smarter. It Made You Comfortable Not Thinking]]></title>
            <link>https://medium.com/@zhonghong9998/ai-didnt-make-you-smarter-it-made-you-comfortable-not-thinking-a89d5fb65391?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a89d5fb65391</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[thinking]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[personal-development]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 08:21:48 GMT</pubDate>
            <atom:updated>2026-06-02T08:21:48.329Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2aGG1BXyqGqE-MKp" /><figcaption>Photo by <a href="https://unsplash.com/@markuswinkler?utm_source=medium&amp;utm_medium=referral">Markus Winkler</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>There’s a specific feeling you get when you paste a messy problem into an AI and it comes back with a tidy explanation.</p><p>Relief. You got the answer without the discomfort of not knowing.</p><p>That relief is the problem.</p><p>The standard pitch for AI is productivity. You move faster. You get more done.</p><p>Your output doubles and your manager notices. Nobody argues with that.</p><p>What the pitch leaves out: speed only compounds good thinking. If you skipped the thinking part, all AI did was get you to a wrong answer faster.</p><p>The pattern I keep seeing is people using AI to fill the gap where thinking is supposed to happen.</p><p>Not after they’ve formed their own read, but instead of forming one. You see an anomaly in the data, a question from a stakeholder, a problem you haven’t framed yet.</p><p>And the first move is to paste it into Claude. Not to sit with it. Not to make a guess. Just to skip the uncomfortable part where you don’t yet know.</p><p>It doesn’t feel like laziness. It feels like efficiency. You have an answer. You’re moving.</p><p>The problem only surfaces later — when someone asks you to defend the output and you realise you can’t, because you never formed your own position in the first place.</p><p>AI doesn’t replace thinking. It replaces the <em>discomfort</em> of thinking. And the discomfort is where the work actually happens.</p><p>Here’s the mechanism: when you prompt before you hypothesise, you anchor to whatever framing comes back.</p><p>Every follow-up question you ask, every interpretation you form, every decision you make after that point is shaped by the AI’s first answer — not your own read.</p><p>You’re not evaluating its output. You’re updating from it. The anchor is set before you’ve started.</p><p><strong>The people who get the most out of AI do it differently. They form a rough hypothesis first — even a bad one — and then use AI to pressure-test it. The hypothesis doesn’t have to be right.</strong></p><p><strong>It just has to be theirs. Because having a position before you prompt means you’re running a test. Not accepting a briefing.</strong></p><p>There’s a real difference between using AI to <em>extend</em> your thinking and using it to <em>replace the start</em> of your thinking. Most people have quietly drifted into the second without noticing.</p><p>There’s a well-documented phenomenon in cognitive science called anchoring: <strong>the first piece of information you receive shapes all subsequent judgment</strong>.</p><p>When AI frames a problem before you have, you don’t evaluate its answer neutrally — you evaluate everything else relative to it.</p><p>That’s not thinking with AI. That’s thinking after AI has already decided where to start.</p><p>The GPS analogy is more uncomfortable. Research on consistent GPS use shows it degrades the brain’s ability to build independent spatial maps over time.</p><p>Not immediately. Quietly, from disuse. Nobody decides to stop navigating — they just stop needing to.</p><p>Until the GPS fails and they realise they genuinely can’t find their way.</p><p>Outsource the first moment of reasoning consistently, across problem after problem, and the same thing happens to analytical judgment. You don’t notice it weakening.</p><p>You notice it when you’re in a meeting without your tools open and someone asks you to think out loud.</p><p>I’ve made a deliberate choice to think before I prompt.</p><p>Not because I’m sceptical of AI — I use it every day — but because I’ve seen clearly what happens when you don’t.</p><p>The sharpest example was a work question involving something domain-specific in fintech data.</p><p>I asked Claude for help and got back output that looked authoritative: structured, detailed, confident. I couldn’t understand it. Not because it was complex.</p><p>Because it was wrong in ways that only became visible if you already understood the domain. The model had filled in the gaps with plausible-sounding nonsense.</p><p>Tf I hadn’t known enough to question it, I would have sent that to my manager as my own analysis.</p><p>After that, I stopped treating AI as a source of answers. I started treating it like a smart junior analyst — someone who has read everything but experienced nothing.</p><p>You don’t hand that person a problem and walk away. You brief them, direct them, check their work. The thinking stays with you. The execution gets faster.</p><p>The people I see using AI well bring a position to the tool. They’ve already done enough thinking to know what they’re testing.</p><p>AI helps them go further — pushes back, finds holes, generates options they hadn’t considered.</p><p>They move faster and they go farther. But the thinking is theirs. AI is running alongside it, not ahead of it.</p><p>The people not getting that benefit are treating the tool as a starting point instead of a sounding board. They get answers. They just can’t tell whether the answers are right.</p><p>One habit. Before you open Claude on a new problem, write one sentence: your own read of what’s happening.</p><p><strong>It doesn’t have to be right. It just has to exist before you prompt.</strong></p><p>That sentence is the line between using AI as a thinking partner and using it as a thinking replacement.</p><p><strong>The goal isn’t to use it less. It’s to make sure there’s still someone thinking when you do.</strong></p><p>📤Thanks for reading! If you found this useful…</p><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a89d5fb65391" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If You Can’t Defend Why Your Dashboard Exists, an AI Agent Will Replace It by End of Year]]></title>
            <link>https://medium.com/@zhonghong9998/if-you-cant-defend-why-your-dashboard-exists-an-ai-agent-will-replace-it-by-end-of-year-8e3db6ac2c5c?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e3db6ac2c5c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[power-bi]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Mon, 11 May 2026 12:56:00 GMT</pubDate>
            <atom:updated>2026-05-11T12:56:00.484Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8EQPyH92vDBVG1Xy" /><figcaption>Photo by <a href="https://unsplash.com/@agk42?utm_source=medium&amp;utm_medium=referral">Alex Knight</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>At my past working experience in WEEX, the marketing team asked me to build a dashboard.</p><p>User growth — new registrations, DAU, activation rate. I built it. Clean data, real-time refresh, exactly what they asked for.</p><p>Three months later, users were growing. Revenue wasn’t.</p><p>They stared at the dashboard. Nobody could explain the gap.</p><p>Not because the data was wrong — it was right. Because the dashboard was never designed to surface the <em>relationship</em> between growth and revenue. It showed one thing in isolation. That’s all anyone asked it to do.</p><h3>The Problem With “Visibility on X”</h3><p>Most dashboard requests come in the same shape:</p><p><em>“We need visibility on X.”</em></p><p>That sounds like a clear requirement. It isn’t.</p><p>Visibility on X is a request for a scoreboard — without specifying what game is being played, what winning looks like, or what anyone would do differently when the score changes.</p><p>The analyst delivers the chart. The stakeholder closes the request. The dashboard goes live.</p><p>And somewhere in that process, nobody asks the one question that matters: <strong>when this number moves, what do we actually do?</strong></p><h3>Why the Weex Dashboard Failed</h3><p>User growth moving up while revenue stays flat is a leading/lagging mismatch.</p><p>User growth is a leading indicator. Revenue is the lag.</p><p>Track one in isolation and you have a number — not a signal.</p><p>The Weex dashboard was technically correct. It just wasn’t connected to the decision that mattered.</p><h3>The Only Question That Should Start a Dashboard Build</h3><blockquote><strong><em>A dashboard is only as useful as the decision it’s attached to.</em></strong></blockquote><p>That sounds obvious. Most teams don’t build like it’s true.</p><p>The question that belongs at the start of every dashboard request isn’t:</p><p><em>“What metric do you want to track?”</em></p><p>It’s:</p><p><em>“When this number changes, what do you do differently?”</em></p><p>If there’s no clear answer, you don’t have a dashboard requirement. You have a report dressed in chart format — running on a server, costing compute, and giving everyone the impression they have data infrastructure.</p><h3>The Dashboard I Can Actually Defend</h3><p>Compare the Weex build to the PNL dashboard I maintain now.</p><p>I can defend that one in one sentence:</p><p><em>“It exists so the trading desk knows their daily and year-to-date position before the market opens.”</em></p><p>Specific person. Specific moment. Specific action triggered by what they see.</p><p>That’s a dashboard.</p><p>The Weex user growth one was a scoreboard with no game attached.</p><p>The difference between them isn’t technical complexity. <strong>It’s whether anyone wrote down the decision before the build started.</strong></p><h3>Agentic BI Is Already Here — And It’s Asking the Same Question</h3><p>This isn’t a distant risk. The tools are already in your stack.</p><p>Traditional dashboards run on a pull model: a human opens it, reads it, decides what to do.</p><p>Agentic analytics inverts that — the system monitors continuously, explains anomalies in plain language, and triggers actions at preset thresholds without waiting for anyone to notice.</p><p><a href="https://www.cio.com/article/4046967/the-end-of-dashboards-genai-and-agentic-workflows-transform-business-intelligence.html">CIO Magazine frames the shift</a> as moving from “time-to-insight” to “time-to-execution.” From systems of record to systems of action.</p><p>One number from that piece is worth sitting with:</p><blockquote><strong><em>58% of business decision-makers still rely on gut feel over data.</em></strong></blockquote><p>Not because dashboards fail to show numbers.</p><p>Because dashboards only answer questions someone anticipated asking.</p><h3>Every Major BI Platform Is Already Embedding This</h3><p>Snowflake Cortex AI, Databricks Genie, Power BI — <a href="https://dataupward.com/how-ai-agents-are-replacing-traditional-dashboards-in-2026/">every major BI vendor is embedding agentic capabilities right now</a>.</p><p>Forrester’s read: GenAI isn’t replacing BI. It’s standardising across every platform at once.</p><p>Which means the agentic layer under your dashboards will eventually ask the same question you should be asking today:</p><p><strong>Does this drive a decision — or just display a number?</strong></p><h3>We’ve Been Here Before</h3><p>This isn’t the first time tools exposed shallow work.</p><p>When self-serve BI arrived, Tableau didn’t eliminate analysts. It eliminated analysts who only knew how to pull data without context.</p><p>The tools didn’t get smarter. They made the cost of shallow work visible.</p><p>That’s what’s happening now — one layer up.</p><p>The dashboards that survive aren’t the most technically impressive ones. They’re the ones with an owner, a question, and a decision attached.</p><p><strong>Everything else is noise with a refresh schedule.</strong></p><h3>How I Handle Dashboard Requests Now</h3><p>Before I agree to build anything, I ask one question:</p><p><em>“What decision does this need to support?”</em></p><p>Not as a bureaucratic gate. As a filter.</p><p>If the answer is “leadership wants visibility” or “we just want to keep an eye on it” — I push back. Not aggressively. But I don’t start building until there’s:</p><ul><li>A real owner</li><li>A real question</li><li>A clear moment when they’d act on what they see</li></ul><p>Stakeholders are used to being handed what they asked for. Being asked <em>why</em> feels like obstruction.</p><p>It’s not obstruction. It’s the question that should have been asked at the start.</p><p>If a stakeholder can’t explain why a dashboard needs to exist when requesting it, they won’t explain it when an audit asks six months later. Better to find the gap before the build than after three months of maintenance.</p><h3>The 60-Second Audit You Should Run This Week</h3><p>Pick five dashboards you own right now.</p><p>For each one, write one sentence:</p><blockquote>“This dashboard exists so [specific person] can make [specific decision] when [specific condition].”</blockquote><p>If you can write it clearly — the dashboard has a reason to exist.</p><p>If the best you can produce is <em>“so the team has visibility on performance”</em> — you have a clarity problem, not a dashboard problem.</p><p>No chart type or colour scheme fixes that.</p><h3>Final Thought</h3><p>The AI audit is coming for every dashboard in your organization.</p><p>It will ask a simple question: does this drive action, or just occupy a screen?</p><p>You still have time to answer it first.</p><p>The dashboards that survive won’t be the prettiest. They’ll be the ones someone can defend in one sentence.</p><p><strong>Write that sentence before someone — or something — asks you for it.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e3db6ac2c5c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Spent Hours Writing Perfect SQL. Then Realized I Solved the Wrong Problem.]]></title>
            <link>https://medium.com/@zhonghong9998/i-spent-hours-writing-perfect-sql-then-realized-i-solved-the-wrong-problem-945328775f65?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/945328775f65</guid>
            <category><![CDATA[database]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[sql]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[data-analyst]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 12:31:01 GMT</pubDate>
            <atom:updated>2026-04-27T12:31:01.553Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Spent Hours Writing Perfect SQL. Then Realized I Solved the Wrong Problem</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E47WOe3KTSLFSWAM" /><figcaption>Photo by <a href="https://unsplash.com/@casparrubin?utm_source=medium&amp;utm_medium=referral">Caspar Camille Rubin</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The Query That Worked. And Still Failed.</p><p>I once wrote a SQL query I was genuinely proud of.</p><p>Clean joins. Correct aggregations. Zero errors.</p><p>I ran it. Reviewed the output. Everything looked right.</p><p>Then I sent it to the stakeholder.</p><p>Their response: <em>“This isn’t what we needed.”</em></p><p>That moment sat with me for a while.</p><p>The code wasn’t broken. The logic was sound. The numbers were accurate.</p><p>And yet — I had completely failed to deliver what was actually needed.</p><p>That’s the kind of failure nobody warns you about when you’re learning data analytics.</p><h3>The Problem Nobody Talks About</h3><p>If you’ve worked with data long enough, you’ve been here:</p><ul><li>The query runs perfectly</li><li>The numbers look reasonable</li><li>The logic checks out</li></ul><p>And still — the stakeholder pushes back.</p><p>Not because you made a technical error. But because you answered a question they weren’t asking.</p><blockquote><strong><em>This is one of the most demoralizing moments in data work. Not a bug. Not a crash. Just a perfectly executed answer to the wrong problem.</em></strong></blockquote><p>We’re trained to obsess over the technical layers — better SQL, cleaner joins, more efficient window functions.</p><p>But in reality, most data failures aren’t technical at all.</p><p><strong>They’re definition failures.</strong></p><h3>What I Actually Got Wrong</h3><p>After that experience, I forced myself to confront something uncomfortable.</p><p>I didn’t misunderstand SQL.</p><p><strong>I misunderstood the problem.</strong></p><p>I was answering a question. Just not the right one. And I had never stopped to check whether my interpretation matched what the business actually needed.</p><p>That distinction — between <em>answering a question</em> and <em>answering the right question</em> — is the gap that separates average analysts from genuinely useful ones.</p><h3>A Real Example: Commission vs. Swap</h3><p>Let me make this concrete.</p><p>We were building a performance metric for trading activity. The task seemed clear: measure the cost impact on each trade.</p><p>Our query logic? Correct. Our calculations? Accurate. Our output? Clean and well-formatted.</p><p>We treated commission and swap as two separate line items — because technically, they are.</p><p><strong>But the business didn’t care about them individually.</strong></p><p>They cared about their <em>combined</em> impact, because that’s what reflects the true profitability of a trade. Splitting them out wasn’t wrong mathematically — it was wrong <em>conceptually</em>.</p><p>Same data. Different definition. Completely different conclusion.</p><p>The query executed our assumption perfectly. The assumption was just never validated.</p><h3>Why This Keeps Happening</h3><p>This kind of mistake is more common than most analysts admit:</p><ul><li><strong>Metrics are assumed, not defined</strong> — everyone thinks they know what “revenue” or “active user” means until you put two people in a room</li><li><strong>Teams use the same words differently</strong> — finance and product might both say “conversion” and mean entirely different things</li><li><strong>Edge cases are never clarified</strong> — what happens with partial records? refunds? time zone differences?</li></ul><p>Here’s the brutal part:</p><p><strong>SQL doesn’t catch definition mistakes. It executes them — perfectly and confidently.</strong></p><p>Your query will return a result either way. The database doesn’t know that your WHERE clause is filtering out something the business considers critical. It just runs.</p><h3>The Framework I Use Now</h3><p>After enough of these painful experiences, I built a simple check I run <em>before</em> writing a single line of SQL.</p><p>It adds maybe 10–15 minutes upfront. It saves hours on the back end.</p><h4>Step 1 — Define the Metric Clearly</h4><p>Don’t assume you know what’s being measured. Ask explicitly:</p><ul><li>What exactly are we calculating?</li><li>What is included in this number? What is excluded?</li><li>Is there an existing definition, or are we creating one?</li></ul><h4>Step 2 — Align on Business Meaning</h4><p>A metric doesn’t exist in isolation. Understand its purpose:</p><ul><li>How does the team interpret this number?</li><li>What decision will it inform?</li><li>Who else uses this metric — and might define it differently?</li></ul><h4>Step 3 — Clarify the Edge Cases</h4><p>This is where most definition failures hide:</p><ul><li>Real-time data or end-of-day snapshots?</li><li>Do we include partial transactions or only completed ones?</li><li>How do we handle refunds, cancellations, or exceptions?</li></ul><h4>Step 4 — Repeat the Logic Back</h4><p>Before touching the database, I say it out loud — or write it in the Slack thread:</p><blockquote>“So we’re calculating net trade cost as commission plus swap, applied at the trade level, excluding internal test accounts — correct?”</blockquote><p>This one habit has saved me more rework than any SQL optimization ever has.</p><h4>Step 5 — Validate Together</h4><p>Don’t just send the results and wait.</p><p>Walk through the numbers with the stakeholder. Check the edge rows together. Ask: <em>“Does this match what you expected?”</em></p><p>Catching a misalignment at this stage takes 10 minutes. Catching it after a dashboard goes live takes days.</p><h3>The Lessons That Actually Stuck</h3><p>After years of data work, these are the truths I keep coming back to:</p><ul><li><strong>Correct SQL doesn’t guarantee correct answers</strong> — it guarantees <em>consistent</em> answers, which is different</li><li><strong>Definitions matter more than queries</strong> — a perfect query built on a fuzzy definition is still wrong</li><li><strong>Assumptions are where the real bugs live</strong> — not in your code, but in what you never thought to question</li><li><strong>Clarity upfront is the highest-leverage habit</strong> — 15 minutes of alignment prevents hours of rework</li><li><strong>Data analysts are translators first, coders second</strong> — the technical skill is table stakes; the translation skill is the differentiator</li></ul><h3>The Mental Shift That Changed My Work</h3><p>I used to believe my job was:</p><p><em>Write accurate queries.</em></p><p>Now I believe my job is:</p><p><em>Define the right problem — then solve it.</em></p><p>That reframe sounds small. It isn’t.</p><p>It changes how you start every analysis. How you read every request. How you ask questions before you answer them.</p><blockquote>“Most data mistakes aren’t technical. They’re conceptual.”</blockquote><h3>Final Thought</h3><p>The biggest risk in data work isn’t a broken query.</p><p>It’s a flawless query solving the wrong problem.</p><p>And the more time I spend in this field, the more I believe this:</p><p>The analysts who are genuinely valuable aren’t the ones who write the best SQL.</p><p><strong>They’re the ones who figure out what actually needs to be measured — before anyone else thinks to ask.</strong></p><p>That’s not a technical skill.</p><p>It’s a thinking skill. And it’s the one worth developing most.</p><p>📤Thanks for reading! If you found this useful…</p><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=945328775f65" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Worked in Crypto — Here’s What It Taught Me About Data, Speed, and Decision-Making]]></title>
            <link>https://medium.com/@zhonghong9998/i-worked-in-crypto-heres-what-it-taught-me-about-data-speed-and-decision-making-d3e8c4d91df8?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/d3e8c4d91df8</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[career-growth]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Mon, 13 Apr 2026 13:16:01 GMT</pubDate>
            <atom:updated>2026-04-13T13:16:01.628Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Worked in Crypto — Here’s What It Taught Me About Data, Speed, and Decision-Making</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CxI1MRm_og6L8CvjpMPpkw.png" /><figcaption>By Zhong Hong</figcaption></figure><p>I thought working in crypto would be all charts, volatility, and big market energy.</p><p>Reality was uglier.</p><p>It was logs. Broken APIs. Urgent tickets at odd hours. Weird edge cases that made no sense until they suddenly made all the sense. And decisions that had to be made before the coffee got cold.</p><p>Honestly? That environment taught me more about data and decision-making than most “clean” analytics work ever did.</p><h3>The Problem: Crypto Makes Bad Data Feel Very Expensive</h3><p>In a normal analytics team, bad data is annoying.</p><p>In crypto, bad data can become a support fire, a product bug, or a decision made five minutes too late.</p><p>That is the part nobody romanticizes enough.</p><p>The pain points were familiar — but amplified:</p><ul><li>Data changed fast</li><li>Systems broke often</li><li>Users expected instant answers</li><li>Small mistakes became visible immediately</li><li>“Later” was usually too late</li></ul><p>I was handling API issues, ticket prioritization, and reproducible debugging across WebSocket, REST, and margin/risk interfaces — while also building Python and SQL workflows for high-volume transaction data.</p><p>It was not glamorous. It was clarifying.</p><h3>What Broke My Illusion</h3><p>At first, I treated speed like a badge of honor.</p><p>Respond fast. Fix fast. Ship fast.</p><p>Then I learned the hard way: fast without discipline is just expensive chaos.</p><p>One lesson hit me brutally.</p><p><strong>Never test API code in the production environment first. Use a virtual environment.</strong></p><p>I learned that the hard way, and it cost me. That kind of mistake does not just hurt the codebase. It can hurt your reputation, your confidence, and your job.</p><p>That was a painful lesson in environment isolation, reproducibility, and not treating production like your personal sandbox.</p><p>Nobody puts that on a résumé. But everybody remembers it.</p><h3>The Real Framework: Speed Is a System, Not a Personality Trait</h3><p>Crypto taught me that speed is not about moving recklessly.</p><p>It is about building a system that lets you move quickly without being stupid.</p><h3>My working framework became:</h3><p><strong>1. Reproduce first, panic second.</strong></p><p>Do not guess. Recreate the issue.</p><p>If you cannot reproduce it, you are doing vibes-based engineering.</p><p><strong>2. Separate the environment before you separate the blame.</strong></p><p>Test in a virtual environment. Keep dev, staging, and production distinct.</p><p>One sloppy shortcut can burn hours later.</p><p><strong>3. Log everything that matters.</strong></p><p>If a bug cannot be traced, it will be rediscovered by the next poor soul on the team.</p><p><strong>4. Use data to rank urgency.</strong></p><p>Not every fire is a fire.</p><p>Some are smoke. Some are just someone yelling because their dashboard blinked.</p><p><strong>5. Make the decision visible.</strong></p><p>Document why you chose one path over another.</p><p>Fast teams do not just act fast. They explain fast.</p><h3>What Crypto Taught Me About Decision-Making</h3><p>Crypto is an extreme case, but the lesson generalizes.</p><p>Good decision-making is not about always being right.</p><p>It is about being right often enough, fast enough, and with enough evidence.</p><p>That means:</p><ul><li>Knowing what you know</li><li>Knowing what you do not know</li><li>Knowing what can wait</li><li>Knowing what cannot wait</li></ul><p>In one stretch of my work, I was resolving API-related issues, managing more than 50 tickets, and building internal knowledge bases so recurring problems would stop eating the team alive.</p><p>That changed how I think about decisions entirely.</p><blockquote><strong><em>A good decision is not only the one that solves today’s problem. It is the one that reduces tomorrow’s noise.</em></strong></blockquote><h3>A Practical Lens for Analysts</h3><p>You do not need to work in crypto to steal the mindset.</p><p>Here is a 4-step filter I now apply to almost every data issue I encounter:</p><p><strong>Is this issue real?</strong> Validate the signal before reacting. Not every alert deserves a war room.</p><p><strong>Is this issue urgent?</strong> Separate “important” from “loud.” They are not the same thing.</p><p><strong>Is this issue repeatable?</strong> If yes, build a system. If no, investigate carefully before drawing conclusions.</p><p><strong>Is this issue preventable?</strong> If yes, improve the workflow. Do not just celebrate a one-time fix like a hero.</p><p>That last one matters most.</p><p>Heroic firefighting looks impressive in the moment. Systems thinking is what makes you valuable over time.</p><h3>Real Example: Why Reproducibility Saved Me</h3><p>When I was deep in API work and transaction data, the difference between a useful fix and a useless one was simple:</p><p>Could I reproduce the problem cleanly?</p><p>That meant:</p><ul><li>Testing with controlled inputs</li><li>Isolating dependencies</li><li>Using the right environment</li><li>Documenting the exact failure path</li></ul><p>Later, I built automated Python ETL pipelines for more than 12 million crypto transactions and used SQL transformations to clean and restructure that data for analytics and model training.</p><p>That kind of work punishes sloppy habits fast.</p><p>If your pipeline is not reproducible, your “solution” is just a temporary mood.</p><h3>Key Lessons</h3><ul><li>Speed without environment control is a trap</li><li>Production is not your testing lab</li><li>Reproducibility is not optional — it is professional hygiene</li><li>Data work gets much harder when systems move fast</li><li>The best analysts do not just answer questions. They reduce future chaos.</li><li>Good decisions are usually boring, documented, and repeatable</li><li>A clean process beats a clever panic move every single time</li></ul><blockquote><strong><em>Fast is good. Reproducible is better. Fast and reckless is how you get humbled by reality.</em></strong></blockquote><h3>The Real Skill Nobody Talks About</h3><p>Crypto did not just teach me about markets.</p><p>It taught me that data work is really about managing uncertainty with discipline.</p><p>Speed matters. But speed without structure is just a quick way to make a mess.</p><p>The real skill is quieter than people expect:</p><p>Stay calm. Test properly. Document clearly. Make decisions the next person can trust.</p><p>Because in the real world, your value is not how fast you can type.</p><p><strong>It is how well you can keep the system from eating itself.</strong></p><p>📤Thanks for reading! If you found this useful…</p><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d3e8c4d91df8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why a High Accuracy Model Can Still Be Useless]]></title>
            <link>https://medium.com/@zhonghong9998/why-a-high-accuracy-model-can-still-be-useless-eb7358e92c19?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb7358e92c19</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Tue, 07 Apr 2026 12:56:01 GMT</pubDate>
            <atom:updated>2026-04-07T12:56:01.449Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O2hl0JzclcI3s1TB9uFZxw.png" /><figcaption>By Zhong Hong</figcaption></figure><p><strong>The Moment I Realized Accuracy Was Lying to Me</strong></p><p>I once built a fraud detection model with <strong>99.2% accuracy</strong>.</p><p>I remember staring at that number like it was a trophy.</p><p>But here’s the part I don’t usually tell people:</p><p>This wasn’t even for a company. It was a personal project. No stakeholders. No business pressure. No one watching over my shoulder.</p><p>And honestly? That made it worse.</p><p>Because for a while, I genuinely believed I was good at this.</p><blockquote>“Damn… I’m actually onto something here.”</blockquote><p>I wasn’t. I was just good at fooling myself.</p><h3>The Silent Problem: No One Challenged Me</h3><p>When you work on a personal project, no one asks the hard questions.</p><ul><li>No PM asking what problem this solves</li><li>No manager questioning your assumptions</li><li>No business metric telling you whether you’re actually useful</li></ul><p>So I did what most self-taught analysts do in that situation.</p><ul><li>Focused on improving the accuracy score</li><li>Tuned hyperparameters until I was above 99%</li><li>Celebrated the number going up</li></ul><p>But I never asked the only question that actually matters:</p><blockquote>“Would this work in the real world?”</blockquote><p>That question didn’t even cross my mind. The number felt like an answer to everything.</p><h3>The Reality Check I Gave Myself</h3><p>One afternoon, I stopped looking at accuracy and actually looked at the predictions.</p><p>That’s when everything fell apart.</p><p>My model was doing something embarrassingly simple — it was predicting almost every transaction as <strong>non-fraud</strong>. Every single one. Legitimate, legitimate, legitimate, all the way down.</p><p>And technically? It was correct.</p><p>Because fraud cases were extremely rare. Less than 1% of the dataset.</p><p>The model hadn’t learned anything. It had found a shortcut.</p><h3>The Analyst Trap: When “Correct” Is Meaningless</h3><p>Here’s the uncomfortable truth most courses skip past:</p><blockquote><strong><em>You can be technically correct and still be completely useless.</em></strong></blockquote><p>The math in my case was brutal in its simplicity:</p><ul><li>Dataset → roughly 99% legitimate transactions</li><li>Model strategy → predict everything as legitimate</li><li>Result → 99% accuracy, with zero fraud caught</li></ul><p>A model that catches no fraud is not a fraud detection model. It’s a very expensive way to do nothing.</p><p>The number looked great. The model was worthless.</p><h3>The Turning Point: Thinking Like a Business, Not a Model</h3><p>That’s when something clicked — something most tutorials never teach you.</p><blockquote><strong><em>Machine learning is not about prediction. It’s about decision-making.</em></strong></blockquote><p>So I reframed the question entirely.</p><p>Instead of asking: <em>“How accurate is my model?”</em></p><p>I started asking: <em>“If this model runs in production, what actually happens?”</em></p><p>The answer was brutal. Fraud would keep happening. Losses would keep accumulating. And we’d have a model proudly reporting 99% accuracy the whole time.</p><p>We’d be losing money. Just very accurately.</p><h3>A Practical Framework I Use for Every Model Now</h3><p>After that experience, I built a mental checklist. I run through it for every model — including personal projects, <em>especially</em> personal projects.</p><h4>1. Define the Real Goal (Even If It’s Just a Side Project)</h4><p>Force yourself to answer: what would a business actually care about here?</p><ul><li>Fraud detection → minimize financial loss</li><li>Churn prediction → maximize retention</li><li>Loan approval → minimize default risk</li></ul><p>If you can’t answer that, you’re not building a model. You’re playing with numbers.</p><h4>2. Simulate Reality — Don’t Trust Your Dataset Blindly</h4><p>Before anything else, ask two questions:</p><ul><li>Is my data imbalanced?</li><li>What does doing <em>nothing</em> achieve?</li></ul><p>In my case, doing nothing — predicting every transaction as legitimate — would have achieved 99% accuracy. My model matched that baseline exactly.</p><p>That’s not a model. That’s the absence of one.</p><h4>3. Break Down the Mistakes</h4><p>Instead of one number, look at what kind of errors you’re making:</p><ul><li><strong>False negatives</strong> — fraud you missed (the expensive ones)</li><li><strong>False positives</strong> — legitimate transactions you flagged (annoying, but cheap)</li></ul><p>Then ask: which error actually matters more in this context? The answer shapes everything.</p><h4>4. Put a Price Tag on Your Errors</h4><p>Even in a personal project, simulate the cost:</p><ul><li>Missed fraud → –$500 per case</li><li>False alarm → –$5 per case</li></ul><p>Now evaluate two models by total estimated loss, not accuracy score.</p><p>Suddenly the conversation shifts from <em>“which model scores higher?”</em> to <em>“which model costs less?”</em> That’s how real decisions get made.</p><h4>5. Optimize for Impact, Not Ego</h4><p>When I adjusted the classification threshold to catch more fraud, my accuracy dropped.</p><p>Old me would have panicked. Old me would have undone that change immediately.</p><p>But now I understood: catching more fraud at slightly lower accuracy is a better outcome. The business wins. That’s the whole point.</p><h3>What This Actually Changed for Me</h3><p>One small project, no deployment, no real stakes — and it taught me more than any course had.</p><ul><li>Metrics can mislead you even when you’re working alone</li><li>High scores often hide low-quality thinking</li><li>Personal projects are dangerous without a reality check</li></ul><p>And the biggest lesson of all:</p><blockquote><strong><em>If no one is challenging your work, you have to challenge yourself.</em></strong></blockquote><p>That’s the discipline that separates analysts who grow from analysts who stagnate.</p><h3>Key Lessons From Someone Who Fell Into the Trap</h3><ul><li>A 99% accurate model can perform worse than doing nothing at all</li><li>Personal projects are risky when no one pushes back on your assumptions</li><li>Always look at what your model is <em>actually</em> predicting — not just the score</li><li>Accuracy is easy to game. Real-world impact is not.</li><li>If it doesn’t change decisions, it doesn’t matter</li></ul><blockquote><strong><em>“If your model isn’t solving a real problem, it’s just a demo.”</em></strong></blockquote><h3>Closing Thought</h3><p>Not every bad model comes from bad technique.</p><p>Sometimes it comes from the absence of someone willing to ask uncomfortable questions.</p><p>If you’re serious about becoming a better analyst, you can’t wait for a manager to do that for you. You have to build the habit of questioning yourself — especially when everything looks fine.</p><p>Next time you hit a number that makes you feel proud, sit with this for a moment:</p><p><strong>“Would anyone actually care if this runs in production?”</strong></p><p>That question alone will make you a sharper analyst than any course ever could.</p><p>📤Thanks for reading! If you found this useful…</p><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb7358e92c19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Precision vs Recall Can Make or Break Your Fraud Detection Model]]></title>
            <link>https://medium.com/@zhonghong9998/why-precision-vs-recall-can-make-or-break-your-fraud-detection-model-75f92c2b83f6?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/75f92c2b83f6</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[metrics]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 12:46:01 GMT</pubDate>
            <atom:updated>2026-04-06T12:46:01.367Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aQYdoGizy-oTGldL" /><figcaption>Photo by <a href="https://unsplash.com/@towfiqu999999?utm_source=medium&amp;utm_medium=referral">Towfiqu barbhuiya</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The first time I looked at my fraud detection model results, I felt a tiny burst of confidence.</p><p>Then I looked again.</p><p>And that confidence died immediately.</p><p>Because the model was doing something very annoying:</p><p>It was catching almost all the fraud cases, but it was also yelling at a ton of innocent wallets.</p><p>That is when precision and recall stopped being theory to me.</p><p>They became real.</p><h3>The problem: the model was acting like an overcaffeinated security guard</h3><p>Imagine a mall security guard who screams “thief!” every time someone walks too fast.</p><p>Sure, he will catch the actual thief sometimes.</p><p>But he will also ruin the day of half the mall.</p><p>That was my fraud model at the beginning.</p><p>In my crypto airdrop fraud detection project, the first version had <strong>very high recall</strong> but <strong>low precision</strong>.</p><p>In plain English:</p><p>It found almost every suspicious wallet.</p><p>But a lot of the wallets it flagged were probably fine.</p><p>So yes, the model looked “smart.”</p><p>But for the people who actually had to review those alerts?</p><p>It was noisy, tiring, and expensive.</p><h3>Why this matters more than people think</h3><p>A lot of analysts fall into the same trap.</p><p>They see one shiny metric and stop there.</p><p>Usually that metric is accuracy.</p><p>And accuracy loves to lie in imbalanced problems.</p><p>Fraud is rare.</p><p>So a model can be “accurate” and still be useless.</p><p>That is why precision and recall matter so much.</p><p>They tell you the thing accuracy cannot.</p><p>They tell you <strong>how often your model is right when it speaks</strong>, and <strong>how many real fraud cases it actually catches</strong>.</p><p>That is the real game.</p><h3>A simple analogy that finally made it click for me</h3><p>Think of precision and recall like fishing.</p><p>If you throw a giant net into the ocean and drag up everything, you may catch almost every fish.</p><p>That is high recall.</p><p>But you will also drag up seaweed, shoes, plastic, and half the ocean floor.</p><p>That is low precision.</p><p>Now imagine you become super selective and only catch fish that look exactly right.</p><p>That increases precision.</p><p>But you may accidentally let a lot of real fish swim away.</p><p>That is the tradeoff.</p><p>Fraud detection is basically that.</p><p>Except the fish are wallets.</p><p>And the ocean is chaos.</p><h3>What I actually did in the project</h3><p>I started by cleaning and preparing wallet behavior data.</p><p>That part matters more than most people admit.</p><p>Because if your features are weak, your model is basically guessing with confidence.</p><p>In the preprocessing stage, I worked with wallet-level behavior signals such as:</p><ul><li>total number of transactions</li><li>number of unique contracts interacted with</li><li>total ETH value</li><li>gas spent</li><li>wallet age</li><li>average transaction value</li><li>transaction diversity</li></ul><p>These kinds of features help the model understand behavior patterns, not just raw activity.</p><p>And in fraud detection, behavior is usually where the story is hiding.</p><p>A wallet that suddenly becomes active after being quiet for a long time does not look the same as a wallet with steady normal activity.</p><p>That difference matters.</p><p>A lot.</p><h3>The turning point: I stopped trusting the default threshold</h3><p>This part is where a lot of people get lazy.</p><p>They train a model, use the default threshold of 0.5, and call it a day.</p><p>That is like using the same volume setting for every conversation in your life.</p><p>No context. No judgment. Just vibes.</p><p>I did not want that.</p><p>So I looked at the precision-recall tradeoff more carefully.</p><p>And here is what happened:</p><p>With the default setup, the fraud class had:</p><ul><li><strong>precision = 0.322</strong></li><li><strong>recall = 0.954</strong></li></ul><p>That means the model caught almost all fraud cases, but most of its alerts were false alarms.</p><p>Then I tuned the threshold upward.</p><p>At a threshold of around <strong>0.974</strong>, the model became much stricter:</p><ul><li><strong>precision = 0.793</strong></li><li><strong>recall = 0.303</strong></li></ul><p>Now the model was far more selective.</p><p>It flagged fewer wallets.</p><p>But the ones it flagged were much more likely to be fraud.</p><p>That was the moment the whole thing made sense.</p><p>I was no longer asking:</p><p>“Can the model find everything?”</p><p>I was asking:</p><p>“Can the model find the right things without making the team miserable?”</p><p>That is a much better question.</p><h3>Precision vs recall in human language</h3><p>Here is the easiest way I explain it:</p><h4>Precision = “How many of my alerts were correct?”</h4><p>If I say 10 wallets are suspicious, how many of those are actually suspicious?</p><p>High precision means I am not wasting people’s time.</p><h4>Recall = “How many of the real fraud cases did I catch?”</h4><p>If there were 100 fraud wallets, how many did I manage to find?</p><p>High recall means I am missing less fraud.</p><p>So one is about <strong>being careful</strong>.</p><p>The other is about <strong>being thorough</strong>.</p><p>And in fraud detection, you rarely get both at the same time for free.</p><p>That is the whole headache.</p><h3>The real lesson: metrics are not decorations</h3><p>This is where analysts sometimes get seduced by dashboards.</p><p>A model with a pretty ROC-AUC score can still be operationally bad.</p><p>Because business teams do not live inside ROC curves.</p><p>They live inside queues, alerts, and limited human attention.</p><p>So the question is not:</p><p>“Which metric is highest?”</p><p>The question is:</p><p>“What kind of mistake hurts more?”</p><p>If false alerts are expensive, then precision matters more.</p><p>If missing fraud is catastrophic, then recall matters more.</p><p>That is not a machine learning question.</p><p>That is a business decision wearing a machine learning costume.</p><h3>What the model taught me about real work</h3><p>This project made one thing painfully clear:</p><p>A good model is not just about prediction.</p><p>It is about usefulness.</p><p>And usefulness depends on tradeoffs.</p><p>In my case, the first version of the model was basically yelling at everyone.</p><p>The improved version was quieter, sharper, and more useful.</p><p>That is what threshold tuning did.</p><p>It turned a noisy detector into something the team could actually trust.</p><h3>Key lessons I would keep on a sticky note</h3><ul><li>Accuracy can fool you in imbalanced fraud problems.</li><li>Precision tells you whether your alerts are worth trusting.</li><li>Recall tells you whether you are missing real fraud.</li><li>Threshold tuning is not optional.</li><li>Better features often beat fancier models.</li><li>A model is only useful if the people using it can handle its output.</li></ul><h3>The quote I keep coming back to</h3><blockquote><strong><em>“A fraud model is not good because it finds a lot. It is good because it finds the right things without creating a mess.”</em></strong></blockquote><p>That is the part people forget.</p><p>Speed is nice.</p><p>Complexity is nice.</p><p>Pretty charts are nice.</p><p>But if the model cannot support a real workflow, it is just expensive theater.</p><h3>Closing thought</h3><p>Precision and recall are not abstract math terms.</p><p>They are a decision about what kind of mistake you are willing to live with.</p><p>That is why fraud detection is such a good example for analysts.</p><p>It forces you to think beyond the metric.</p><p>Beyond the notebook.</p><p>Beyond the ego.</p><p>And once you understand that, you stop building models that only look smart.</p><p>You start building models that actually matter.</p><h3>📤Thanks for reading! If you found this useful…</h3><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=75f92c2b83f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What I Learned From Building a Fraud Detection Side Project on Crypto Wallets]]></title>
            <link>https://medium.com/@zhonghong9998/what-i-learned-from-building-a-fraud-detection-side-project-on-crypto-wallets-e549b123f592?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/e549b123f592</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[fraud-detection]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 12:26:01 GMT</pubDate>
            <atom:updated>2026-04-03T12:26:01.766Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zM-e3U-XQrH04SfN" /><figcaption>Photo by <a href="https://unsplash.com/@jefflssantos?utm_source=medium&amp;utm_medium=referral">Jefferson Santos</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><strong>It Started With a Pattern I Couldn’t Unsee</strong></p><p>When you work as a data analyst in crypto, you spend a lot of time staring at dashboards.</p><p>User counts. Transaction volumes. Wallet activity. Retention curves.</p><p>And after a while, you start noticing things that don’t add up.</p><p>I kept seeing spikes — sharp, sudden surges in new wallet activity — that would appear right around airdrop events. Then disappear just as fast. The wallets would claim tokens, dump immediately, and go silent.</p><p>No organic user behaves like that.</p><p>I flagged it internally. Discussed it with the team. But there was no systematic way to catch it at scale.</p><p>That curiosity stayed with me. So I decided to build something on my own time to answer the question properly.</p><h3>What Airdrop Fraud Actually Looks Like From the Inside</h3><p>If you’ve never worked in crypto, here’s what airdrop farming looks like in practice.</p><p>A project distributes free tokens to wallets that meet certain criteria — early users, active traders, community members. The goal is to reward loyalty and grow adoption.</p><p>But bad actors reverse-engineer the eligibility rules.</p><p>They spin up hundreds — sometimes thousands — of wallets. They simulate just enough activity to qualify. They claim the airdrop. They dump the tokens within minutes. And they disappear.</p><p><strong>The damage is real:</strong></p><ul><li>Token prices tank from coordinated sell pressure</li><li>Legitimate users get diluted</li><li>Projects lose credibility and funding momentum</li><li>The data analyst watching the dashboard is left wondering what just happened</li></ul><p>I watched versions of this play out. And the frustrating part was that the signals were always there in the data — buried in wallet behavior patterns that nobody had built a proper model to catch.</p><h3>Why I Built This as a Side Project</h3><p>I want to be upfront: the dataset I used came from an online source, not from my employer.</p><p>I wasn’t going to risk leaking proprietary data. But I understood the problem domain deeply enough to know what signals to look for — because I’d seen them at work.</p><p>That real-world context is what made the project meaningful.</p><p>I wasn’t following a tutorial. I was trying to recreate — in a clean, reproducible way — the exact fraud detection logic I wished had existed at my job.</p><blockquote><em>“The best side projects aren’t invented in a vacuum. They’re born from a problem you’ve already lived.”</em></blockquote><h3>How I Thought About Fraud Detection as a Data Analyst</h3><p>Before writing a single line of code, I asked myself: <em>if I were trying to game an airdrop, what would my wallet behavior look like?</em></p><p>This is the analyst’s secret weapon. Model the adversary first.</p><p>Here’s the behavioral logic I worked from:</p><p><strong>A bot wallet tends to:</strong></p><ul><li>Be newly created, right before or during the airdrop window</li><li>Have very low activity diversity — it only touches the airdrop contract</li><li>Dump tokens fast, often within minutes or hours of receipt</li><li>Show uniform, automated transaction patterns (same amounts, same intervals)</li><li>Have no meaningful history before or after the event</li></ul><p><strong>A real user tends to:</strong></p><ul><li>Have a wallet with genuine age and varied on-chain history</li><li>Interact with multiple protocols over time</li><li>Hold or gradually sell tokens, not dump immediately</li><li>Show irregular, human-looking transaction timing</li></ul><p>The gap between these two profiles is where the model lives.</p><h3>The Project Structure I Built</h3><p>I separated the work into two notebooks with one job each.</p><p><strong>Notebook 1 — </strong><a href="https://github.com/Zhonghong99/crypto-airdrop-fraud-detection-ml/blob/main/data_preprocessing_wallet_behavior.ipynb"><strong>data_preprocessing_wallet_behavior.ipynb</strong></a></p><p>This is where I did the analytical thinking.</p><ul><li>Defined what “suspicious” actually means in behavioral terms</li><li>Engineered features: transaction volume, wallet age, activity diversity, dump velocity</li><li>Built a rule-based labeling system to assign is_suspicious flags</li><li>Produced a clean 10,000-wallet dataset with 11 core features</li></ul><p><strong>Notebook 2 — </strong><a href="https://github.com/Zhonghong99/crypto-airdrop-fraud-detection-ml/blob/main/airdrop_fraud_detection_ml.ipynb"><strong>airdrop_fraud_detection_ml.ipynb</strong></a></p><p>This is where the ML pipeline ran.</p><ul><li>Train/test split with stratification (critical for imbalanced fraud data)</li><li>Baseline classification models</li><li>Evaluated using precision/recall, not just accuracy</li><li>Outputs documented clearly for non-technical readers</li></ul><p>Keeping them separate was a deliberate choice. Preprocessing and modeling are different types of thinking.</p><p>Mixing them creates notebooks that are hard to debug, hard to explain, and hard to hand off.</p><h3>The Feature That Surprised Me Most</h3><p>I expected wallet age to be the strongest fraud signal.</p><p>It wasn’t.</p><p><strong>Dump velocity</strong> — how fast a wallet moves tokens after receiving them — turned out to be far more discriminating.</p><p>A human might sell tokens within a week.</p><p>A bot sells within 30 minutes. That gap is enormous, and it shows up cleanly in the data.</p><p>This is the kind of insight you only get when you build the thing yourself. Reading about fraud detection gives you frameworks.</p><p>Building it gives you intuition.</p><h3>The Hardest Part: Labeling</h3><p>Here’s what nobody tells you about fraud detection projects.</p><p>The model is the easy part.</p><p>Labeling is where you earn your money.</p><p>When there are no ground-truth labels — no confirmed fraud cases from a compliance team — you have to define suspicious behavior yourself. And every choice you make in that labeling logic directly shapes what the model learns.</p><p>I spent more time designing the rule-based labeling system than I did training the model.</p><p><strong>My labeling approach:</strong></p><ul><li>No single signal flags a wallet as suspicious</li><li>Fraud is defined by the <em>combination</em> of signals</li><li>Thresholds were set conservatively to minimize false positives</li><li>The logic is fully documented so anyone can challenge or improve it</li></ul><p>That last point matters. Reproducibility isn’t just a technical virtue. It’s intellectual honesty.</p><h3>What I’d Do Differently at Work</h3><p>If I had this framework inside my actual job, here’s how I’d operationalize it:</p><ol><li><strong>Set up a real-time scoring pipeline</strong> — score every wallet at claim time, not after</li><li><strong>Flag clusters, not just individuals</strong> — coordinated fraud leaves network patterns; look for wallets that share behavioral fingerprints</li><li><strong>Tune the threshold with a business owner</strong> — deciding what precision/recall trade-off to accept is a product decision, not a data decision</li><li><strong>Build feedback loops</strong> — confirmed fraud cases should retrain the model continuously</li><li><strong>Document everything</strong> — compliance and legal teams need to understand <em>why</em> a wallet was flagged, not just that it was</li></ol><h3>Key Lessons</h3><ul><li><strong>Working in the industry gave me the domain knowledge. The side project gave me the proof.</strong></li><li><strong>Feature engineering beats algorithm selection.</strong> A smart feature in a simple model will outperform a dumb feature in a complex one.</li><li><strong>Labeling is a design problem.</strong> Treat it with the same rigor you’d give to any analytical framework.</li><li><strong>Precision/recall is a conversation, not a calculation.</strong> The right threshold depends on what kind of mistakes your business can tolerate.</li><li><strong>Reproducibility is a professional signal.</strong> A clean repo with documented logic tells a hiring manager more than a Kaggle leaderboard score.</li></ul><blockquote><em>“Fraud doesn’t hide in the data. It hides in the behavior. Your job as a data analyst is to make behavior visible.”</em></blockquote><p>That’s the lens I brought from my time in crypto. And it’s what made this project feel real rather than academic.</p><h3>If You’re a Data Analyst Thinking About a Side Project</h3><p>Don’t look for a dataset first.</p><p>Look for a problem you already understand.</p><p>The advantage you have over someone following a tutorial is lived context. You know what the anomalies feel like before you can articulate why. That intuition is the starting point — the data and the code come after.</p><p>I built this because I worked in crypto and saw something that bothered me. The side project was just my way of thinking it through properly.</p><p>That’s enough of a reason to start.</p><p><em>Full project on GitHub: </em><a href="https://github.com/Zhonghong99/crypto-airdrop-fraud-detection-ml"><em>crypto-airdrop-fraud-detection-ml</em></a><em> — notebooks, dataset, and data quality scripts included.</em></p><h3>📤Thanks for reading! If you found this useful…</h3><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e549b123f592" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[7 Smart Ways Data Analysts Use ChatGPT to Save Time]]></title>
            <link>https://medium.com/@zhonghong9998/7-smart-ways-data-analysts-use-chatgpt-to-save-time-6022143e958d?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/6022143e958d</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Thu, 02 Apr 2026 12:41:01 GMT</pubDate>
            <atom:updated>2026-04-08T15:42:12.395Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QvlZN3KzuiNHxPlyUtgDvQ.png" /><figcaption>By Zhong Hong</figcaption></figure><p>A while ago, I watched a data analyst spend 45 minutes doing something ChatGPT could have helped her do in 8.</p><p>Not writing the final SQL.<br>Not building the final dashboard.<br>Just untangling the mess before the real work began.</p><p>That’s the part most analysts underestimate.</p><p><strong>ChatGPT is not just a faster Google. It’s a friction remover.</strong></p><p>And for data analysts, friction is everywhere.</p><h3>The Real Problem: Data Analysts Don’t Usually Lose Time on “Analysis”</h3><p>Most analysts don’t burn their day on the final insight.</p><p>They burn it on everything around the insight.</p><p>Things like:</p><ul><li>translating vague stakeholder requests</li><li>cleaning up broken SQL logic</li><li>writing repetitive documentation</li><li>turning messy notes into structured summaries</li><li>debugging small but annoying errors</li><li>switching between technical work and communication work</li></ul><p>The work is not always intellectually hard.</p><p>It’s just fragmented.</p><p>And fragmented work is where time gets wrecked.</p><h3>My Turning Point</h3><p>I used to think ChatGPT was mainly useful for writing.</p><p>Emails. Summaries. Maybe a quick explanation.</p><p>But the real unlock came when I stopped treating it like a content tool and started treating it like a thinking assistant.</p><p>That changed everything.</p><p>Because the best use of ChatGPT is not to replace analytical judgment.</p><p>It’s to reduce the setup cost of analytical work.</p><blockquote><strong><em>The smartest analysts don’t use ChatGPT to avoid thinking. They use it to get to the thinking faster.</em></strong></blockquote><h3>1. Turning Vague Requests Into Clear Analytical Questions</h3><p>One of the biggest time drains in analytics is bad problem framing.</p><p>A stakeholder says:</p><p>“We want to understand customer behavior better.”</p><p>Cool. That means absolutely nothing.</p><p>Strong data analysts know that before SQL, before dashboards, before any modeling, the real job is to sharpen the question.</p><p>This is where ChatGPT becomes useful.</p><p>I’ll drop in a raw stakeholder request and ask it to help me turn it into:</p><ul><li>a clear business question</li><li>possible KPIs</li><li>dimensions to segment by</li><li>follow-up questions I should ask before starting</li></ul><p>That doesn’t replace my judgment.</p><p>It gives me a faster first draft of the analytical framing.</p><p>And that matters because half of bad analysis comes from a bad starting question.</p><h4>Example angle</h4><p>Instead of starting with<br>“analyze user retention,”</p><p>you quickly get to:</p><ul><li>What retention definition are we using?</li><li>Cohort by signup date or first purchase?</li><li>Are we measuring D7, D30, or rolling retention?</li><li>Should we segment by channel, region, or product type?</li></ul><p>That’s not a small improvement.</p><p>That’s the difference between useful analysis and dashboard cosplay.</p><h3>2. Drafting SQL Logic Before Touching the Database</h3><p>Let’s be honest.</p><p>A lot of SQL work is not typing.<br>It’s thinking through joins, filters, edge cases, grouping logic, and granularity.</p><p>ChatGPT is surprisingly useful here.</p><p>Not because it always writes perfect SQL. It doesn’t.</p><p>But because it helps me think through the skeleton faster.</p><p>I’ll ask it to draft:</p><ul><li>the rough query structure</li><li>the likely joins</li><li>window function ideas</li><li>edge cases that might create duplicates</li><li>how to validate the result</li></ul><p>That last one is underrated.</p><p>A mediocre analyst asks, “Can ChatGPT write the query?”</p><p>A strong data analyst asks, “Can ChatGPT help me pressure-test the query before I trust it?”</p><h4>Why this works</h4><p>When you’re staring at a blank SQL editor, the hardest part is often the first 30%.</p><p>Once the structure is there, your own experience takes over.</p><p>ChatGPT is not the pilot.<br>It’s the intern who gets the whiteboard started.</p><h3>3. Debugging Faster When the Logic Feels Off</h3><p>Sometimes the query runs.</p><p>That’s the dangerous part.</p><p>Because “it runs” is not the same as “it’s right.”</p><p>This is where analysts lose hours.</p><p>Not on syntax errors.<br>On silent logic errors.</p><p>Wrong grain.<br>Accidental duplication after join.<br>Bad date filter.<br>Incorrect denominator.<br>A metric that looks plausible but is completely wrong.</p><p>I use ChatGPT to explain my query logic back to me in plain English.</p><p>That sounds simple, but it’s insanely effective.</p><p>When the logic is rewritten clearly, broken assumptions become obvious.</p><h4>Example prompt direction</h4><p>You can paste in a query and ask:</p><blockquote>“Explain what this query is actually calculating, step by step. Where could duplicates or metric inflation happen?”</blockquote><p>Now you’re using ChatGPT less like a generator and more like a reviewer.</p><p>That’s a much smarter use case.</p><h3>4. Summarizing Data Findings for Non-Technical Stakeholders</h3><p>This is one of the highest ROI uses.</p><p>A lot of analysts can find insights.</p><p>Far fewer can communicate them cleanly.</p><p>And if stakeholders don’t understand your insight, the analysis might as well be a decorative PNG.</p><p>I’ll often feed ChatGPT:</p><ul><li>a few bullet-point findings</li><li>chart context</li><li>stakeholder audience</li><li>tone I want</li></ul><p>Then I’ll ask it to turn that into:</p><ul><li>an executive summary</li><li>a Lark update</li><li>a meeting recap</li><li>a concise insight memo</li></ul><p>The key is that I do <strong>not</strong> ask it to invent conclusions.</p><p>I give it the facts.</p><p>It helps me package them.</p><blockquote><strong><em>Analysis creates value only when someone else can act on it.</em></strong></blockquote><p>That’s where a lot of technically strong analysts still fumble.</p><h3>5. Creating Documentation Without Hating Your Life</h3><p>Nobody becomes a data analyst because they love writing documentation.</p><p>But bad documentation quietly taxes everyone.</p><p>Especially when teams scale.</p><p>ChatGPT is great for turning rough technical notes into cleaner artifacts like:</p><ul><li>metric definitions</li><li>dashboard descriptions</li><li>SQL query comments</li><li>data validation checklists</li><li>project handover notes</li></ul><p>This is especially useful after a long project when your brain is fried and the last thing you want to do is explain your own work.</p><p>I’ve found this matters more than people think.</p><p>Because one of the most valuable analysts on any team is not just the one who solves problems.</p><p>It’s the one who leaves behind clarity.</p><p>That person compounds team speed.</p><h3>6. Speeding Up Exploratory Thinking</h3><p>This one is subtle, but powerful.</p><p>Sometimes you don’t need an answer yet.</p><p>You need angles.</p><p>You need hypotheses.</p><p>You need better ways to think about the problem.</p><p>ChatGPT is useful for idea expansion during EDA or problem exploration.</p><p>I’ll ask things like:</p><ul><li>what hypotheses could explain this trend?</li><li>what dimensions should I segment by?</li><li>what confounders might be affecting this metric?</li><li>what are possible reasons this experiment underperformed?</li></ul><p>Again, this is not about outsourcing judgment.</p><p>It’s about widening the search space faster.</p><p>A lot of analysts get stuck because they only explore the first obvious path.</p><p>ChatGPT can help you surface the second, third, and fourth path sooner.</p><p>And that’s where better analysis often lives.</p><h3>7. It Helps Me Handle the Mess Around the Analysis</h3><p>One of the most useful ways I use ChatGPT is not for one big task.</p><p>It’s for all the small tasks around the analysis.</p><p>For example, in one project, I might need to:</p><ul><li>turn a vague stakeholder message into a clear analysis question</li><li>sketch the first draft of a SQL approach</li><li>sanity-check why the numbers look off</li><li>rewrite my findings into a simple business summary</li><li>turn rough notes into clean documentation</li></ul><p>None of these tasks is huge on its own.</p><p>But switching between them is tiring.</p><p>I might spend 30 minutes writing SQL, then suddenly need to reply to a stakeholder, then go back to debugging, then prepare a short summary for a meeting.</p><p>That constant switching slows me down more than I realize.</p><p>So instead of carrying all of that in my head, I use ChatGPT as a working partner throughout the project.</p><p>For example:</p><ul><li>I paste in a stakeholder request and ask, “What are the key business questions hidden here?”</li><li>I paste in my SQL logic and ask, “Where could duplicates or wrong aggregation happen?”</li><li>I paste in my bullet points and ask, “Turn this into a short update for a product manager.”</li><li>I paste in rough notes and ask, “Convert this into clear documentation with sections.”</li></ul><p>This is where the time savings become real.</p><p>Not because ChatGPT does the analysis for me.</p><p>But because it helps me move faster through all the support work around the analysis.</p><p>It reminds me of the <a href="https://gettingthingsdone.com/2010/07/10-big-ideas-from-gtd/">GTD idea</a>: <strong>your brain is better for thinking than for holding too many open loops.</strong></p><p>ChatGPT helps me get those open loops out of my head and into something I can work with.</p><p>That means less friction, less restarting, and more energy for the part that actually matters: solving the problem.</p><h3>A Real Analyst Workflow Example</h3><p>Here’s what this looks like in practice.</p><p>Say a product manager asks:</p><p>“Can we understand why weekly active users dropped last month?”</p><p>A smart ChatGPT-assisted workflow might look like this:</p><p>First, use it to sharpen the business question:<br>What exactly counts as active? Which users? Which markets? Which time window?</p><p>Then, use it to sketch analysis directions:<br>segment by acquisition channel, geography, app version, device, and cohort.</p><p>Next, use it to help draft SQL structure:<br>CTEs for weekly aggregation, lag comparison, and segment deltas.</p><p>Then, use it to review the logic:<br>check for tracking changes, denominator shifts, and missing event data.</p><p>After that, use it to summarize the findings:<br>“WAU decline was concentrated in Android users in one region after version X release.”</p><p>Finally, use it to turn notes into a stakeholder update:<br>clear, concise, action-oriented.</p><p>Same analyst.<br>Same brain.<br>Better leverage.</p><h3>What ChatGPT Should Not Do for Analysts</h3><p>Let’s not get delusional.</p><p>There are things analysts should be careful about.</p><p>ChatGPT should not be trusted blindly with:</p><ul><li>final metric definitions</li><li>production SQL without review</li><li>business conclusions without source evidence</li><li>sensitive data pasted carelessly into prompts</li><li>statistical reasoning you don’t understand yourself</li></ul><p>This is where lazy usage backfires.</p><p>Because the danger is not that ChatGPT is useless.</p><p>The danger is that it’s useful enough to make bad analysts overconfident.</p><p>That’s a brutal combo.</p><h3>The Bigger Lesson</h3><p>The analysts who benefit most from ChatGPT are not the ones trying to automate their whole job.</p><p>They’re the ones who understand their job well enough to know where the bottlenecks actually are.</p><p>That’s the move.</p><p>Use ChatGPT for:</p><ul><li>structure</li><li>speed</li><li>synthesis</li><li>communication</li><li>first drafts</li></ul><p>But keep the judgment for yourself.</p><p>Because your real value is not typing faster.</p><p>It’s knowing what matters, what’s wrong, and what to do next.</p><blockquote><strong><em>ChatGPT won’t make a weak analyst great. But it can make a strong analyst a lot faster.</em></strong></blockquote><h3>Closing Insight</h3><p>The future of analytics probably won’t belong to the analysts who resist AI.</p><p>And it also won’t belong to the ones who blindly depend on it.</p><p>It’ll belong to the analysts who learn how to collaborate with it well.</p><p>That means knowing when to use it.</p><p>Knowing when not to.</p><p>And knowing that the real edge is not automation alone.</p><p>It’s amplified judgment.</p><p>Because in the end, stakeholders don’t pay for prompts.</p><p>They pay for clarity, decisions, and outcomes.</p><p>And that part is still on you.</p><p>If you’re a data analyst, the best thing you can do this week is simple:</p><p>Pick one annoying, repetitive part of your workflow and test ChatGPT on it.</p><p>Not to replace your work.<br>To remove friction from it.</p><p>That’s where the real gains start.</p><h3>📤Thanks for reading! If you found this useful…</h3><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6022143e958d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Most Data Analysts Use ChatGPT for the Easy Stuff. That’s Why They Stay Slow]]></title>
            <link>https://medium.com/@zhonghong9998/most-data-analysts-use-chatgpt-for-the-easy-stuff-thats-why-they-stay-slow-a538a0fded83?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a538a0fded83</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Tue, 31 Mar 2026 13:16:01 GMT</pubDate>
            <atom:updated>2026-03-31T13:16:01.319Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CIdvEw0b3hHZ0xS-N322gg.png" /><figcaption>By Zhong Hong</figcaption></figure><p>I’ve seen a lot of data analysts use ChatGPT like a fancy search bar.</p><p>They paste in an error message.<br>Ask for an Excel formula.<br>Maybe get a quick SQL fix.</p><p>That feels productive.</p><p>But it’s also kind of a trap.</p><p>Because if you only use ChatGPT for the small stuff, you’re saving minutes — while missing the chance to save hours.</p><blockquote><strong><em>Most analysts use ChatGPT like a shortcut. The best analysts use it like a thinking partner.</em></strong></blockquote><h3>The real problem: analysts are underusing leverage</h3><p>A lot of data analysts think using ChatGPT means:</p><ul><li>fixing syntax</li><li>rewriting emails</li><li>summarizing meetings</li><li>explaining Python errors</li><li>generating basic SQL queries</li></ul><p>None of that is bad.</p><p>But let’s be honest: that’s entry-level leverage.</p><p>It helps you move a little faster.<br>It does not change the level you operate at.</p><p>The analysts who stay slow usually make one big mistake:</p><h3>They only use ChatGPT after they already know what to do</h3><p>That means they’re using it as a helper.</p><p>Not as a collaborator.</p><p>And that’s why their workflow still looks like this:</p><ul><li>think alone</li><li>get stuck alone</li><li>try random fixes</li><li>search 19 tabs</li><li>ask ChatGPT one tiny question</li><li>repeat until mentally cooked</li></ul><p>That’s not smart work.<br>That’s just chaos with Wi-Fi.</p><h3>My turning point: I stopped asking for answers first</h3><p>What changed for me was simple.</p><p>I stopped treating ChatGPT like a vending machine for instant outputs.</p><p>Instead, I started using it the same way I’d use a strong teammate:</p><ul><li>describe what I’m trying to do</li><li>talk through the messy context</li><li>go back and forth until the problem is clearer</li><li>align on the exact goal</li><li>then ask for the plan</li><li>then ask it to execute the coding part</li><li>then keep asking questions anywhere I’m confused</li></ul><p>That one shift changed everything.</p><p>Because the biggest bottleneck in analysis work usually isn’t typing.</p><p>It’s thinking clearly.</p><p>And ChatGPT becomes way more useful when you bring it in <strong>before</strong> the task is fully formed.</p><h3>The better way: use ChatGPT across the full workflow</h3><p>Here’s the framework I use now.</p><h4>1. Start with the messy brief</h4><p>Before asking for code, I explain the situation first.</p><p>I’ll say things like:</p><ul><li>what I’m trying to build</li><li>what the business goal is</li><li>what data I have</li><li>what feels confusing</li><li>what constraints matter</li><li>what “good” output should look like</li></ul><p>At this stage, I’m not asking for final answers yet.</p><p>I’m pressure-testing the problem.</p><p>This matters because vague input creates fake productivity.</p><p>You’ll get an answer fast.<br>But it might be answering the wrong question.</p><h4>2. Go back and forth until the goal is sharp</h4><p>This is the underrated part.</p><p>I’ll challenge the direction with ChatGPT until we reach clear alignment.</p><p>That might include:</p><ul><li>refining metric definitions</li><li>spotting blind spots</li><li>identifying missing assumptions</li><li>comparing different approaches</li><li>clarifying what I actually need versus what sounds impressive</li></ul><p>This is where a lot of analysts skip too early.</p><p>They want the output immediately.</p><p>Bad move.</p><p>Because if your thinking is fuzzy, your dashboard, notebook, or SQL logic will just become a beautifully formatted mess.</p><blockquote><strong><em>Clarity first. Speed second. Otherwise you’re just automating confusion.</em></strong></blockquote><h4>3. Ask for a plan only after alignment</h4><p>Once the objective is clear, then I ask for the plan.</p><p>Not before.</p><p>Now the plan is actually useful because it’s built on shared context.</p><p>A solid plan usually includes:</p><ul><li>problem framing</li><li>step-by-step approach</li><li>logic flow</li><li>edge cases to watch</li><li>what to validate</li><li>what to deliver first</li></ul><p>This is where ChatGPT stops being a chatbot and starts feeling like a project copilot.</p><h4>4. Let ChatGPT do the coding heavy lifting</h4><p>Once the path is agreed, I’ll ask it to complete the coding part.</p><p>That can mean:</p><ul><li>SQL queries</li><li>pandas transformations</li><li>feature engineering steps</li><li>data cleaning scripts</li><li>Power BI DAX logic</li><li>chart-building code</li><li>reusable functions</li><li>notebook structure</li></ul><p>This is the part most people think is the main use case.</p><p>It’s not.</p><p>It’s only powerful <strong>because the thinking before it got stronger</strong>.</p><p>Code generated from a bad understanding is still bad code.</p><p>Just faster.</p><h4>5. Ask for clarification aggressively</h4><p>Whenever anything feels off, I ask.</p><p>Not later. Immediately.</p><p>I’ll ask things like:</p><ul><li>Why did you choose this approach?</li><li>What assumption is this based on?</li><li>What happens if there are nulls?</li><li>Why use a left join instead of inner join?</li><li>Can you simplify this logic?</li><li>Show me with a small example table.</li><li>What’s the business risk if this is wrong?</li></ul><p>That part matters a lot.</p><p>Because using ChatGPT well is not passive.</p><p>You don’t just copy-paste and pray.</p><p>You interrogate the output until it makes sense.</p><p>That’s how you learn faster <strong>and</strong> ship better work.</p><h3>Real example: two analysts, same task, very different speed</h3><p>Let’s say both analysts need to build a churn analysis workflow.</p><h4>Analyst A uses ChatGPT for the easy stuff</h4><p>They ask:</p><ul><li>“write sql for churned users”</li><li>“fix this error”</li><li>“explain group by”</li><li>“write python chart code”</li></ul><p>They get fragments.</p><p>Useful fragments, sure.</p><p>But they still have to stitch everything together themselves while figuring out the business definition on the fly.</p><p>Result:</p><ul><li>more backtracking</li><li>more inconsistent logic</li><li>more rework</li><li>more hidden mistakes</li></ul><h4>Analyst B uses ChatGPT like a collaborator</h4><p>They start with:</p><p>“I want to analyze churn for a subscription product. Here’s how I think churn should be defined, here’s the data available, here’s what stakeholders care about, and here’s what I’m unsure about.”</p><p>Then they go back and forth.</p><p>Then they ask for a plan.</p><p>Then they ask for the SQL, Python, and interpretation support.</p><p>Then they question the logic until it’s clean.</p><p>Result:</p><ul><li>clearer metric definition</li><li>better logic before code</li><li>faster execution</li><li>fewer dumb mistakes</li><li>stronger business explanation</li></ul><p>Same tool.</p><p>Different game.</p><h3>Why this makes analysts faster in the long run</h3><p>Using ChatGPT this way improves more than output speed.</p><p>It improves decision speed.</p><p>That’s the real flex.</p><p>You get faster at:</p><ul><li>framing ambiguous problems</li><li>turning messy requests into clear tasks</li><li>seeing edge cases earlier</li><li>structuring analysis logically</li><li>learning while doing</li><li>reducing rework</li></ul><p>That last one is huge.</p><p>Because “fast” analysts are rarely the ones who type fastest.</p><p>They’re the ones who make fewer wrong turns.</p><h3>What most analysts still get wrong</h3><p>Here are the big misses I keep seeing:</p><h4>They ask tiny questions without giving context</h4><p>So the answer is technically correct but practically useless.</p><h4>They ask for code before defining the problem</h4><p>Classic clown behavior. Respectfully.</p><h4>They treat the first answer like truth</h4><p>Instead of challenging assumptions and checking logic.</p><h4>They use ChatGPT only to reduce effort</h4><p>Instead of increasing quality and speed at the same time.</p><h4>They don’t build a repeatable workflow</h4><p>So every task starts from zero again.</p><p>That’s the opportunity cost.</p><p>If you only use ChatGPT as a calculator, you’ll keep working like a calculator.</p><h3>The framework I’d recommend to any analyst</h3><h4>Use this sequence:</h4><ol><li><strong>Describe the task in plain English</strong></li><li><strong>Go back and forth until the goal is fully clear</strong></li><li><strong>Ask for the plan</strong></li><li><strong>Ask for full execution on the coding parts</strong></li><li><strong>Ask clarification on anything confusing</strong></li><li><strong>Refine until the logic is both correct and usable</strong></li></ol><p>That’s it.</p><p>Simple, but not basic.</p><blockquote><strong><em>If you only use ChatGPT to save effort, you’ll stay average. If you use it to sharpen thinking, you become dangerous.</em></strong></blockquote><h3>Closing insight</h3><p>The point is not to outsource your brain.</p><p>The point is to upgrade how your brain works with tools.</p><p>That’s the difference.</p><p>Weak analysts ask ChatGPT to finish tasks faster.<br>Strong analysts use ChatGPT to think better, decide faster, and execute with less waste.</p><p>And in data work, that compounds hard.</p><p>Because once you can turn messy problems into clean workflows, you stop looking like “the person who writes SQL.”</p><p>You start looking like the person who can actually solve things.</p><h3>📤Thanks for reading! If you found this useful…</h3><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a538a0fded83" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If You Keep Making SQL Mistakes, Your Problem May Not Be SQL]]></title>
            <link>https://medium.com/@zhonghong9998/if-you-keep-making-sql-mistakes-your-problem-may-not-be-sql-9dca14d1fa2a?source=rss-c99fe335779c------2</link>
            <guid isPermaLink="false">https://medium.com/p/9dca14d1fa2a</guid>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[sql]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Zhong Hong]]></dc:creator>
            <pubDate>Fri, 27 Mar 2026 13:06:00 GMT</pubDate>
            <atom:updated>2026-03-27T13:06:00.601Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_L4HrqzQDPHzJECvwGZDxw.png" /><figcaption>By Zhong Hong</figcaption></figure><p>In my last article, I talked about t</p><p>he <a href="https://medium.com/@zhonghong9998/956e9e244f02">5 SQL mistakes that keep junior data analysts from leveling up</a>.</p><p>Things like jumping into SQL too fast, ignoring data grain, trusting joins too easily, writing messy queries, and believing the output too quickly.</p><p>But knowing the mistakes is only half useful.</p><p>Because if we are being real, most people do not improve just because they know what not to do.<br>They improve when they have a <strong>practical way</strong> to do things better.</p><p>That is what I want to share here.</p><p>Not some perfect textbook method.<br>Just a simple SQL workflow that helps me think better before, during, and after I write a query.</p><h3>The solution is not better SQL first. It is better thinking first.</h3><p>The biggest shift for me was realizing this:</p><blockquote><strong><em>Good SQL usually starts before SQL.</em></strong></blockquote><p>That sounds obvious, but a lot of junior analysts still skip that part.</p><p>They treat the query editor like a battlefield.<br>As if speed is proof of competence.</p><p>But in real work, the people who level up faster are often the ones who pause at the right moment.</p><p>Not to overthink forever.<br>Just enough to avoid building the wrong thing beautifully.</p><p>So here is the workflow I now use.</p><h3>Step 1: Rewrite the request in plain English</h3><p>Before I touch the database, I try to rewrite the business request in one or two simple sentences.</p><p>Not in technical language.<br>In normal language.</p><p>For example, if someone asks:</p><p>“Can you show monthly active users for the last quarter?”</p><p>I do not let that sentence pass too casually.</p><p>I rewrite it for myself:</p><p>“I need the number of unique users who performed a defined ‘active’ action in each month of the last quarter.”</p><p>That rewrite matters.</p><p>Because vague requests sound harmless at first.<br>But they are where bad analysis quietly begins.</p><p>If I cannot explain the request clearly in plain English, I know I am not ready to write the query yet.</p><h4>What I often clarify here</h4><p>Usually only a few things matter most:</p><ul><li>what exactly the metric means</li><li>which date field should define the time period</li><li>what should be included or excluded</li><li>what level the result should be shown at</li></ul><p>This first step already solves a big part of Mistake #1.</p><p>Because the goal is not to start fast.</p><p>The goal is to start clean.</p><h3>Step 2: State the grain before writing anything</h3><p>This habit has saved me from a lot of nonsense.</p><p>Before I query, I literally ask myself:</p><p><strong>What does one row represent in this table?</strong><br>And: <strong>What does one row need to represent in my final output?</strong></p><p>That is the grain question.</p><p>And grain is one of those things that sounds boring until it ruins your KPI.</p><p>A user table, transaction table, and event table do not behave the same way.<br>If I forget that, I can inflate counts, distort averages, or duplicate records without even noticing.</p><p>So I try to write a quick note like this:</p><ul><li>source table grain = one row per transaction</li><li>final output grain = one row per country per month</li></ul><p>That tiny note helps my brain stay honest.</p><p>Because once the grain is clear, a lot of query decisions become easier.</p><p>This step is what protects me from Mistake #2.</p><p>And frankly, many junior analysts do not need more SQL functions.<br>They need more grain awareness.</p><h3>Step 3: Sketch the logic before writing the full query</h3><p>This is one of the most practical things I do now.</p><p>I do not always jump straight into full SQL anymore.<br>Sometimes I sketch the logic first.</p><p>Just rough steps. Nothing fancy.</p><p>For example:</p><ol><li>filter the relevant date range</li><li>remove test or invalid records</li><li>aggregate to the right grain</li><li>join only what is needed</li><li>calculate the final metric</li></ol><p>This keeps me from getting lost halfway through the query and stitching together random logic later.</p><p>It also helps me see where things might break.</p><p>If the workflow still feels fuzzy at this stage, this is exactly where AI can be useful.</p><p>Not to dump the whole task onto it like a helpless intern manager.<br>But to think with it.</p><p>I might ask:</p><ul><li>“Does this logic flow make sense?”</li><li>“What assumptions am I missing?”</li><li>“Where could duplication happen here?”</li><li>“What would a safer query structure look like?”</li></ul><p>Sometimes AI gives a decent thought starter.<br>Sometimes it says something mid.<br>Either way, I stay responsible for the judgment.</p><p>That is the key.</p><p>Use AI to sharpen your thinking, not replace it.</p><h3>Step 4: Build the query in layers, not in one chaotic block</h3><p>This changed a lot for me.</p><p>Earlier on, I used to write SQL in one long messy swing.</p><p>Filter here.<br>Join there.<br>Case when somewhere in the middle.<br>Aggregate at the end.<br>Pray to the data gods.</p><p>Now I prefer building the query in layers.</p><p>Usually through clean CTEs or clearly separated sections.</p><p>Each part should do one job.</p><p>For example:</p><ul><li>one step to filter raw data</li><li>one step to standardize or clean fields</li><li>one step to aggregate</li><li>one step to join supporting context</li><li>one final step for the output</li></ul><p>This makes the logic easier to debug and explain.</p><p>It also makes mistakes easier to catch, because I can inspect each stage instead of staring at one giant block of SQL like it is ancient scripture.</p><p>This step directly helps with Mistake #4.</p><p>Readable SQL is not just for aesthetics.</p><p>Readable SQL protects thinking.</p><h3>Step 5: Treat joins like risk points, not admin work</h3><p>This one deserves its own mindset.</p><p>I no longer treat joins as a casual step.</p><p>Every join is a possible source of duplication, distortion, or confusion.</p><p>So before I join anything, I ask:</p><p>What is the relationship between these two tables?</p><p>One-to-one?<br>One-to-many?<br>Many-to-many?</p><p>And more importantly:</p><p>What will this do to the row count?</p><p>That question sounds simple, but it catches a lot.</p><p>If I am joining a transaction table to a payments table, I know there is danger there.<br>If I am joining users to events, same story.</p><p>When I am unsure, I often test small counts before and after the join.<br>That alone can reveal whether I am about to create fake growth and embarrass myself for free.</p><p>If needed, I might also use AI here as a second brain.</p><p>For example:</p><p>“I have a users table and an events table. I want one row per user. What duplication risks should I watch for?”</p><p>That kind of prompt can help surface risks I should manually verify.</p><p>Again, AI is not the analyst.<br>I am.</p><p>But it can still help me challenge lazy assumptions.</p><p>This step is the direct antidote to Mistake #3.</p><h3>Step 6: Explain the query back to myself like I am teaching someone</h3><p>This habit is underrated.</p><p>Once the query is mostly done, I try to explain what it does in simple language.</p><p>Not line by line.<br>Logic by logic.</p><p>Something like:</p><p>“First I filtered valid orders from last month. Then I aggregated revenue at the order level. Then I joined the country dimension. Then I summed net revenue by country.”</p><p>If I cannot explain it clearly, that usually means the SQL is still too messy or the logic still is not fully clean.</p><p>Sometimes I even use AI for this part too.</p><p>I paste the query and ask it to explain the logic back in plain English.</p><p>Not because I need it to understand SQL for me.<br>But because if the explanation comes back muddled, that is often a sign my structure is muddled too.</p><p>That is such a practical trick.</p><p>Because clarity in explanation usually reflects clarity in thought.</p><h3>Step 7: Validate before I emotionally commit to the result</h3><p>This is probably the most important step of all.</p><p>Because once we see a number, our brain wants closure.</p><p>We want it to be right.<br>We want to move on.<br>We want to feel done.</p><p>That is exactly why validation matters.</p><p>Now I try to do at least a few checks before I trust the output:</p><ul><li>does the number roughly make sense?</li><li>does the trend look suspicious?</li><li>do sample rows look correct?</li><li>did row counts jump after the join?</li><li>does it align with any known totals or prior reports?</li></ul><p>And when I want a second round of thinking, I may ask AI something like:</p><ul><li>“What are 3 ways this result could still be wrong?”</li><li>“What sanity checks would a strong analyst run here?”</li><li>“Which edge cases am I most likely missing?”</li></ul><p>That helps me think wider.</p><p>Not because AI knows the truth automatically, but because it forces me to pause and challenge my own confidence.</p><p>This is what protects me from Mistake #5.</p><p>And honestly, this is where junior analysts start becoming more senior in how they think.</p><p>Not when they write fancier code.</p><p>When they become harder to fool.</p><h3>A simple example of this workflow in real life</h3><p>Let’s say I get this request:</p><p>“Can you show last month’s revenue by country?”</p><p>Old version of me might have opened SQL instantly.</p><p>New version of me slows down and walks through the workflow.</p><p>First, I rewrite the request clearly.<br>Do we mean gross revenue or net revenue? Are refunds excluded? Which date defines last month?</p><p>Then I define the grain.<br>Maybe my orders table is one row per order, while my payments table has multiple payment events per order.</p><p>Then I sketch the logic.<br>Filter last month’s valid orders, calculate revenue at the correct level, map country correctly, then aggregate.</p><p>Then I build in layers.<br>Not one giant mess.</p><p>Then I validate.<br>Do totals roughly match finance? Does country mapping look weird? Are there duplicated orders after the join?</p><p>That whole process is not “slower.”</p><p>It is safer.</p><p>And safer usually becomes faster over time, because you spend less time cleaning up your own avoidable mistakes.</p><h3>The real upgrade is not technical. It is operational.</h3><p>That is probably the main thing I would want readers to take away.</p><p>A lot of junior analysts think growth comes from learning more advanced SQL syntax.</p><p>That helps, sure.</p><p>But in my experience, the real improvement often comes from building a more reliable operating system around your SQL.</p><p>How you define the question.<br>How you respect grain.<br>How you handle joins.<br>How you structure logic.<br>How you validate.</p><p>Those habits do not look flashy.</p><p>But they compound hard.</p><blockquote><strong><em>The best analysts are not just good at writing queries. They are good at reducing avoidable mistakes before the query is even finished.</em></strong></blockquote><h3>Closing insight</h3><p>In the last article, I focused on the mistakes.</p><p>This time, I wanted to focus on what actually helps.</p><p>Because most analysts do not need more guilt.<br>They need a better process.</p><p>That is what helped me.</p><p>Not perfection.<br>Not genius.<br>Just a cleaner way of thinking before I let myself believe the output.</p><p>And I think that is one of the most practical ways to level up in analytics.</p><p>Not by trying to look impressive.</p><p>By becoming more trustworthy, one query at a time.</p><h3>📤Thanks for reading! If you found this useful…</h3><ul><li>☕ <a href="https://www.buymeacoffee.com/zhonghong99">Buy me a coffee</a> 😁</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9dca14d1fa2a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>