<?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 Aayush Gupta on Medium]]></title>
        <description><![CDATA[Stories by Aayush Gupta on Medium]]></description>
        <link>https://medium.com/@ca.aayushgupta?source=rss-dd01b0c52e97------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*VPbLY7n8-nXZ6W77nlL_Iw.jpeg</url>
            <title>Stories by Aayush Gupta on Medium</title>
            <link>https://medium.com/@ca.aayushgupta?source=rss-dd01b0c52e97------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:37:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ca.aayushgupta/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[Scaffolding Holds the Shape, Engineering Holds the Weight: Three Lines for a High-Growth Business]]></title>
            <link>https://medium.com/@ca.aayushgupta/scaffolding-holds-the-shape-engineering-holds-the-weight-three-lines-for-a-high-growth-business-c5b9b1c9c13c?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5b9b1c9c13c</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[internal-control]]></category>
            <category><![CDATA[ipo]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[compliance]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 17:52:29 GMT</pubDate>
            <atom:updated>2026-11-03T17:52:29.347Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6ThHEKcKM_QAXt6pwEThWQ.png" /></figure><h3>A founder once summarised an entire pre-IPO controls team in four lines and a waveform. Every company I have seen since has drawn the three lines of defence instead, and found that the diagram does no work on its own. One of those drawings is a blueprint. The other is scaffolding.</h3><p><em>By </em><a href="https://aayush-gupta.com/"><em>Aayush Gupta</em></a></p><p>The founder drew four lines and a waveform.</p><p>He was summarising the whole mandate of a controls team at a fintech preparing to list, and he did it on one board in under a minute. Two axes. Two boundary lines, both rising. Between them a wave, climbing up and to the right. At the two points where the wave poked above the top line and dipped below the bottom one, he wrote a single word. Alert.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EbyuxScGo70y-yMfeNQVxw.png" /></figure><p>“My business is growing,” he said, “and my ops and business leaders will move like this wave. You are my two lines. Not to contain them. To alert the leaders and me when they are about to breach a line, or, as happens sometimes, when they already have and need to be brought back within.”</p><p>I have carried a different drawing into a lot of rooms since. The famous three lines, with one statement underneath: impact falls as the control moves closer to the first line. Nobody has ever argued with that statement. Organisational board committees nod hardest of all. Then the company goes away and builds an org chart, with a head for each line and a reporting line into somebody senior, and on the slide it looks finished.</p><p>The three lines you are being asked to build for listing readiness can be drawn as an org structure or engineered as a control system. And trust me, only one of them works.</p><h3>We implemented what you asked for</h3><p>All our theory says the first line owns the control, and on the chart it does. Somewhere between the chart and the code, being the control owner gets confused with being accountable for the control.</p><p>Nobody does anything wrong on the way there. The second line writes a control and hands it to product, engineering or operations, who build it exactly as written, because a control built exactly as written cannot become a finding in the next cycle. When something goes wrong anyway, the answer is polite and completely true. We implemented what you asked for, so there cannot be a finding on this. Ownership sits with the first line. Accountability sits with whoever wrote the requirement, and that person, theoretically and realistically, cannot be accountable for whether a control functions, because the second line exists to provide assurance over it.</p><p>I have written my share of those requirements, and the one that taught me the most was a promotion code.</p><p>Growth teams wrote campaign briefs in business language, around thirty parameters each. Who qualifies, on what, for how much, how often. A brief like that reads perfectly well to a person and means nothing to a system, so somebody else had to turn it into settings in the promotions engine.</p><p>That engine had four hundred and twenty-six settings. Each is narrow and literal: how many times a code may be used per device per week, the daily budget per user, which merchant identifiers it may fire on. A single line of a brief can take four or five of them to express, and the engine has no opinion about whether the combination somebody picked matches the brief somebody wrote. A missing setting does not throw an error. It pays out.</p><p>And so the money went out wrong. A five hundred rupee code written for a small category paid on large appliances. Campaigns fired on merchant IDs outside their own construct. A recharge offer built for one telecom operator paid out on others, because a failed selection was re-initiated from the backend against the last operator that had worked. Each one was implemented exactly as asked, against a brief that was not wrong.</p><p>My answer, at the time, was a dashboard. Configuration mismatches, money spent, top beneficiaries, campaign by campaign, visible to anyone who looked. It was a good dashboard and it found real money. It was also not a control, and it took me a while to admit the difference.</p><h3>An org chart is scaffolding</h3><p>None of that was a failure of structure. The company had a controls office, an internal audit function and a board committee that met on schedule. The structure was complete, and the money still went out wrong.</p><p>An org chart is scaffolding. It holds the shape while you build the thing, and it is not the thing.</p><p>Safety was never in the safety department, and control has never lived in the compliance team. It lives in the field that will not accept a malformed PAN, the screen that holds a withdrawal until the name on the bank account matches the name on the KYC, and the alert that goes to somebody by name when neither of those behaves. Those are built by the people who build the product, and they run whether or not anyone is watching.</p><p>So the question for a company preparing to list is not whether it has the three lines. It is whether anything inside the building is wired.</p><h3>The translator in the middle</h3><p>Wiring it means accepting that org structure is not a control system, and that a control system which has to scale as fast as the product must be written in the first line’s language.</p><p>Business models now change faster than any second line can write controls for them, and expecting one team to mitigate everything is unreal. This is where the real change has to happen, in structure and in method. The second line becomes a translator first.</p><p>Translation means taking a requirement, from a regulator or from the company’s own appetite, and working out what it does inside this particular organisation. Not the second-order effect. The third and the fourth.</p><p>On one account I worked as a consultant, a retail chain had done the first part properly. Its delegation of authority was written down, and the approval thresholds were configured into the system so it would enforce them by itself. Then the purchase orders began arriving in pieces, each sitting just under the limit that would have sent it upward. Nobody broke a rule as written. Splitting a payment to stay below a threshold is the most ordinary thing people do when a number stands between them and their day, and it is the first thing a translator should expect.</p><p>A chief executive once asked me what would actually make his people work compliantly, and offered to put it into everybody’s OKRs. It was a generous offer and the wrong instrument. An OKR makes compliance a target somebody carries for a quarter. What changes a company is narrower: every person understanding that implementing a control, and keeping it working, belongs to them rather than to the risk manager who asked for it. Do that and you see the change. Put it in an OKR and you see a number.</p><p>Which is why, when I scoped a product-controls programme at the same company, the plan said it in a line. The product managers would write the controls for their own modules, I would run a session on how to write a control, and then work with the analysts on the test steps. The second line teaching the first line to write its own controls is the whole thesis in one sentence.</p><p>Do that for a few cycles and the second line stops being the team that writes controls. It becomes a walking risk register, filled to the brim with context, working out appetite and residual exposure on the fly and deciding what matters this week. It stops asking tough questions and starts asking real ones. What happens when a player’s location cannot be fetched for a cash game? What should the name-match score be between a borrower and the bank account the loan is about to land in?</p><h3>A dashboard is not an alert</h3><p>Only then does the second line earn the part everyone thinks it does from day one: the monitoring, alerting, prioritising and remediating layer.</p><p>And this is where the next fallacy sits, the one my promotions dashboard walked straight into. A dashboard is not an alert, and an alert is not a control.</p><p>A dashboard is for study, with context, by somebody who knows what normal looks like. An alert should fire for a critical error, or for the signs of one brewing in the system, and for nothing else. Most importantly, an alert has to mean an action, or a sequence of actions, that ends in a control doing its job. An alert that ends in a look is a dashboard with a worse delivery mechanism.</p><p>That office eventually got it right, at scale. More than fifty live reports across six and more businesses, from payments and commerce to lending, refreshed every week, each owned by the business team that had to act on it. The discipline that made it work was unglamorous: the reports that kept finding nothing were switched off rather than admired. And the promotions problem was finally closed not by my dashboard but by restricting who could configure a campaign at all, which is the sort of fix only the first line can build.</p><p>That loop has a shape. Know what controls exist, module by module, each with the alert that watches it. Test them on a calendar rather than on a hunch. Report the fixes into the product roadmap. Product owns the control, risk owns the assurance, and what comes out of the far end is the only thing the committee actually wanted.</p><h3>Fewer controls, each one owned</h3><p>Once the alert has to end in an action, what counts as a control gets wider.</p><p>A control is not only the step that prevents a risk. It can be a process that reduces the impact, delays it, or shortens the recovery. Our fraternity has a jargon for each of those, from velocity to recovery time objectives, and underneath every one of them is a requirement that can be translated and engineered by the people who run the process.</p><p>It also makes the case for fewer of them, which is the hardest sell in this profession. My own belief is that frauds survive less often because a company had too few controls than because it had too many, stacked on top of each other with nobody owning any single one. A control with four owners has none, and the honest fix is usually to retire three and put a name against the fourth.</p><p>The same test, turned on my own register, did not flatter it. The group framework was clear about what every row had to carry: a named owner, how often the control runs, whether it prevents or detects, and a rating of how well it worked, backed by a defined test. Our rows carried the first three in patches. Several controls had no owner written against them. The effectiveness rating, the column that answers whether the thing actually works, was blank the whole way down.</p><p>That did not mean the controls were failing. It meant the evidence that they worked was living in the alert loop and the monthly reviews, and the register had quietly become a description of the estate rather than an instrument for running it.</p><p>The register was scaffolding too.</p><h3>Who finds out, and how fast</h3><p>At about this point a CFO makes a fair objection, and it deserves to be put in their own words. We have the three lines. We have a risk committee, a compliance head, an internal audit function with a charter. The structure is in place, so what we need now is time.</p><p>Most of that deserves to be granted. The structure is genuinely necessary, a listed parent or a regulator will ask for exactly that diagram, and the people in those seats are usually excellent. None of what goes wrong underneath it is the auditor’s fault, or the compliance head’s. It is what happens when a model built for assurance is asked to do detection.</p><p>But time is not the variable, and one question settles it. When a control fails, does the first line find out from a running alert with a named owner and a clock, or from an auditor months later?</p><p>If the honest answer is the auditor, the company is learning about its failures on the auditor’s calendar, which is the one schedule fraud does not consult. That is not a verdict on anybody in those three boxes. It is a reading of what has been wired and what has only been drawn.</p><p>I still carry the three lines into rooms, and I still write the same statement underneath. What I have come to understand is that the founder was not drawing an org chart that day, and he was not drawing a metaphor either. He was handing us a blueprint. A business that will keep growing, two lines that rise with it instead of holding it down, and a single word at every crossing, meaning somebody moves.</p><p>Build that, and the second line finally gets to do what everyone already believes it does all day.</p><p>Sip tea.</p><p><em>Aayush Gupta works in risk, controls, compliance, and Trust &amp; Safety for high-growth, digitally native companies across gaming, payments, and digital lending. More at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5b9b1c9c13c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Wet Towel and the Courtier]]></title>
            <link>https://medium.com/@ca.aayushgupta/the-wet-towel-and-the-courtier-e215e8fc748f?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/e215e8fc748f</guid>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[governance]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Wed, 26 Aug 2026 17:01:03 GMT</pubDate>
            <atom:updated>2026-08-26T17:01:03.459Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qHIczWyK5YLwhUfsyynajw.png" /></figure><h4>Once, the person who watched for what could go wrong was the most trusted voice in the room. Somewhere along the way the counsel turned into a caveat, and the caveat turned into a formality that gets bought and quietly filed. The way back is older than it looks.</h4><p><em>By </em><a href="https://aayush-gupta.com/"><em>Aayush Gupta</em></a></p><p>A friend who runs a fast-growing lending business calls his risk and compliance head a wet towel. Nothing gets past him. Every deal comes back damp. He says it the way you complain about a strict parent, with a laugh and a little real irritation.</p><p>What rarely comes up in the same conversation is that his book keeps growing and his commissions keep landing, and some part of that is the towel doing its work. The fires that did not start and the frauds that never scaled leave nothing behind to point at. His best work leaves no trace, so it earns no thanks, and he made his peace with that a long time ago. It is the quiet economy of the job. You are remembered for the disasters that arrive and seldom for the ones that never do.</p><p>The fear, though, is real, and it is worth being honest about where it comes from. Founders who completely understand why risk and compliance matters still tense up when the risk person walks in. Not because they doubt the work. Because the work has come to feel like a tax on momentum.</p><p>It is worth telling the story of how that happened, because it did not begin here, and it does not have to end here.</p><h3>The person who used to have the king’s ear</h3><p>Go back far enough and this was the most coveted seat in the building. Every court kept one. Not the treasurer, not the general, but a single trusted person whose whole job was to keep his eyes and ears open, connect the dots the specialists could never see past their own walls, and tell the ruler the truth about what was coming.</p><p>In the Mauryan court, Chanakya took a nobody and made him an emperor, then wrote the manual for running the state he had built. In Akbar’s darbar, Birbal was the voice the emperor kept closest, valued for candour over flattery, and he put his own neck on the line for his king more than once. In Tudor England, Thomas Cromwell rebuilt the machinery of an entire kingdom for Henry VIII, until the same king turned on him. William Cecil steadied Elizabeth I for forty years, the quiet hand behind the reign, rarely loud and never absent.</p><p>Strip the costumes away and these may just be the risk and controls people of their age. They watched for what could go wrong. They held the whole picture while everyone around them held a fragment. And their word was received as counsel, something a ruler acted on, not as a caveat to be filed and forgotten. That last part is what the profession lost.</p><h3>How the counsel became a caveat</h3><p>Somewhere across the centuries the trusted advisor shrank into a formality. It is worth asking why, and the honest answer implicates my own side of the table first.</p><p>A risk manager gets a call wrong once and the business remembers it for years. So the language quietly changes the next time, and the clean recommendation grows a hedge, then a sub-clause beneath the hedge.</p><p>Over a career this hardens into habit, and it is worth being fair about how it forms. A person hired to give clear counsel gradually learns to give cover instead, and cover comes with a vocabulary of its own. Precise terminology is genuinely useful, which is exactly what makes it such a comfortable place to shelter, because nobody can accuse you of sloppiness when every sentence is technically correct. We told ourselves this was rigour, and a great deal of it honestly was. Some of it was also protection. The same instinct built the aura around the role, the studied distance that lets a “no” pass unchallenged and signals that this person is carrying a heavier load than anyone else in the room. Very little of that comes from arrogance, whatever it looks like from the outside. It comes from being the one who gets named when something goes wrong, and from the entirely human wish not to be named that way again. The shelter works. It also has a price, because language built so that it can never be held against you is language nobody can act on either.</p><p>And then the loop closes on everybody. The more we hedge, the less the founder listens. The less the founder listens, the more exposed we feel, so the more we hedge. Both sides slowly prove each other right. The founder concludes that risk is a brake. The risk manager concludes that the founder is reckless. Nobody is lying. The trust just leaves the room without a sound.</p><h3>The fear of the two, and the loss of the eight</h3><p>Under all the jargon sits one small and very human fear, and saying it plainly is where the cure starts.</p><p>A good risk manager carries an unusual amount of context. The intent behind the product, the shape of the last incident, the mood of the regulator this quarter, the shortcut the ops team takes when nobody is looking. With that much in your head, you can be right most of the time. Call it eight times out of ten. But the profession has trained itself to be so afraid of the two it might get wrong that it stops making the eight it would get right. Every call gets softened into mush so that none of them can ever be pinned on you. The dread of being wrong twice quietly forfeits the value of being right eight times. It is a poor trade, and risk people are the ones who keep making it.</p><h3>Why the stamp gets bought</h3><p>Faced with a function that hedges, a founder does something rational. He asks for a stamp.</p><p>Sometimes that stamp is external and sometimes it is internal, and often it arrives conditional, wrapped in qualifying language of its own. It has real value. It brings method, an outside pair of eyes, and a documented basis for a decision that was going to be made anyway. What no stamp can do is carry the decision with you afterwards. Most founders understand this perfectly well, and they will say privately that one person with real context and real ownership would move things further than a bound report does. What stops them is that they have so rarely met that person. They have met the wet towel, and they have met the formality, and the trusted advisor has become genuinely hard to find.</p><h3>Conscience and context</h3><p>So what is the thing they are quietly looking for, under all the noise?</p><p>Take away the frameworks and the acronyms and top-class risk management comes down to two words. Conscience and context. The judgment to see what is actually right, and the awareness of the exact situation in front of you. Everything else, the models and the maturity ladders and the three-letter standards, is vocabulary that grew over those two words and slowly buried them.</p><p>You can hear the difference in how the counsel sounds. A trusted advisor says yes far more often than the wet towel ever does, because he knows which risks are cheap to carry, and he clips a light guardrail onto them instead of dropping a heavy gate. When he does say no, he does not stop at no. He hands over a simpler route, a version of the idea that keeps most of the upside and sheds most of the exposure. A yes that arrives quickly. A no that arrives with a door in it. Any founder recognises that voice inside a single conversation, and it is the sound of someone who has decided to own the outcome.</p><h3>A profession of Chanakyas</h3><p>Chanakya never advised from a safe distance. He picked an untested boy, staked his own standing and his safety on that judgement, and then stayed on to build the state he had argued for. Had the call been wrong, he would not have survived being wrong. Nobody is asking a risk manager to go that far. But in this craft reputation is the nearest thing we have to a life, and the whole argument narrows to one question. Will you stake it?</p><p>Staking it means owning the call. Saying, I have looked at this honestly, the path is sound and the goal is sound, and I will put my name on it and carry the result with you. It means taking a risk on other people’s behalf when you are truly confident, so the business speeds up instead of stalling.</p><p>It also means putting down the belief that the entire world, board and regulator and colleagues, is watching only you and waiting for the slip. Every function in a company is judged on something. Sales carries a number it can miss in public. Engineering carries outages that everyone can see. Scrutiny is the ordinary condition of working somewhere serious, and it was never a special burden handed to risk alone. Carrying it as though it were is what produces the guarded, humourless version of this role, the one that finds it hard to sit in a room and simply be useful to the people in it. Collaboration is not a distraction from the work. Risk is one part of the machine, an important part, and holding the job that way is far lighter than holding up the aura.</p><p>Founders, for their part, are readier for this than the profession assumes. Most will gladly grant the grace to be wrong once in a while, the freedom to change a view when the facts change, the room to be human, on a single condition. That they can see the thinking was genuinely done, the context was weighed, the call was honest. Give them that, and the two out of ten stops being a hanging offence and becomes what it always was in the old courts. The ordinary price of keeping someone close who tells you the truth.</p><p>That is the whole bargain. The advisor owns the outcome. The ruler grants the room. Fear turns back into trust, and the formality turns back into a person worth listening to.</p><p>The wet towel was never really the problem. It was an advisor who had forgotten he was allowed to speak plainly, working for a court that had forgotten how to listen. Both of those have been fixed before. They can be fixed again.</p><p><em>Aayush Gupta works in risk, controls, compliance, and Trust &amp; Safety for high-growth, digitally native companies across gaming, payments, and digital lending. More at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e215e8fc748f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Triage, Not Surveillance: Warehouse Video Analytics Without a Model]]></title>
            <link>https://medium.com/@ca.aayushgupta/triage-not-surveillance-warehouse-video-analytics-without-a-model-b37ae4ee7342?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/b37ae4ee7342</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[warehouse-management]]></category>
            <category><![CDATA[ffmpeg]]></category>
            <category><![CDATA[loss-prevention]]></category>
            <category><![CDATA[video-analytics]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 21:44:51 GMT</pubDate>
            <atom:updated>2026-08-13T21:44:51.789Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vuPVp6Xblgw1qIEIsRKJ1g.jpeg" /></figure><p><em>Most cameras record everything and surface nothing. The gap is not intelligence, it is triage. Here is a method that reduced 277 hours of footage to twenty minutes, measured at every step, with the code in public.</em></p><p><em>By </em><a href="https://aayush-gupta.com/"><em>Aayush Gupta</em></a></p><p>My son keeps a gullak. It is a clay pot the size of a football, and he fills it by charging us for his chores. Every eight to twelve months we break it open and the money becomes a gold coin.</p><p>On a Sunday morning it was gone.</p><p>Two hours of searching the flat found nothing, which left the camera on our landing. Six years old, Xiaomi, an SD card inside. I pulled the card, ordered a reader, and told myself I would run the footage at fifteen times speed and be done before lunch.</p><p>The camera writes one file per minute. Thirteen days had produced 16,294 clips and 277 hours of video.</p><p>What follows is how I got from there to the moment that mattered. It is a domestic story, but the shape of the problem is not. It is the same shape as a warehouse aisle nobody has time to watch, a till nobody reviews until the count is short, and a loading bay with ninety days of retention and no way to search it.</p><h3>The instinct that costs you a day</h3><p>The obvious move is to make it play faster. I lost most of a day there, and the arithmetic is worth writing down because it does not improve with a better laptop.</p><p>At sixteen times speed, 277 hours is still 17.3 hours of watching. That alone should end the conversation, but there is a second problem underneath it. The source runs at twenty frames per second, so sixteen times speed demands 320 frames per second of decoding. A 60Hz display can present sixty. Four frames in five are discarded before your eye ever receives them.</p><p>So even in the best case you are paying seventeen hours of attention to see a fifth of the evidence. Speed was never the constraint. The constraint is that you are asking a machine to show you everything, when what you want is the part where something happened.</p><p>That is a triage problem, and triage has a well understood shape: sort first, look second, and accept a defined risk in exchange for a workable volume.</p><h3>What the camera is actually handing you</h3><p>Before any of this works you have to know what is on the card, because most consumer and prosumer cameras hand you two nasty surprises.</p><p>The first is that a meaningful fraction of the files are broken. An MP4 stores its media payload in one box and its index in another. The index, called moov, holds the byte offsets and decoder configuration that make the file navigable, and the camera writes it only when the clip closes. Cut the power mid clip, which happens constantly, and you get real video data with no index. It is unplayable by normal means.</p><p>In my archive <strong>330 of 16,294 clips were in that state, slightly over two per cent</strong>. This matters more than the percentage suggests, because the tool that joins clips together aborts the entire job on one malformed input. An unfiltered run had essentially no chance of completing. Every clip has to be probed first, and every exclusion logged, before anything else can happen.</p><p>The second surprise is the one that shapes the whole method. Video is not a stack of independent pictures. Compression stores occasional complete frames, called keyframes, and expresses everything between them as changes from what came before. The interval between keyframes is the finest resolution at which you can cheaply sample the footage.</p><p>Measure yours before you design anything:</p><pre>ffprobe -v error -select_streams v:0 -show_entries packet=flags -of csv=p=0 day.mp4 \<br> | awk -F, &#39;{n++; if($1 ~ /^K/) k++} END{printf &quot;1 keyframe per %.1f frames\n&quot;, n/k}&#39;</pre><p>Mine returned one keyframe per 59.2 frames. At twenty frames per second that is <strong>one complete picture every 3.0 seconds</strong>. Every design decision after this point, and every limitation, follows from that number.</p><h3>One flag, and ten times the speed</h3><p>Here is where most of the gain lives, and it is a single argument.</p><p>Scanning 277 hours means decoding. There are three ways to approach it, and I benchmarked all three on the same one hour slice of the same file on the same four cores:</p><p>MethodWhere it actsSpeed277 hours would takeFull decoden/a59.7x realtimeabout 5.2 hours-skip_frame nokeyDecoder157x realtimeabout 2.0 hours-discard nokey<strong>Demuxer590x realtimeabout 28 minutes</strong></p><p>The two flags sound identical and are not. -skip_frame is a hint to the decoder: packets still arrive, headers are still parsed, and only the expensive reconstruction is skipped. -discard acts one layer earlier, at the demuxer, which drops the non key packets so the decoder never receives them at all.</p><p>That distinction is worth 3.8 times against the decoder level flag and ten times against full decode. It is the difference between a method you will actually run and one you will keep meaning to.</p><p>Joining the clips beforehand is similarly cheap if you refuse to re-encode. Copying the compressed stream ran at roughly 3,500 times realtime, 4.17 GB in 19.9 seconds, and the output is bit for bit the camera’s own encoding. Re-encoding the same material would have been about <strong>8.5 days of continuous CPU</strong>, and lossy. Re-encoding here is not a slow option. It is a wrong one.</p><h3>Frame differencing is naive on its own</h3><p>With keyframes arriving cheaply, detection is simple in principle. Compare each keyframe to the previous one. A static aisle scores near zero. A person does not.</p><p>In practice that alone produces an unusable number of false positives, because a camera cannot tell the difference between a person and a cloud crossing the sun. Three gates handle the classes of noise that actually appear, and each exists because of a specific failure I could name.</p><p><strong>The illumination gate.</strong> A light switching on, headlights sweeping a wall, an auto exposure correction: every pixel changes at once and the difference score spikes hard. The tell is that <em>mean luminance</em> moves with it. A person occupies a fraction of the frame and barely shifts the mean. So a large difference accompanied by a large brightness jump is a lighting change, not an intruder. This one matters most on any camera that sees daylight.</p><p><strong>The persistence gate.</strong> An insect crossing the lens, a rain streak, a shadow edge flickering: these trip exactly one keyframe. A human crossing a space is present for several seconds and therefore several keyframes. Requiring an event to span a minimum number of keyframes discards that entire class. The cost is explicit and you should hold it consciously: at a three second sampling grid, requiring two keyframes means an event must last about three seconds to be reported.</p><p><strong>Hysteresis.</strong> One threshold fragments events, because a subject mid stride dips below it for a frame and reappears. Two thresholds fix it, a higher one to open an event and a lower one to keep frames inside it. This is standard edge detection practice, borrowed wholesale.</p><p>These three are the difference between a demo and something you would leave running. They are also the part worth testing, because they are the only place in the pipeline where a judgement is encoded. Mine has twelve assertions covering each failure class, including the case that catches people out: a slow auto exposure ramp must survive the illumination gate, because a subject walking into a dim room produces the same gentle curve.</p><h3>What came out, and what it cost</h3><p>Across thirteen days the pipeline kept <strong>7,238 frames</strong>, grouped into <strong>1,841 distinct events</strong>, rendered as a twenty minute reel at six frames per second. That is a reduction of <strong>827 times</strong>. Precision, sampled by eye across thirty tiles drawn evenly from a six hour window, was about <strong>87 per cent</strong>: twenty six contained a person, and the four misses were door movements with nobody visible.</p><p>The gullak turned up fifteen minutes into reviewing that reel.</p><p>The efficiency did not come from a faster machine. It came from doing less work at every step. One frame in fifty nine is ever decoded. None are re-encoded. The comparison runs at quarter resolution, because shrinking an image before differencing is the cheapest noise rejection available: fine grained detail like insects and rain disappears while a person survives intact. And it makes a single pass, because the frames and their scores are written in the same order, so filtering never needs a second decode.</p><p>Reading a keyframe costs the same whether you look at it or not. The engineering was choosing what not to look at.</p><h3>Where this breaks</h3><p>Any method worth using should tell you where it fails, so here are the three that matter.</p><p><strong>The sampling grid is your keyframe interval.</strong> Mine is three seconds, so a person who enters and leaves frame faster than that can be missed entirely. On a doorway, where someone approaches, opens and passes through, dwell time runs four to ten seconds and the risk is low. On a corridor crossed in a second and a half it is unacceptable. Two fixes exist: shorten the camera’s keyframe interval, which costs perhaps ten to fifteen per cent more storage and is the right answer if you control the hardware, or decode every frame, which costs roughly ten times the compute.</p><p><strong>This detects motion, not people.</strong> A swinging door, a cat, a trolley and a delivery all register. The gates reduce noise, they do not classify. If you need classification, this is the correct thing to put <em>in front of</em> a model rather than a replacement for one. Run your detector over 7,238 candidate frames instead of 22 million and the economics change completely.</p><p><strong>My defaults were tuned on the easiest possible scene:</strong> a fixed indoor landing under constant artificial light with no windows. An outdoor camera needs a region of interest defined before anything else, cropping out sky, foliage and any public road, and that single change removes more false positives than any threshold adjustment.</p><p>One more, because it is the kind of error that hides. My first verification pass checked the merged files at container level, validating the index without decoding, and reported them clean. A later full decode surfaced <strong>166 bitstream errors</strong>: damaged data inside clips whose index was valid, so they had passed the health filter. Nothing was structurally broken, but “verified” at container level is a narrower claim than it sounds, and I had reported it as though it were not.</p><h3>The same shape at a different scale</h3><p>In the mid 2010s the consulting pitch was that one day video analytics would flag a till opening while the operator walked out of shot. The capability arrived. What did not, for most operators, is a way to review what the cameras already captured without committing a person’s week to it.</p><p>The shape recurs everywhere footage accumulates faster than anyone can watch it. A warehouse aisle where stock walks and nobody knows when. A loading bay with ninety days of retention and no index. A cinema counting free refills. A back door that opens at odd hours.</p><p>None of that needs a model to start. It needs a region of interest drawn tightly around the thing you care about, a persistence gate set for how long a real event lasts in your frame, and an illumination gate set for how much your light moves. Then it needs someone to look at the contact sheets and check the settings against a window where they already know what happened. A tuning you have not validated against known ground truth is a guess wearing a number.</p><p>The code that does this is public and free to take, at <a href="https://github.com/ATOIH/cctv-motion-triage">github.com/ATOIH/cctv-motion-triage</a>. It is about four hundred lines of shell with one dependency, and it runs on synthetic footage out of the box so you can see the behaviour before pointing it at anything of your own. The documentation is honest about where it fails, which I think is the more useful half.</p><p>My son got his gullak back. The method that found it was not clever. It was triage: refuse to look at everything, define what earns attention, and accept a stated risk in exchange for a volume a human can actually review. Every camera you own is already recording. The question is only ever which few minutes you are willing to look at.</p><p><em>Aayush Gupta writes about risk, controls and Trust &amp; Safety at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a><em>. He has built trust scoring, payments risk engines and fraud controls for lending and payments businesses.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b37ae4ee7342" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hope to Be Wrong. Prepare to Be Right.]]></title>
            <link>https://medium.com/@ca.aayushgupta/hope-to-be-wrong-prepare-to-be-right-9b1511c0df69?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b1511c0df69</guid>
            <category><![CDATA[personal-growth]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[resilience]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 21:31:38 GMT</pubDate>
            <atom:updated>2026-07-22T17:30:23.544Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OMyPKrsAYb44G-eNHeU-xw.jpeg" /></figure><h3>Somewhere in every company sits a person whose trophy cabinet holds only the disasters that never arrived — who engineers the silence of quarters where nothing happens, swallows every “I told you so”, and knows a secret the rest of the building doesn’t: <em>the invisible win is the best dopamine in business</em>.</h3><p><em>By </em><a href="https://aayush-gupta.com/"><em>Aayush Gupta</em></a></p><p>Nothing happened last quarter.</p><p>No headline. No war room at two in the morning. No regulator’s letter opening with “it has come to our attention”. In most careers that sentence is filler, the part of the review everyone skips. In risk management, it is the trophy cabinet. Because that “nothing” was built, assessment by assessment, control by control, drill by drill, someone engineered all that silence.</p><p>This is the strangest scorecard in business, and once it gets under your skin, the most addictive. Most professions win in public — the launch, the deal, the quarter. <em>A risk manager’s finest work is a no-impact event</em>: the crisis that called ahead and was told, politely, that there was nothing here for it.</p><p>The catch, of course, is that you can’t frame silence. So the profession runs on a different fuel and that fuel is what this piece is about.</p><h3>Resilience, wearing a lanyard</h3><p>Strip the frameworks away and risk management is one thing: resilience, practised professionally. You are always preparing for disasters, which means you are truly ready only when one finally strikes. And you were <em>good </em>only if what you prepared actually worked when it did.</p><p>So the honest test of this career is not the audit rating or the committee minutes. It is the day one of your top five risks stops being a slide in a deck and becomes a Tuesday. On that day, there are exactly two versions of you.</p><p>The prepared version watches the thing you built take a hit and hold the fort. Systems fail over. Playbooks open. People know their first move without being told. And you get to feel, for about two seconds, entirely in private, the purest sentence in professional life: <em>I told you so.</em></p><p>The unprepared version hurts more and, done right, is worth almost as much. Something you didn’t see, or saw and undersold, gets through. You take the hit, you take the lesson, and you move on, sharper, with one more pattern in the pool. No self-flagellation, no theatre. In this craft, the learning <em>is</em> part of the compensation. The only unforgivable outcome is the third one: neither prepared nor taught.</p><p>Prepared, you collect a told-you-so. Unprepared, you collect a lesson. Either way, you compound.</p><h3>The pulse that doesn’t show</h3><p>Those test days build something else, quietly: a temperament. Triage wants a steady hand; escalation wants a flat voice; a crisis wants at least one person in the room whose pulse doesn’t show. Being unshakeable at work is a superpower, and like every superpower it charges tuition. The steadiness follows you home, and the people who love you will sometimes read expressionless as emotionless — calm where there is, in fact, deep feeling.</p><p>Because the best in every profession are emotional about the work; this one just demands a demeanour, and nerves of steel to hold it. Two settings. One face.</p><h3>A scorecard you keep with yourself</h3><p>The other habit worth starting young costs nothing: keeping a journal of your told-you-sos; the two said out loud and the twenty swallowed.</p><p>Swallowed, because the moments rarely deserve the sentence. Mid-incident, commentary helps no one; the incident needs handling, not narration. And on the rare day it must be said, the craft lies in saying it without sounding like a prick — usually by pointing at the fix, not the forecast.</p><p>Write them down anyway, all of them, said and swallowed. A few years of honest entries become something no appraisal cycle will ever hand you: proof of your own judgment getting better. A profile page tracks your promotions; this tracks whether you were right. It gives you a strike rate, and read it as a rate, never a count.</p><p>A count flatters the person who makes one safe call a year and slanders the one who makes fifty hard ones.</p><p>Played straight, this is one of the most satisfying games in professional life: a scoreboard nobody can inflate for you, and nobody can take away.</p><h3>The ladder above being right</h3><p>For a long time I assumed a high strike rate was the whole game. Then one of the most self-assured, possibly brilliant men I’ve crossed paths with compressed the real hierarchy into two lines.</p><p>A very good risk manager has a high strike rate.</p><p>An outstanding one has a high strike rate and a low regret rate (which only a few people ever bother to track).</p><p>Of all the times you were right, how many still ended as the organisation’s “we should have listened”? Being right and being useful, it turns out, are different professions. The outstanding ones convert, just like the great marketeers or product folks: they get the fix done, or do it themselves — irrespective of their role, the span of that role, the roadmap, or what the annual internal audit plan says is in scope.</p><p>Conversion has a mechanism, and it is the least glamorous word in this piece: hands-on. Strip away the title and this job is one act — you recommend, somebody else decides. Whether they do as advised, ignore you, or do the exact opposite rarely depends on the quality of the analysis. It depends on whether you have skin in the game. Taleb wrote the book; this profession lives it. You flagged it, you own it, you stay in the room. Advice from behind glass gets treated like glass, advice with skin attached gets treated like instruction.</p><p>And here sits the graduation moment, the point where the two-second wage stops mattering. Because eventually you understand what being proven right by a real disaster means: you, your colleagues, your families’ stability, sometimes your whole industry, in the soup together. Winning by catastrophe is a losing way to win. So the outstanding ones stop collecting told-you-sos and start collecting something better — said by the other side, on your behalf, in a far better tense: <em>“You told us so. We believed you. And now we are better off than anyone else.”</em></p><p>That sentence is worth more than “the two seconds”.</p><h3>Intuition is a supply chain</h3><p>Which leaves the question every young professional actually cares about: where does a strike rate come from? The real answer is the least professional-sounding word in the building: intuition. At work we dress it up as experience-based judgment, and the dressing is fair — it is pattern recognition running slightly ahead of your ability to explain it. The better question is where the patterns come from.</p><p>Someone handed me the answer in college, as a brutal little taxonomy: there are people who learn from their own mistakes, people who learn from other people’s mistakes, and people who don’t learn at all. The emphasis sits on both words, mistake <em>and</em> learning. The world is full of people rich in one and bankrupt in the other. The first two categories compound. The third stays busy.</p><p>This is why exposure is the most underrated habit in the profession. Intuition eats whatever you feed it, and it is not a fussy eater: a history chapter, a courtroom judgment, a founder’s post-mortem, the way Mumbai reroutes itself in the rain. Read wide of your industry. Travel wider than your comfort. Keep at least one interest with no obvious return on investment. Nothing you truly absorb is wasted; it drops into the pool as one more pattern and waits, sometimes for years, for its moment.</p><p>Because risks never announce their domain. A loan collections mess, behaves like a queueing problem. A fraud ring grows exactly like a referral programme. A culture decaying quietly presents like deferred maintenance. Stay read-up only on your own industry and you will meet each of these as a stranger; build the wider pool and you have met them all before — somewhere else, wearing different clothes.</p><p>Run the loop for years and the pool turns deep enough that your calls come lightning-quick, on subjects nowhere near your lane, quick enough that people won’t quite believe it at first. Let them. The real jewelers, the few who can tell a stone from glass, will spot the trait in one meeting and cherish your company for life.</p><h3>A better game</h3><p>Put the pieces together and you can see what this career has been building all along. Risk forces preparedness. Preparedness, tested and retested, matures into resilience. And resilience, carried long enough, stops being a competency on a review form and becomes character. Beneath it runs the quieter loop, learning into experience, experience into adaptability, adaptability into the settled conviction that everything is connected: the risks, the systems, the people running both.</p><p>You can spot the people in whom the loop has been running for decades. They have an odd relationship with applause. Somewhere along the way the dopamine changed address: great work, done quietly, started outbidding every cheaper thrill. Reels bore them. Validation doesn’t move them. Status reads as what it always was — a mythic tradition someone else needs them to believe in. They aren’t performing humility. They’ve simply found a better game, and they’re still keeping score privately, as a rate.</p><p>That is the arc, and it is a brighter one than the job descriptions suggest: years spent hoping to be wrong, preparing as if right, and building teams and systems that hold either way. The work is invisible exactly when it is perfect and, the people doing it came to prefer it that way.</p><p>Nothing happened last quarter. Someone engineered that silence, added one quiet line to a journal, and slept well.</p><p><em>Aayush Gupta works in risk, controls, compliance, and Trust &amp; Safety for high-growth, digitally native companies across gaming, payments, and digital lending. More of his writing and case studies at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b1511c0df69" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context, Not Control: When Audit Dropped Its Defences]]></title>
            <link>https://medium.com/@ca.aayushgupta/context-not-control-when-audit-dropped-its-defences-8a0a7d87be34?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a0a7d87be34</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[internal-audit]]></category>
            <category><![CDATA[adaptability]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[trust-and-safety]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 14:55:57 GMT</pubDate>
            <atom:updated>2026-06-16T15:01:23.035Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cQB1nyoOqJqEG9G_R13AfA.jpeg" /></figure><h3>In 2020, The Institute of Internal Auditors quietly deleted a single word: Defence. That edit is the whole story of how the people who seemed as a cost centre become the ones who own trust, growth, and yes, the funnel in digital native companies.</h3><p><em>By </em><a href="https://aayush-gupta.com/"><em>Aayush Gupta</em></a></p><p>The finding was immaculate.</p><p>Every exception was evidenced. Every control gap was mapped to a clause, every clause to a risk, every risk to an owner. It was the kind of work that takes a sharp mind three weeks to build and a lifetime to learn how to build. By the time it landed, the product had shipped twice more. The gap it flagged had already been patched by an engineer who never read the report, or worse, the feature it described no longer existed.</p><p>The auditor did everything right. The operating model let them down.</p><p>This is the quiet tragedy of the assurance function in a fast company: some of the best systems-thinkers in the building, asked to comment on the game after the final whistle, and then quietly filed under “cost centre.” This is the the story of how that changes, not because audit failed, but because audit is growing up. And the new name it grows into is Trust &amp; Safety.</p><h3>A noble origin story</h3><p>Start with respect, because the lineage deserves it.</p><p>Internal controls and internal audit exist for a reason that has never stopped being true: to protect value, and to give the people at the top — the CEO, the board, the shareholder, an honest, independent read on whether the company is actually doing what it says it does. The discipline of evidence over opinion, of independence over convenience, of “show me” over “trust me,” is one of the genuinely great inventions of modern management. It is also unglamorous, which is precisely why it matters.</p><p>For decades the architecture held: the first line owns the risk, the second line (internal control, compliance) sets the guardrails, the third line (internal audit) checks the whole thing independently. The Three Lines. It was built to safeguard a company against itself, and it did.</p><p>Then the clock sped up.</p><h3>When the SOP can’t keep up</h3><p>A high-growth, digitally native company does not move at the speed of an audit cycle. It moves at the speed of a deploy.</p><p>By the time you’ve written the SOP, established the process, sampled it, tested it, and drafted the finding, the process has already changed — sometimes three times. Post-facto assurance, the very thing that made the old model trustworthy, becomes the thing that makes it late. And “late” in a digital business isn’t a scheduling problem; it’s a relevance problem.</p><p>There’s a deeper issue underneath the speed: interconnectedness. In a platform business, a routing tweak becomes a liquidity event becomes a conduct problem becomes a regulatory one, and it does all of that in a weekend. The classic audit response to interconnectedness is to bury it in caveats: <em>subject to the scope of our review, during the period examined, based on the samples selected.</em> Honest caveats. Also the sound of a function quietly stepping back from the very interconnections that now decide whether the company lives or dies.</p><p>None of this is the auditor’s fault. The people are excellent. The <em>model</em> hit a ceiling.</p><h3>The tell: a profession edits itself</h3><p>Here is the detail I love, because the profession diagnosed itself before anyone else did.</p><p>For years the framework was called the <em>Three Lines of Defence</em>. In July 2020, the Institute of Internal Auditors retired that name. It is now simply the <em>Three Lines Model</em>. The word that got deleted was <strong>“defence.”</strong></p><p>The IIA’s own reasoning was that “defence” framed risk as something purely bad, something to be guarded against in an era when risk-based decisions are as much about seizing opportunity as avoiding harm. The role of management, the board, and internal audit, they said, is to <em>enhance</em> value, not only protect it. To create and to protect.</p><p>Read that again, because it’s the whole thesis in one edit. The most conservative standard-setter in the assurance world looked at its own flagship model and decided the posture of pure defence had run its course. The vocabulary evolved first. The operating models are still catching up.</p><h3>Context, not control</h3><p>What replaces defence?</p><p>Reed Hastings has the cleanest answer I’ve found, and it has nothing to do with audit. Netflix’s organising principle is <em>lead with context, not control.</em> Don’t govern people through approvals, gates, and rules; give them so much context — the strategy, the numbers, the customer reality, the assumptions, that they make great decisions on their own, in real time, without stopping to ask. Hastings adds the part that should be tattooed on every assurance professional’s wall: when someone does something dumb, don’t ask who to blame, ask <em>what context you failed to set.</em></p><p>That sentence is the bridge from the old world to the new one.</p><p>The legacy assurance model led with control: the gate, the checklist, the sign-off, the after-the-fact finding. It worked when the business moved slowly enough to stop at each gate. It cannot work when there are ten thousand decisions a day and the gate is a queue. So the function does something radical — it stops trying to control every decision and starts embedding context into every decision. The control doesn’t disappear; it changes state. It moves out of the audit file and into the product one-pager, the UX, the rules engine, the real-time signal. The best controls in a digital-native company stop feeling like controls at all. They feel like the product working properly.</p><p>And the moment control monitoring moves to the first line — into the product, owned alongside the people who build it, something happens that has never quite happened in the assurance world before: the assurance professional can no longer hide behind a caveat. You own the outcome <em>with</em> the business, or you don’t own it at all. That shift, from commentary to accountability, is the birth of Trust &amp; Safety.</p><h3>The cheapest growth lever nobody used to audit: trust</h3><p>Why did this discipline emerge now, and why under that name?</p><p>Because digital-native companies, for all their dashboards, optimise for two numbers — what it costs to acquire a customer (CAC), and what that customer is worth over time (LTV). And the cheapest, highest-leverage input to both turns out to be the softest-sounding one: Trust. Trust that I won’t be cheated. That my data won’t be quietly sold. That I won’t be defrauded on your watch. That I’ll be dealt with ethically and not be mis-sold the thing you were measured on selling me.</p><p>Trust is brutally asymmetric. It is hard to earn, easy to break, and almost impossible to rebuild. So a discipline grew up specifically to protect it: high-ownership, high-accountability, allergic to patchwork — and it pulled the best instincts of audit, controls, and risk management onto the first line to do it.</p><p>This isn’t a soft claim. PwC’s 2025 Trust &amp; Safety Outlook found that introducing a single new trust-and-safety feature made 72% of surveyed users more likely to engage and 68% more likely to consider a purchase. PwC’s own framing is that T&amp;S leaders are no longer just risk managers; they are “architects of user confidence, business continuity and long-term value.” Trust isn’t the cost of doing business. In a digital business, trust <em>is</em> the funnel.</p><h3>The heresy: a control that grows the funnel</h3><p>Which brings us to the belief that keeps the old function in its box: <em>controls don’t improve funnels. Controls are friction. Pick a side, growth or safety.</em></p><p>A study of interconnectedness dissolves that fallacy completely. The trick is to stop asking “does this control reduce risk?” and start asking “does this control reduce risk <em>and</em> improve the funnel?” — and then to refuse to ship the ones that only do the first. Let me give you the examples I’ve actually watched work.</p><p>A customer trust score at a lending business that validated genuine customers logically and systemically — and lifted conversion, because the honest customer got through faster while the fraudulent one didn’t. A rules engine in a payments business that paid genuine merchants out faster, reducing false positives in half (and corresponding complaints), which earned word of mouth — a control that ran the growth loop. A payment-success screen that displayed a live payment counter to defeat the fake-payment-screenshot apps conning small merchants, years before the soundbox made it fashionable. A simple “are you on a call with someone right now?” prompt placed in the payment flow in 2020 to interrupt the social-engineering frauds that were emptying accounts.</p><p>None of those is a checkbox. Each is a control that a customer would <em>thank</em> you for. That is the entire craft: finding the guardrail that the customer experiences as care.</p><h3>Shift left, and read the wiring</h3><p>Operationally, this has a shape, and PwC names it well: <em>shift left.</em> Engage trust and safety at the design stage, not the launch review ensuring safety designed in from the product one-pager, not retrofitted after the incident. Treat safety as a core business requirement rather than a compliance checkbox.</p><p>But shifting left only pays off if you can read the wiring underneath, and this is where the audit mind becomes a superpower rather than a relic. Consider three traps that look compliant and aren’t:</p><p>The app store that <em>requires</em> you to offer account deletion and yet permits retaining some data for fraud prevention and legal duty. The real question isn’t “do we have a delete button.” It’s: who decided that the user’s contacts, their referral graph, their stored payment details were ever needed for any of it and therefore must go?</p><p>A privacy law that says encrypt data at rest and in transit. Fine. Except data can be encrypted at the source and at the destination and still travel naked in between. “We use SSL” is comforting until you remember that every developer on the team knows the credentials, so what exactly is the control protecting, and from whom?</p><p>Reading those interconnections, separating the <em>appearance</em> of safety from the <em>fact</em> of it, is exactly what a great auditor has always done. The skill didn’t become obsolete. It found a bigger job.</p><h3>You don’t need a hoodie to do this</h3><p>So is Trust &amp; Safety only for tenured product-and-engineering insiders? No. But it does have non-negotiable ingredients, and tenure isn’t one of them.</p><p>It needs context, because context is what creates interconnectedness, drives impact, and lets you tell a harmless wobble from the early edge of a wave. It needs customer-centricity so genuine that you become the customer — you actually use the product and the competitors’, and you can benchmark every flow, every nudge, every control against the company’s risk appetite, over and over, until you feel the friction the user feels. And it needs high agency: the temperament that treats every company problem and every customer problem as your own, and solves it structurally and scalably instead of with another patch.</p><p>An internal auditor or a consultant who can do those things, who will own the outcome shoulder-to-shoulder with the product and process owners, rather than grading them afterwards — makes an outstanding Trust &amp; Safety professional. The independence, the skepticism, the controls literacy, the systems thinking: that’s not baggage to leave behind. That’s the raw material. What actually changes is posture — from defending to building, from caveats to ownership, from the third line to the first, from cost centre to the team that demonstrably moves the number.</p><h3>A promotion you have to accept</h3><p>Organisations evolve. Roles evolve. The functions that survive are not the ones that defend their old mandate hardest; they’re the ones that notice the environment has changed and choose, deliberately, to grow into it. Trust &amp; Safety isn’t an exotic import that arrived to replace audit, controls, and risk. It’s what those three become when a high-velocity, customer-owned business asks more of them than a post-facto opinion — and they decide to give it.</p><p>The auditors who make the move don’t lose their identity. They finally get to own the outcome they were once only ever allowed to comment on.</p><p>The defences came down in 2020. What’s been waiting on the other side, all along, was the front line.</p><p><em>Aayush Gupta works in Trust &amp; Safety along with being a risk, controls, and compliance officer for high-growth, digitally native companies across gaming, payments, and digital lending. More at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a0a7d87be34" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Wobble to Wave]]></title>
            <link>https://medium.com/@ca.aayushgupta/from-wobble-to-wave-5efefcd378e6?source=rss-dd01b0c52e97------2</link>
            <guid isPermaLink="false">https://medium.com/p/5efefcd378e6</guid>
            <category><![CDATA[interconnectedness]]></category>
            <category><![CDATA[enterprise-risk]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[operations-management]]></category>
            <dc:creator><![CDATA[Aayush Gupta]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 20:21:15 GMT</pubDate>
            <atom:updated>2026-06-11T17:56:27.869Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vA4UOEc8iHtn5z4BkFENzg.jpeg" /></figure><p>Your organs aren’t standalone systems — and neither are your risks. Why a rising alert queue, a 98%-healthy API, and a “winning” referral campaign were each an enterprise crisis in disguise.</p><p>By <a href="https://aayush-gupta.com/">Aayush Gupta</a> <em>risk and compliance head for B2C digital startups across digital gaming, fintech payments and fintech lending; previously a consultant with Big 4 advisory firms.</em></p><p>I have made expensive mistakes by trusting neat dashboards.</p><p>A good physician never reads one number in isolation. A liver enzyme runs high, and the question isn’t “what pill lowers it” — it’s what your diet, your drinking, and sometimes even your back pain are quietly saying about the same body. Our bodies are not standalone systems. Neither are organizations. So why do we keep building risk-management systems as if they were — strategy in one box, liquidity in another, regulation, reputation, and cyber each on a separate slide with a separate owner?</p><p>This piece is about reducing the rate at which that fiction hurts you, not eliminating incidents. Anyone promising zero is selling something, but detecting trouble earlier, naming who acts, and refusing to treat a small operational tremor as a small operational problem is at the core of understanding interconnectedness. One discipline does most of the work: <strong>if one KRI deviates, you should know what else can move, how fast, and who acts first.</strong></p><h3>Interconnected risk is a network effect problem where interactions matter more than isolated nodes</h3><p>The pattern is stubborn: most major crises are interconnected, not isolated. Failures compound across domains rather than queuing up politely to take turns. Digital platforms make it worse, because they amplify the <em>velocity</em> of risk. A signal that once took a quarter to travel now crosses your business in days.</p><p>So you have to study risk as a probability of systemic exposure, not as a catalogue of isolated incidents. In calm periods, the links between strategy, liquidity, regulation, reputation, and cyber look weak — almost theoretical. Under stress, the correlations jump and everything moves together. That is exactly when single-event assumptions fail you.</p><p>Take a mundane signal. The queue of cases waiting for the ops-risk team to review starts climbing. The reflexive response is a staffing answer — hire more reviewers, automate the alerts, point AI at it. But the queue was a symptom. Underneath sat limited oversight or annual review of the risk rules and thresholds feeding it. Treated as a manpower problem, the lengthening queue meant slower alert handling from more frauds slipping through, which meant a regulator’s attention.</p><p>And the early signal is usually already in your data. In most fintech fraud events, the first weak signals appear roughly <strong>30 days</strong> before the financial loss — buried in operational metrics. It’s why sharp business teams obsess over 30-day-past-due and day-30 retention. The wave is visible as a wobble, a month out, if you’re willing to read it.</p><blockquote><em>If one metric wiggles, ask what else can wobble.</em></blockquote><h3>The small inconveniences are the story</h3><p>If networks explain why risk travels, operations explain where it starts. ORM is where the weak signals surface first, long before anything reaches a loss line. Risk rarely arrives as one dramatic event; it arrives as ten small inconveniences. Queues lengthen, overrides spike, retries climb, handoffs slip. None of it dramatic. All of it diagnostic.</p><p>So track control health, not just losses. Losses are a lagging confession; control health is a leading tell — who bypassed which control, how often, and why. Treat the near-miss log as a predictive asset, because repeated “almost incidents” precede real ones. Put plainly: <strong>ORM signal quality determines ERM response quality.</strong> Weak operational signals guarantee a slow, panicked enterprise response; sharp ones buy you the rarest thing in a crisis — time.</p><p>Consider a real interconnection. A risk team celebrated a milestone: they were blocking a high volume of requests from sanctioned territories. Good news, surely. Except a critical API was running at 98% (well above the Critical Threshold) — and that quiet 2% was letting some of those very transactions through resulting in services being rendered in sanctioned territories with a default “Null Replacer” updating the region instead of country resulting in revenue being recognized in sanctioned countries for a SEC listed company, which meant a compliance and regulatory exposure nobody had on a dashboard. “API performance” had quietly become a question of business continuity and corporate existence — because no one was reviewing system thresholds at the steering-committee level. A 2% gap, read as a success.</p><h3>Read the change, not just the drop</h3><p>“Watch your weak signals” is useless without a method. Here is a repeatable one. Start with incident archaeology: replay every past incident and near-miss, and find the <em>earliest detectable anomaly</em> — not when it hurt, when it first showed. Convert that into a trigger statement: what changed, by how much, for how long. “Refunds felt high” is a feeling; “refund rate ran at twice baseline for three days” is a trigger. Pair internal triggers (exceptions, retries, overrides) with external ones (advisories, outages, sentiment), because the interesting failures live in the overlap. Validate the causal pathway with reverse stress testing before you set a threshold. And give every trigger an owner, a playbook, and an escalation clock — a trigger with no clock is just a rumour.</p><p>The subtle part: it’s about the <em>change</em> in a metric, not just a degrading one. A consumer business revamped its referral bonus and watched new-user onboarding surge. Marketing celebrated. But that same onboarding KRI was simultaneously flashing a failure in fraud-prevention controls — the spike wasn’t demand, it was abuse climbing through an open door. The wobble in marketing spend was, in fact, a building wave of liquidity risk. The number everyone wanted to see was the warning.</p><h3>A KRI without an owner is decoration</h3><p>This is the operational core, so I’ll be concrete. A key risk indicator is only useful as a complete chain: <strong>Trigger → Metric → Threshold → Owner → Action.</strong> Miss any link and it’s decoration. A dashboard without an action owner is just expensive wallpaper.</p><p>Watch one chain run across the whole enterprise. The trigger: resignations clustering among key-person dependencies in the DevOps team. The metric: unplanned attrition in critical-skill cohorts. The threshold: more than <strong>15% inside a 90-day window</strong>, owned by engineering leadership with a pre-agreed knowledge-transfer action. Left unaddressed, that “people” wobble became an enterprise wave — release cycles slowed by roughly <strong>40%</strong>, the team missed the market window for a core feature, and about <strong>12%</strong> of projected annual revenue was suddenly at risk. High attrition in IT was never an HR line item. It was knowledge risk, and then it was a sales miss. Leading indicators exist precisely to cut that decision latency, and with it the tail loss.</p><h3>When the wave breaks in minutes: Bithumb</h3><p>Some waves don’t take 30 days. On 6 February 2026, the South Korean exchange Bithumb ran a “random box” promotion. An employee meant to credit prizes denominated in won — about ₩620,000 in total — but entered the amounts in Bitcoin. The exchange airdropped roughly <strong>620,000 BTC</strong>, around <strong>$40 billion</strong>, into customer accounts. Bitcoin’s price fell about 17% as the error rippled out. Bithumb caught it within roughly half an hour and froze the affected accounts, recovering about <strong>99.7%</strong> — but not before close to <strong>1,800 BTC</strong> had been sold and were gone.</p><p>Read it as an interconnection, not a blooper. Low probability, high impact, very high velocity. The immediate ORM job was containment — a kill switch, exposure caps, frozen withdrawals. But the failure underneath was a tangle of risks usually managed in separate rooms: human capital (a single manual entry), liquidity (no cap on internal credits versus actual reserves), system controls (no input-validation hard stop, a 24-hour delay in balance updates), and incident response. There was no board-level KRI on financial-system integrity, and no intraday modelling of what an internal mis-posting could do to liquidity. The missing control wasn’t technical sophistication; it was the wiring between tech risk and financial risk.</p><p>The sequencing is the leadership lesson. From minute zero to hour two, the job is ORM containment — stop the bleed. From hour two onward, it becomes ERM consequence management — quantify the enterprise impact and steer it. Through both, the behaviour that separates handled from handled-badly is the same: communicate early, quantify your uncertainty honestly instead of faking precision, and update your assumptions in public as facts arrive.</p><h3>This isn’t hindsight theater</h3><p>It’s easy to draw straight lines backward, so let me be precise about the claim. The point isn’t that these endings were obvious in advance. It is narrower and more useful: in each case, early operational signals genuinely existed and were not converted into decisive governance action. The failure wasn’t a missing signal — it was a missing <em>conversion</em>, from signal to decision. That you can fix.</p><p>A risk manager’s real job, in fact, is to be the organization’s intelligence layer: to pair every risk with a response, because boards don’t want a risk list, they want risk-and-response together. That work lives on the wall, in how you build the dashboard. Stop reporting tiles and start reporting in three movements — <strong>Movement</strong> (what’s shifting), <strong>Causality</strong> (why), and <strong>Action status</strong> (who has it, and whether it’s moving). Link each ORM metric to exactly one ERM consequence pathway: override frequency, recurrence rate, and exception aging on one side; an appetite-breach index, liquidity stress, and regulatory heat on the other. Override frequency isn’t interesting alone — override frequency <em>as the leading edge of a conduct breach</em> is. And keep a challenge culture around the board itself, because the most dangerous dashboard is the one that’s reassuring every single week. If everything is green, you are either lucky or blind.</p><h3>Stay curious. Stay calm. Stay measurable.</h3><p>I keep theory close, but only where it sharpens a decision. Theory guides judgment; operations reveals truth.</p><p>If you take one thing from this, make it a habit rather than a memo. This week, pick a single volatile KRI. Map two plausible downstream ramifications — what <em>else</em> moves if this moves. Then pre-agree one action with one owner. That single exercise does more for risk maturity than any framework I can name.</p><p>We are not chasing zero incidents. We are chasing faster detection and better recovery — the discipline of reading the wobble while it is still just a wobble, across the organization and, if you are paying attention, across your own body too. Keep building context.</p><p><em>Aayush Gupta is a risk, controls and compliance leader with 15+ years institutionalising enterprise risk, product risk, AML/CFT and internal audit across gaming, payments and digital lending. He currently leads Risk, Controls &amp; Compliance at Flutter Entertainment (Junglee Games). More of his writing and case studies at </em><a href="https://aayush-gupta.com/"><em>aayush-gupta.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5efefcd378e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>