<?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 Amrita Sarkar on Medium]]></title>
        <description><![CDATA[Stories by Amrita Sarkar on Medium]]></description>
        <link>https://medium.com/@amritasarkar?source=rss-96212bbd166e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*TtYJOl0nnx7WmZV_AYDajg@2x.jpeg</url>
            <title>Stories by Amrita Sarkar on Medium</title>
            <link>https://medium.com/@amritasarkar?source=rss-96212bbd166e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 21:51:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@amritasarkar/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[How a Camera Company Became a Billion Dollar Business]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@amritasarkar/how-a-camera-company-became-a-billion-dollar-business-addb90e9d5f2?source=rss-96212bbd166e------2"><img src="https://cdn-images-1.medium.com/max/1199/1*_GhegS0gsQ8gVtI_h1JZPQ.png" width="1199"></a></p><p class="medium-feed-snippet">And Why Its Next Chapter Depends on Growing With Its Users</p><p class="medium-feed-link"><a href="https://medium.com/@amritasarkar/how-a-camera-company-became-a-billion-dollar-business-addb90e9d5f2?source=rss-96212bbd166e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@amritasarkar/how-a-camera-company-became-a-billion-dollar-business-addb90e9d5f2?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/addb90e9d5f2</guid>
            <category><![CDATA[snapchat-stories]]></category>
            <category><![CDATA[commercialization]]></category>
            <category><![CDATA[snapchat]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[product-marketing]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 06:28:59 GMT</pubDate>
            <atom:updated>2026-10-14T06:28:59.410Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[From Payment Rails to Payment Promises]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@amritasarkar/from-payment-rails-to-payment-promises-3823a8f3ab91?source=rss-96212bbd166e------2"><img src="https://cdn-images-1.medium.com/max/1920/1*deYhXVfz4TRzt2YIdMyMTQ.jpeg" width="1920"></a></p><p class="medium-feed-snippet">Why the next generation of payment resilience may need to start with intent.</p><p class="medium-feed-link"><a href="https://medium.com/@amritasarkar/from-payment-rails-to-payment-promises-3823a8f3ab91?source=rss-96212bbd166e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@amritasarkar/from-payment-rails-to-payment-promises-3823a8f3ab91?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3823a8f3ab91</guid>
            <category><![CDATA[payments]]></category>
            <category><![CDATA[payment-rails]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[singapore]]></category>
            <category><![CDATA[payment-gateway]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 07:32:25 GMT</pubDate>
            <atom:updated>2026-10-08T07:32:25.657Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Consent Architecture for Agentic Payments]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@amritasarkar/the-consent-architecture-for-agentic-payments-3c9e741d8b7a?source=rss-96212bbd166e------2"><img src="https://cdn-images-1.medium.com/max/1766/1*bF6Ulm8hFTHLOqTDZq7sxg.png" width="1766"></a></p><p class="medium-feed-snippet">Who Owns Delegated Financial Authority in Agentic Payments</p><p class="medium-feed-link"><a href="https://medium.com/@amritasarkar/the-consent-architecture-for-agentic-payments-3c9e741d8b7a?source=rss-96212bbd166e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@amritasarkar/the-consent-architecture-for-agentic-payments-3c9e741d8b7a?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c9e741d8b7a</guid>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 06:37:24 GMT</pubDate>
            <atom:updated>2026-07-09T06:37:24.323Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Product Scientist’s Rubric: When to Code the Rule and When to Train the Model]]></title>
            <link>https://medium.com/@amritasarkar/the-product-scientists-rubric-when-to-code-the-rule-and-when-to-train-the-model-084533c56856?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/084533c56856</guid>
            <category><![CDATA[probabilitist]]></category>
            <category><![CDATA[deterministic-systems]]></category>
            <category><![CDATA[regulated-ai]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Sat, 06 Jun 2026 13:22:26 GMT</pubDate>
            <atom:updated>2026-06-06T13:22:26.264Z</atom:updated>
            <content:encoded><![CDATA[<p>By The Product Scientist</p><p>There is a particular kind of meeting that happens in enterprise boardrooms right now. You have probably been in one.</p><p>A senior leader — sharp, genuinely curious, well-intentioned — stands up and says something like this:</p><p><em>We should use GenAI to automate compliance, payments, underwriting, customer support, legal review, clinical triage, and fraud detection.”</em></p><p>Everyone nods. It sounds modern. It sounds brave. And somewhere in the room, someone quietly types “AI strategy” into a slide template and calls it a vision.</p><p>But beneath the excitement is a confusion that will cost that company dearly — not in a dramatic, headlines-generating way, but in the slow, expensive, auditable way that regulated enterprises eventually pay for bad architectural decisions.</p><p>The confusion is this: they are treating every problem as if it belongs to the same class.</p><p>It does not.</p><h3>The First Failure of AI Leadership: Mistaking Intelligence for Instruction</h3><p>Some enterprise problems are not reasoning problems. They are enforcement problems.</p><p>A payment message either conforms to ISO 20022 formatting, or it does not.</p><p>A user either gave consent for their health data to be processed, or they did not.</p><p>A jurisdiction either permits cross-border data movement, or it does not.</p><p>A customer either appears on an explicit sanctions blacklist, or they do not.</p><p>These are not places where you want a model to <em>infer</em>, <em>interpret</em>, or <em>creatively reason</em>. These are places where you need the system to be boring, deterministic, auditable, and correct.</p><p>The first mistake immature AI leaders make is assuming that intelligence is always superior to instruction. In enterprise systems, sometimes the most intelligent thing you can do is hard-code the boundary.</p><p>That is not a failure of imagination. That is architectural maturity.</p><h3>The Real Question Is Not “Can AI Solve This?”</h3><p>There is a better question. It is the question mature Product Scientists ask before a single architecture decision is made:</p><h4>What kind of failure can this system tolerate?</h4><p>That question changes everything — because it forces you to classify decisions by their failure severity before you select your tooling.</p><p>Deterministic rules are powerful precisely because they are predictable, auditable, and enforceable. But they are also rigid. They fail when the world becomes messy, ambiguous, and unstructured.</p><p>Probabilistic AI is powerful because it can interpret nuance, synthesize context, and handle noisy, heterogeneous inputs. But it is not inherently safe. It produces likelihoods, not guarantees.</p><p>So the real architecture question is not <em>rules versus AI</em>. It is: <strong>what is the correct relationship between certainty and flexibility in this specific decision context?</strong></p><h3>The Metaphor That Changes How You See the Problem</h3><p>Let me give you the image I return to when I am designing enterprise AI systems.</p><blockquote><strong><em>Deterministic rules are the containment vessel.</em></strong></blockquote><blockquote><strong><em>Probabilistic AI is the adaptive liquid inside it.</em></strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/0*5uqYoLogwxIDmsE_" /></figure><p>The vessel defines what cannot be breached: regulatory compliance, consent boundaries, payment validation, data residency, explicit blacklists, audit thresholds, escalation triggers.</p><p>The liquid fills the space inside the vessel: summarization, exception reconciliation, ambiguous case interpretation, clinical knowledge extraction, anomaly clustering, risk explanation, decision support.</p><blockquote><strong><em>The vessel does not change shape when the liquid moves. That is the point.</em></strong></blockquote><p>And here is the line I want you to carry out of this:</p><blockquote><strong><em>Rules decide what must never happen. AI helps decide what might be happening.</em></strong></blockquote><p>That is not a semantic distinction. It is an architectural one — and getting it wrong is how regulated enterprises end up with governance failures that no one intended and everyone should have predicted.</p><h3>Why “AI-Only Architecture” Breaks in Regulated Environments</h3><p>In consumer products, probabilistic behavior can often be tolerated. A recommendation system can be slightly wrong. A chatbot can misread a preference. A search ranking can be imperfect.</p><p>But in regulated systems, wrongness has categories.</p><p>Some errors are annoying. Some are expensive. Some are illegal. Some are existential.</p><blockquote><em>A hallucinated product recommendation is a UX issue.</em></blockquote><blockquote><em>A hallucinated legal citation is a trust issue.</em></blockquote><blockquote><em>A hallucinated compliance decision is a governance failure.</em></blockquote><blockquote><em>A hallucinated consent interpretation is a regulatory breach.</em></blockquote><blockquote><em>The enterprise AI leader who does not distinguish between these categories is not building AI strategy. They are accumulating liability.</em></blockquote><h3>The Product Scientist’s Selection Rubric</h3><p>This is the practical tool I use. Not theory — a working decision framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MI1gSjGWxe4Mc8PH" /></figure><p>Then distill it to three sentences:</p><blockquote><em>If the decision must be defensible in court, </em><strong><em>code the rule.</em></strong></blockquote><blockquote><em>If the decision must make sense of messy reality, </em><strong><em>train or prompt the model.</em></strong></blockquote><blockquote><em>If the decision must do both, </em><strong><em>build a hybrid — and be deliberate about where each layer ends.</em></strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/0*hbPog09b-_-kLHsv" /></figure><h3>What This Looks Like in Practice</h3><p>Abstract frameworks are useful. Real examples are what engineers and product leaders actually remember.</p><p><strong>Payments:</strong> A PayNow or ISO 20022 payment message should never be validated by an LLM. The schema, mandatory fields, rail-specific formatting, and account identifiers belong in deterministic logic. Full stop.</p><p>But when a payment <em>fails</em> because of an ambiguous beneficiary name, inconsistent remittance context, or unusual transaction behavior — that is where AI becomes genuinely valuable.</p><blockquote><em>The rule says: </em>“This payment cannot pass because field X is invalid.”</blockquote><blockquote><em>AI says: </em>“This failure likely occurred because the beneficiary name differs from three prior successful transactions by one token, and the remittance note suggests a recurring supplier relationship. Here is what a reviewer should examine.”</blockquote><p>The rule protects the rail. AI resolves the exception.</p><p><strong>Healthcare:</strong> Consent, data residency, clinical disclaimers, biomarker reference ranges, and escalation boundaries must be deterministic. If a user has not consented to process their partner’s data, the system cannot use partner-level inputs. That is not a model decision. That is a hard boundary — and it must be enforced upstream, not left to a model’s judgment.</p><p>But AI can meaningfully synthesize complex clinical inputs: cycle history, semen analysis notes, lifestyle data, biomarker trends, stress patterns, and patient narratives. It can surface patterns that no rigid schema would catch.</p><blockquote><em>The rule says: </em>“Do not process this data without consent.”</blockquote><blockquote><em>AI says: </em>“Given the available consented data, these are the reproductive-readiness patterns worth discussing with a clinician.”</blockquote><p>The rule protects the patient. AI improves the interpretation layer.</p><p><strong>Legal and knowledge systems:</strong> A legal AI product should never invent whether a source is valid, whether a jurisdiction applies, or whether a citation exists. Source traceability, document permissions, access rights, and citation validation need deterministic controls. These are not candidates for probabilistic reasoning.</p><p>But AI is genuinely powerful for summarizing case law, extracting arguments, comparing contract clauses, or flagging possible contradictions across document sets.</p><blockquote><em>The rule says: </em>“Only retrieve documents from approved sources the user has access to.”</blockquote><blockquote><em>AI says: </em>“Across these retrieved documents, the strongest argument appears to be X, with supporting evidence from Y and Z.”</blockquote><p>The rule protects truth boundaries. AI accelerates reasoning.</p><h3>The Architecture Pattern That Actually Works</h3><p>A mature enterprise AI system is not a single model call. It is a layered decision system. Here is the pattern:</p><p><strong>Layer 1 — Input validation:</strong> Checks schemas, required fields, user permissions, consent status, jurisdiction, format integrity, and access rights. Nothing proceeds without passing this layer.</p><p><strong>Layer 2 — Deterministic policy:</strong> Enforces the non-negotiables. Regulatory rules, blacklists, data boundaries, approval gates, escalation thresholds. This is where the hard lines live.</p><p><strong>Layer 3 — Retrieval and context:</strong> Pulls only allowed, traceable, permissioned data into the reasoning environment. The model sees only what it is authorized to see.</p><p><strong>Layer 4 — Probabilistic AI:</strong> Performs synthesis, classification, anomaly detection, summarization, drafting, or recommendation. This is where intelligence operates — inside the vessel, not in spite of it.</p><p><strong>Layer 5 — Confidence and risk scoring:</strong> Evaluates uncertainty, missing evidence, contradictions, source quality, and decision severity before anything reaches the next layer.</p><p><strong>Layer 6 — Human-in-the-loop:</strong> Routes high-risk or low-confidence cases to a reviewer. The model flags; a human decides.</p><p><strong>Layer 7 — Audit and learning:</strong> Records which rule fired, what data was used, what the model generated, what the human approved, and how the system should improve over time.</p><p>This is the architecture that separates AI product leadership from AI theater.</p><p>The model should never be the system. The model should be one reasoning component inside a governed decision architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*shiYS3BbNrqQbfgX" /></figure><h3>What This Means for Leaders</h3><p>The AI maximalist sees a model and immediately looks for places to deploy it.</p><p>The Product Scientist sees a decision system and asks three questions first: Where must certainty be preserved? Where can judgment be assisted? Where must risk be contained?</p><p>These are not engineering questions. They are leadership questions. And they belong to product leaders, not just architects.</p><p><strong>The organizations that will win the next five years of enterprise AI are not the ones who replace the most rules engines with models. They are the ones who build the most thoughtful boundaries between where rules must live and where models earn their place.</strong></p><p>Deterministic architecture gives the enterprise its control — its compliance, its consent, its auditability, its defensibility.</p><p>Probabilistic AI gives the enterprise its adaptability — its interpretation speed, its exception intelligence, its ability to navigate the world as it actually is rather than as the rule-writers imagined it would be.</p><p>The Product Scientist’s job is to design the operating boundary between both. And to know — with precision, not instinct — which side of that boundary every decision belongs on.</p><p><strong>In mature systems, rules own the red lines. Models own the grey zones.</strong></p><p>That is not a philosophical preference. It is the design principle that keeps regulated enterprises both intelligent and safe.</p><p>The question is whether your AI architecture reflects it — or whether someone in your boardroom is still nodding at the word “GenAI” and calling it a strategy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=084533c56856" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Singapore’s Fertility Floor Is a Product Design Problem, Not a Healthcare Problem]]></title>
            <link>https://medium.com/@amritasarkar/singapores-fertility-floor-is-a-product-design-problem-not-a-healthcare-problem-3b9c98937e24?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b9c98937e24</guid>
            <category><![CDATA[reproductive-health]]></category>
            <category><![CDATA[ivf]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[singapore]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Thu, 04 Jun 2026 07:58:45 GMT</pubDate>
            <atom:updated>2026-06-04T07:58:45.714Z</atom:updated>
            <content:encoded><![CDATA[<p>By Amrita Sarkar- Founder Ovviia Inc.</p><p>Singapore’s resident fertility rate hit 0.87 in 2025. A decade ago it was 1.24. In Parliament, DPM Gan Kim Yong put it in terms no spreadsheet softens: at this rate, for every 100 residents today, there will be 44 children and 19 grandchildren.</p><p>The reflex explanation arrives instantly. Housing is expensive. Careers are demanding. Work culture is intense. Childcare costs too much. People are choosing themselves.</p><p>None of that is wrong. It is also the most predictable conversation in the country — the one we have had, in some form, for thirty years. So let me propose a different one.</p><p>What if the problem is not only that couples don’t want children, but that the entire system gives them clarity too late?</p><p>Because here is what is strange about the 0.87 number. Singapore has not been passive. It has built one of the most generous family-formation toolkits in the world.</p><h3>We built a serious toolkit. It points the wrong way in time.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nKqw1tRdfdnAtOW4S6CJ3g.png" /></figure><p>Run down the list. Enhanced Baby Bonus cash gifts. Larger government contributions to Child Development Accounts. Shared Parental Leave, scaling to ten weeks. Doubled paternity leave. Childcare fee caps cut year on year. And the newest, headline measure: the Large Families Scheme, worth up to $16,000 in additional support for every third or subsequent child.</p><p>This is real money and real intent. But look closely at <em>where in the journey</em> it lands.</p><p>Almost every dollar is disbursed <strong>after the decision has already been made</strong> — and most of it after the child already exists. Baby Bonus rewards a birth. Parental leave supports a parent. The Large Families Scheme, the boldest recent move, is aimed at couples deciding whether to have a <em>third</em> child. In other words, the state’s freshest and largest incentives reward people who have already crossed the bridge into parenthood — twice.</p><p>Meanwhile, the couple stuck <em>before</em> their first child gets almost nothing. No clarity on their timeline. No read on their combined biology. No sense of what waiting two more years actually costs them. The support kicks in exactly one stage too late to change the decision that matters most.</p><p>You cannot subsidize your way out of a journey-design failure.</p><h3>The journey is designed like an emergency room, not a planning system</h3><p>Picture how a couple in Singapore actually experiences fertility today.</p><p>It starts as life uncertainty — <em>not yet, maybe in a couple of years.</em> That hardens into delay. Delay curdles into a vague anxiety neither partner says out loud. Eventually someone Googles AMH at midnight. They book a test here, a different test there, with no one to interpret the whole picture. The first serious, structured experience they get is a fertility clinic — by which point the system has already quietly reframed their situation as illness.</p><p>Life uncertainty → delay → anxiety → fragmented testing → clinic escalation → expensive, emotionally loaded late decisions.</p><p>At no point in that funnel does anyone hand the couple a simple answer to the only questions they actually have. The system is built to respond once you are already in trouble. That is not a healthcare gap. An emergency room is supposed to wait for emergencies. The gap is that <strong>there is no planning system upstream of it</strong> — and a planning system is a product, not a policy line item.</p><h3>Three institutions are looking at the same couple and seeing three different problems</h3><p>This is the heart of it.</p><p>The <strong>state</strong> looks at that couple and sees a birth-rate problem — a number to lift with incentives.</p><p>The <strong>clinics</strong> look at them and see an infertility problem — something to diagnose and treat once it presents.</p><p>But the couple themselves are not experiencing a birth-rate problem or an infertility problem. They are experiencing a <strong>clarity problem.</strong> They don’t know where they stand, how much time they realistically have, or what to do next.</p><p>And nobody owns the clarity problem. The state owns the endpoint. The clinics own the clinical endpoint. The messy middle — pre-marriage ambiguity, delayed timing, silent reproductive decline, couple disagreement, the fear of walking into a clinic and being made a patient — has no owner at all.</p><h3>Couples don’t need more data. They need decisions.</h3><p>Even when couples <em>do</em> gather information, the system hands them fragments, not answers. A lab report says AMH = X. A semen analysis says motility = Y. A wearable says sleep = Z. Three numbers, three apps, zero interpretation.</p><p>What a couple actually needs to know is short:</p><ul><li><strong>Where are we right now?</strong></li><li><strong>How much time do we realistically have?</strong></li><li><strong>What does it cost us if we wait?</strong></li><li><strong>What is the single best next step — change something, test something, consult someone, or just start?</strong></li></ul><p>Data without decision design does not reduce anxiety. It manufactures it.</p><p>And there is a structural blind spot inside even this. Most fertility products are built around one body — the woman’s cycle, her egg reserve, her age. But conception is a joint system. Male factors, lifestyle, metabolic health, and timing all sit inside the same equation, and they are largely invisible in how the journey is designed today. If your product models only one partner, you hand the couple a distorted map and call it guidance. The real decision-making unit was never the patient. It was the couple.</p><h3>The missing layer</h3><p>Singapore’s fertility stack has two strong floors and a hole between them.</p><p><strong>Layer 1 — Policy incentives:</strong> cash, leave, childcare, housing, large-family support. Heavily invested.</p><p><strong>Layer 2 — Clinical intervention:</strong> diagnostics, specialists, IVF, assisted reproduction. Strong capacity.</p><p><strong>The missing layer — Couple decision infrastructure:</strong> a trusted, early, couple-level clarity layer that tells two people where they stand, how much runway they have, what to improve in the next 90 days, and <em>when</em> — and only when — to escalate to a clinician.</p><p>This is not another period tracker. It is not an IVF marketplace. It is not another baby campaign. It is the decision layer that should sit <em>before</em> a couple ever needs the clinic, turning a high-uncertainty funnel into an informed one.</p><p>The redesigned journey looks like this:</p><p>Partnership → a couple-level readiness read → a clear biological timeline → a concrete near-term plan → evidence-based escalation only when warranted → an informed decision, made early enough to still be a decision.</p><h3>What this means for the people who can build it</h3><p>Singapore is the right place to prove this, and the timing is rare. The country has the policy will, the digital-health maturity, the clinical depth, and — at 0.87 — an undeniable mandate. The one thing it has not yet built is the layer that meets couples <em>before</em> crisis.</p><p>For founders, that is a category waiting to be created: reproductive decision infrastructure, not fertility software. For clinicians, it is a way to receive patients who arrive informed and on time, instead of late and frightened. For policymakers, it reframes the whole debate — the next frontier is not more support after birth, but more clarity before conception.</p><p>The state sees a birth-rate problem. The clinics see an infertility problem. Couples experience a clarity problem.</p><p>Until we design for the third one, the first two will keep spending against a number that refuses to move.</p><p>That missing layer is exactly what I’m building Ovviia around.</p><p><em>Sources: Total fertility rate figures — Singapore Department of Statistics (TFR 2025) and Parliamentary remarks by DPM Gan Kim Yong, Feb 2026. Family support measures — Ministry of Social and Family Development, Strategy Group (Large Families Scheme), and LifeSG.</em></p><h3>LinkedIn version</h3><p><em>(Tighter cut for the feed — first two lines are written to land above the “see more” fold.)</em></p><p>Singapore’s fertility rate just hit 0.87 — a record low.</p><p>The reflex is to blame cost, careers, and culture. I think we’re misdiagnosing the problem.</p><p>Singapore has not been passive. Enhanced Baby Bonus. Shared Parental Leave. A new Large Families Scheme worth up to $16,000 per third child.</p><p>But look at <em>where in the journey</em> that money lands. Almost all of it is disbursed after the decision is already made — and the biggest, newest measure rewards couples deciding on their <em>third</em> child. The couple stuck before their <em>first</em> one gets almost nothing: no read on their timeline, no sense of what waiting two years actually costs them.</p><p>You cannot subsidize your way out of a journey-design failure.</p><p>Here’s how a couple actually experiences fertility today:</p><p>Uncertainty → delay → quiet anxiety → a midnight Google of “AMH” → scattered tests no one interprets → a fertility clinic.</p><p>The first structured experience they get is a clinic — by which point the system has already reframed them as patients. It’s built like an emergency room: it waits until you’re in trouble, then offers clarity. There is no planning system upstream of it.</p><p>And that’s the real gap. Three institutions look at the same couple:</p><p>→ The <strong>state</strong> sees a birth-rate problem. → The <strong>clinics</strong> see an infertility problem. → The <strong>couple</strong> experiences a clarity problem.</p><p>Nobody owns the clarity problem.</p><p>Couples don’t need more data — they have AMH numbers, semen reports, sleep scores in three different apps. They need four answers: Where are we? How much time do we have? What does waiting cost? What’s the next best step?</p><p>That missing layer isn’t another period tracker or IVF marketplace. It’s couple-level reproductive decision infrastructure — and it should exist <em>before</em> anyone needs a clinic.</p><p>The next frontier in fertility isn’t more support after birth. It’s more clarity before conception.</p><p>It’s what I’m building at Ovviia.</p><p>What would change if couples got clarity at 28 instead of a diagnosis at 38?</p><p>#FemTech #ReproductiveHealth #ProductDesign #Singapore #HealthTech</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b9c98937e24" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Reduced STP Failures 60% on PayNow Corridors While Implementing MAS-Aligned AI Controls]]></title>
            <link>https://medium.com/@amritasarkar/how-i-reduced-stp-failures-60-on-paynow-corridors-while-implementing-mas-aligned-ai-controls-40d7b6d96048?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/40d7b6d96048</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[singapore]]></category>
            <category><![CDATA[ma]]></category>
            <category><![CDATA[upi]]></category>
            <category><![CDATA[paynow-upi]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Sun, 31 May 2026 07:52:01 GMT</pubDate>
            <atom:updated>2026-05-31T07:52:01.227Z</atom:updated>
            <content:encoded><![CDATA[<p>By the Product Scientists</p><p><em>Most teams treat instant-payment speed and AI governance as opposing forces. On a $6B-a-day corridor, I learned they were the same problem — and solving one was the only way to solve the other.</em></p><p>When you run real-time rails — Singapore’s PayNow, India’s UPI, Thailand’s PromptPay — straight-through processing is the entire economic model. A payment that clears untouched costs cents. One that drops out of STP into a manual exception queue costs dollars, hours, and a measurable amount of customer trust. Across the APAC corridor I led at HSBC, that exception tax was quietly eroding $6 billion in daily payment volume.</p><p>The reflex is to treat STP failures as a payments-engineering problem: patch the rails, the connectivity, the message formats. My hypothesis was different, and it’s the one that turned out to matter: <strong>most STP failures are not payment failures. They are information failures wearing a payments costume.</strong></p><p>So before re-architecting anything, I instrumented the corridor and decomposed every exception by root cause.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e0Hb2-4PAe8JoCcN58dkSw.png" /></figure><p>The breakdown reframed the whole program. The single largest bucket wasn’t the network or rail behaviour — it was screening false positives: legitimate payments halted because a beneficiary name fuzzily matched a sanctions list. The second was structured-data loss in the ISO 20022 / MT translation layer — truncated names, dropped purpose codes, malformed remittance fields. Genuine rail-level faults were the <em>smallest</em> slice.</p><p>That meant the highest-leverage intervention wasn’t in the payment rail at all. It was upstream — a pre-flight layer that inspected each message <em>before</em> it entered the corridor, predicted whether it would clear, repaired the structured fields ISO 20022 actually cares about, and ranked screening alerts by genuine risk instead of raw string similarity.</p><p>This is where most fintech stories end. It’s also where most regulators start asking questions.</p><p>Because in a MAS-supervised environment, you cannot drop an opaque model into the payment path. An AI that “usually” repairs a beneficiary field, or “probably” downgrades a sanctions alert, is not an efficiency — it’s an unbounded liability. The same model that cuts your exception rate can just as easily clear a payment that should have been stopped.</p><p>So governance wasn’t a compliance layer bolted on at the end. It <em>was</em> the architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W8xUDHI-w5uUKtBC8EofoQ.png" /></figure><p>I built every AI decision in the corridor to satisfy four requirements, mapped directly to MAS’s FEAT principles — Fairness, Ethics, Accountability, Transparency:</p><ul><li><strong>Reasoning traces.</strong> Every repair and every alert score emitted a structured, human-readable rationale: which field changed, why, and on what evidence. No silent edits.</li><li><strong>Human-in-the-loop thresholds.</strong> The model never auto-cleared anything above a defined risk band. It triaged; people decided the cases that mattered.</li><li><strong>Immutable audit trail.</strong> Every intervention was logged end-to-end, reproducible and replayable for an examiner — the same traceability discipline I’d later carry into agentic workflows at Thomson Reuters.</li><li><strong>Explainability over accuracy theatre.</strong> A slightly less aggressive model I could defend in front of a regulator beat a marginally better one I couldn’t.</li></ul><p>The governance layer did something counterintuitive: it made the AI <em>more</em> useful, not less. Because every decision was inspectable, we could safely raise automation thresholds over time instead of capping them out of fear.</p><p>The numbers followed the architecture.</p><p>On the best-performing corridors, STP failure rates fell by <strong>60%</strong>, with the program landing in a 40–60% band across rails. Compliance cost-per-alert dropped from <strong>$15.00 to under $0.50</strong> — roughly a 97% reduction — as the screening layer learned to suppress false positives without suppressing real risk. Transparency improved across the full $6B daily flow, which mattered as much to treasury and audit as the cost line did to finance.</p><p>But the metric I care about most isn’t on that list. It’s that the control layer stopped being treated as the brake and became the moat. The governance work compliance had initially seen as overhead is exactly what made the automation defensible enough to scale — and exactly what a MAS-aligned market will pay a premium for.</p><p>That’s the pattern I keep returning to as a product leader in regulated AI: <strong>governance is not the tax you pay for automation. It’s the feature that lets automation compound.</strong> Teams that treat trust as friction ship slower and stall at the first regulator conversation. Teams that treat trust as architecture earn the right to move fast — precisely because they can show their work.</p><p>On instant rails, the bank that can prove <em>why</em> a payment cleared will always out-scale the one that can only prove <em>that</em> it did.</p><blockquote><em>Amrita Sarkar leads AI, data and payments products across regulated APAC and GCC markets, with 13+ years across HSBC, Thomson Reuters and Deloitte. She writes as </em><strong><em>The Product Scientist</em></strong><em> on evidence-driven product thinking at the intersection of fintech, AI governance and trust.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40d7b6d96048" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Couple Score Problem: Why AI in Reproductive Health Needs a Different Compliance Architecture]]></title>
            <link>https://pub.towardsai.net/the-couple-score-problem-why-ai-in-reproductive-health-needs-a-different-compliance-architecture-e3e82c26e675?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3e82c26e675</guid>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Sat, 23 May 2026 15:01:01 GMT</pubDate>
            <atom:updated>2026-05-23T15:01:01.253Z</atom:updated>
            <content:encoded><![CDATA[<p>By the Product Scientist</p><p>There is a moment in product development when you realise that what you have built is not just technically novel — it is legally uncharted. For me, that moment came while designing the couple-level scoring methodology at the core of Ovviia.</p><p>Most health AI operates on a single subject. One patient. One risk score. One consent form. The compliance frameworks — PDPA, DPDP Act, HIPAA, the EU AI Act — were written with that model in mind. They assume one data subject, one controller relationship, one consent event.</p><p>A couple score breaks all three assumptions simultaneously.</p><p>When you compute a reproductive compatibility score across two individuals, you are ingesting health data from both, generating an output that neither party produced alone, and creating a record that belongs to both — and to neither. The moment you store that score, you have created a new data artefact with no clear owner and no established regulatory precedent.</p><p>This is not a hypothetical edge case. It is the central product surface of any platform serious about couples-level fertility intelligence. And getting the compliance architecture wrong here does not just create legal exposure. It destroys the clinical trust that reproductive health platforms live and die by.</p><h3>Why Reproductive Health AI Is Not Just “Sensitive Data With Extra Steps”</h3><p>The standard guidance for sensitive health data in AI systems runs something like this: encrypt it, minimise it, get consent, log access, and build a human review step for high-stakes outputs. All of that is correct. None of it is sufficient for a couple-level signal.</p><p>The problem is that most compliance frameworks treat data sensitivity as a property of the <em>input</em> — the menstrual cycle data, the hormone levels, the sperm motility scores. What they do not account for is output sensitivity: the fact that a derived inference can be more revealing, more consequential, and more contested than any of the raw inputs that produced it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4H4uEc6MFazoNkIsm3LRNQ.png" /></figure><p>A couple score is an output-sensitive artefact. It tells two people something about the combination of their bodies that neither of them could derive independently. It can inform decisions about when to try for a pregnancy, whether to pursue clinical investigation, and how to interpret years of unexplained outcomes. The score has weight — clinical, emotional, relational. And unlike a cholesterol reading, it has two subjects, two sets of expectations, and potentially two different interpretations of what the number means.</p><p>That asymmetry is where the compliance risk lives.</p><h3>The Three Structural Vulnerabilities</h3><h3>1. Consent Is Not Transferable</h3><p>In single-subject health AI, consent flows cleanly: one person authorises their data for a defined purpose, and the system respects that authorisation until it is withdrawn. The data subject and the data controller are in a direct, legible relationship.</p><p>In a couple score, Partner A’s consent does not authorise the use of Partner B’s data, even if their data is being combined to produce an output Partner A will see. This is not a technicality. It is the foundational principle of data sovereignty — each person’s health data belongs to them, and their partner’s consent does not extend it.</p><p>This means a couple score can only be lawfully generated when both parties have independently consented, under the same consent version, for the same stated purpose. If either consent is withdrawn, the score must not be regenerable — even if the other party’s data remains in the system.</p><p>Most platforms handle this badly, because they bolt consent onto registration flows designed for single users. The fix is architectural: consent state must be a precondition checked at inference time, not a checkbox captured at signup.</p><h3>2. Explainability Has Two Audiences</h3><p>When a single patient receives an AI-generated health insight, the explainability requirement is reasonably clear: the system should be able to say, in plain language, what data drove the output and why. Clinical AI governance frameworks — from Singapore’s MOH guidelines to the EU’s forthcoming AI Act medical device provisions — increasingly codify this as a minimum standard.</p><p>A couple score has two audiences for that explanation, and they will not always interpret it the same way.</p><p>If the score is low, each partner may want to understand their individual contribution. If the score changes between cycles, each may have different ideas about what changed and why. If one partner’s data is the dominant driver of a poor outcome, surfacing that in a shared context without appropriate framing is not just a UX problem — it is a potential harm vector. Blame architectures inside relationships are fragile and consequential.</p><p>This does not mean you hide the signal. It means your explainability layer has to be designed with the relational context in mind. Feature attribution should distinguish between couple-level signals (those that emerge from the combination) and individual signals (those that derive from one partner’s data). The plain-language summary should be written to inform both parties without creating a forensic accounting of whose body is “responsible” for a given score.</p><h3>3. The Audit Trail Has Two Subjects</h3><p>Standard audit trail design for health AI focuses on traceability: what data was used, which model ran, what output was produced. This is necessary but not sufficient when two data subjects are involved.</p><p>An audit log that is readable by one partner should not inadvertently expose the other partner’s raw inputs. A data access request from one party (increasingly a legal right under GDPR, DPDP, and PDPA) should not compel disclosure of the other party’s health data. A data deletion request from one party needs to be handled carefully — removing their inputs does not automatically invalidate the score, but regenerating the score without their data creates a different artefact with different legal standing.</p><p>These are not edge cases that a competent legal team can patch after launch. They are structural properties of the data model that have to be resolved at architecture stage.</p><h3>What a Compliant Couple Score Architecture Looks Like</h3><p>I have been building toward this with Ovviia, and the design I have settled on has four layers.</p><p><strong>Layer 1: Consent as a Runtime Precondition</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*220RGlbGgTDB0_Dm2QdlAw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RtFl5cX0gkyWO4oHQpacUQ.png" /></figure><p>Consent is not stored as a user attribute. It is stored as a versioned, timestamped record for each partner, indexed to a specific consent scope (fertility planning, research use, clinical sharing — each scoped separately). At inference time, the scoring pipeline queries both consent records before processing any data. If either record is missing, expired, or mismatched to the current consent version, the inference does not run. The consent check is logged as the first event in the audit chain.</p><p>This design means consent withdrawal is immediate and enforceable. It also means consent version upgrades cannot silently apply to existing users — each partner must re-consent independently before the new version takes effect.</p><p><strong>Layer 2: Derived Features Only, at the Inference Boundary</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AQ_Dz8XN0UixCdkDPlsoqg.png" /></figure><p>Raw health data — cycle logs, hormonal markers, sperm parameters — never enters the scoring pipeline directly. At ingestion, raw data is transformed into derived features: cycle irregularity index, stress quartile, phenotype category. The scoring model operates exclusively on these derived features. Raw data is stored separately, encrypted, with its own access controls and retention policy.</p><p>This matters for two reasons. First, it limits the sensitivity of what sits inside the audit log — feature vectors are far less revealing than raw health data. Second, it creates a natural data minimisation boundary that regulators can audit clearly.</p><p><strong>Layer 3: SHAP Attribution as a First-Class Output</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eRDns9I8o2uKsryol90Z7g.png" /></figure><p>Every couple score is accompanied by a SHAP (SHapley Additive exPlanations) feature attribution vector, computed at inference time and stored alongside the score. This is not a post-hoc explanation — it is part of the output contract. The attribution distinguishes couple-level signals from individual signals, and it feeds both the technical audit record and the user-facing plain-language explanation.</p><p>The confidence interval is also a first-class output. A score without a confidence band is not a health insight — it is a number. The confidence band communicates the model’s uncertainty, flags when inference is running on sparse data, and triggers different UX states for high-uncertainty outputs (more prominent caveating, shorter recommendation horizon, explicit suggestion to revisit when more data is available).</p><p><strong>Layer 4: Guardrails as Infrastructure, Not Policy</strong></p><p>Clinical action thresholds, purpose limitation checks, data residency verification, model version pinning — these are not documented in a policy PDF. They are runtime checks that run before every output is delivered, logged in the audit trail, and observable by any compliance review. A guardrail that exists only in documentation has no enforcement value.</p><p>The threshold design is the hardest part. For Ovviia, I have been working toward a framework where scores above a defined ceiling trigger a “clinician review recommended” flag — not a block, but a visible signal that this output warrants professional interpretation before it drives clinical action. Where that ceiling sits is a clinical question, not a product question, which is why getting a clinical advisory relationship in place before this feature launches is non-negotiable.</p><h3>What This Means for the Field</h3><p>Ovviia is not the only company building couple-level fertility intelligence. The intersection of AI, reproductive health, and couples data is a growing product category — and it is one that regulators in Singapore, India, and the UAE are beginning to pay attention to, partly because of how badly the consumer fertility tracking industry has handled data historically.</p><p>The companies that will earn clinical partnerships and regulatory confidence in this space will not be the ones with the most sophisticated models. They will be the ones that can demonstrate, with working infrastructure rather than policy documents, that they have thought through the consent, explainability, audit, and guardrail requirements specific to couples-level health data.</p><p>That is a higher bar than most teams are building toward. It is also a significant moat for the teams that get there first.</p><p><em>Amrita Sarkar is a product leader with 13 years across regulated digital platforms in financial services and health technology. She is the founder of Ovviia, a reproductive health intelligence platform with a patent-pending couple-level scoring methodology. She writes at the intersection of AI governance, clinical product design, and the evidence-based approach to building in sensitive domains. Try out Ovviia at os.ovviia.com . I would love to hear your thoughts on the MVP</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3e82c26e675" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/the-couple-score-problem-why-ai-in-reproductive-health-needs-a-different-compliance-architecture-e3e82c26e675">The Couple Score Problem: Why AI in Reproductive Health Needs a Different Compliance Architecture</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Death of the Rule Engine: Why the Next Payments Moat is “Cost-Per-Decision” (CPD)]]></title>
            <link>https://medium.com/@amritasarkar/the-death-of-the-rule-engine-why-the-next-payments-moat-is-cost-per-decision-cpd-07f3eebe15d5?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/07f3eebe15d5</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[payments]]></category>
            <category><![CDATA[agentic-workflow]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Mon, 18 May 2026 10:08:44 GMT</pubDate>
            <atom:updated>2026-05-18T10:08:44.678Z</atom:updated>
            <content:encoded><![CDATA[<p>The Product Scientist Take on Payment Moat.</p><p>Every second, a silent war is fought at the checkout counter.</p><p>Imagine it is 2:14 AM. A user in Tokyo attempts to buy a $2,500 luxury leather bag from a Shopify merchant based in New York. The transaction is routed through a maze of payment APIs, risk databases, and banking rails.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fToqbAK6YFsqXyv2dyiunw.png" /><figcaption>Try the Product Scientist CPD Framework here — <a href="https://pay-optima.lovable.app/">https://pay-optima.lovable.app/</a></figcaption></figure><p>In a traditional payments stack, this high-stakes moment is governed by a hard-coded line of code written three quarters ago by a tired engineer:</p><blockquote>if (transaction.amount &gt; 500 &amp;&amp; user.isNewIP) {<br> trigger3DS();<br>}</blockquote><p>This line of code is not just brittle; it is an invisible tax. It is a blunt instrument attempting to solve a high-dimensional, dynamic problem. It is the reason why perfectly legitimate customers get insulted with false declines, and why sophisticated fraud rings slip through the cracks.</p><p>For the last decade, fintech companies competed on <strong>features</strong>: more local payment methods (APMs), multi-acquirer failovers, and faster settlement times. But today, the payment stack is commoditized. You can buy feature-parity off the shelf from Stripe, Adyen, or Primis in an afternoon.</p><p>The next payments moat isn’t what your system <em>can</em> do. It is how efficiently it <em>decides</em> what to do.</p><blockquote>We are entering the era of <strong>Cost-Per-Decision (CPD)</strong>.</blockquote><p>We are entering the era of <strong>Cost-Per-Decision (CPD)</strong>.</p><h3>The Physics of a Payment Decision</h3><p>To understand why this is a paradigm shift, we must look at a payment transaction not as a plumbing problem, but as an optimization problem. Every decision to approve, decline, route, or step-up a transaction has an associated cost.</p><p>As a Product Scientist, we can formalize the <strong>Cost-Per-Decision (CPD)</strong> as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zp17b8X1cBD80fgLwZ0mxQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VcDyaYbxaE1Rl0qJfg0dog.png" /></figure><h3>From Brittle Rules to Agentic Loops</h3><p>Static rule engines operate on <strong>determinism</strong>. Agentic workflows operate on <strong>probabilistic optimization</strong>.</p><p>When you replace static rules with an autonomous payment agent, the linear payment path transforms into a continuous, self-correcting loop:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/660/1*HBt1Eh1axYQruhDD8B2zqQ.png" /></figure><h3>1. Context Gathering (Beyond the Transaction)</h3><p>An agent doesn’t just look at the metadata inside the JSON payload. It reaches out to gather real-time macro-context:</p><ul><li>What is the current authorization rate of this specific bank identification number (BIN) range on Adyen vs. Stripe over the last 10 minutes?</li><li>Is there a localized internet outage or latency spike in the user’s region?</li><li>What are the latest velocity patterns of this device fingerprint across the wider network?</li></ul><h3>2. Micro-Simulation (Evaluating the Expected Value)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RxUfH8jXRG8lX1FD31Ui0A.png" /></figure><h3>3. Execution &amp; Observation (Continuous Learning)</h3><p>The agent executes the optimal route. If the bank declines the transaction with a code 05: Do Not Honor, the agent doesn’t just display an error to the user. It immediately ingests the failure, updates its internal probability matrices, and retries the transaction through an alternative network with adjusted parameters (e.g., down-stepping from 3DS to a non-3DS path if the bank’s 3DS server is lagging).</p><h3>Why Agentic Workflows Are an Unassailable Moat</h3><p>In software, a “moat” is anything that makes it progressively harder for a competitor to take your market share. Agentic payments create a compounding, self-defending moat through three distinct feedback loops:</p><h3>1. The Data Density Flywheel</h3><p>Every time the agent makes a decision, it acquires a high-fidelity data point.</p><p>Because the agent learns in <strong>milliseconds</strong> rather than quarterly engineering sprints, its ability to price risk and optimize routing compounds exponentially. A competitor using a legacy rule engine is playing chess turn-by-turn, while the agentic loop is playing at 10,000 simulations per second.</p><p>[More Volume] ──&gt; [More Decision Feedback] ──&gt; [Lower CPD] ──&gt; [Higher Margins] ──&gt; [Win More Volume]</p><h3>2. Micro-Routing Mastery</h3><p>Payment networks are incredibly noisy. A specific card issuer in Spain might temporarily change their fraud threshold on Wednesday afternoons. A human risk analyst will never spot this. A static rule engine cannot adapt to it.</p><p>An agentic workflow, however, detects the sudden dip in Spanish BIN authorizations and dynamically micro-routes those transactions through a local Spanish acquirer, recovering up to</p><p>in lost conversion overnight.</p><h3>3. Dynamic Margin Management</h3><p>During high-volume events (like Black Friday), the optimal balance between fraud prevention and transaction speed shifts. An agentic system understands that transaction throughput is the priority. It automatically loosens secondary risk checks on low-risk buckets to maximize conversion velocity, while dynamically tightening them on high-risk vectors. It manages risk as a dial, not a light switch.</p><h3>The Product Scientist’s Roadmap to CPD Optimization</h3><p>If you are a product leader or payments engineer looking to survive the transition from the “Feature Era” to the “Decision Era,” you cannot do it overnight. You need a structured migration path:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GiRqnjcizAT6wB5lt9ZsFg.png" /><figcaption>Check out the detailed Dashboard here -<a href="https://adaptive-vision-loops.lovable.app/">https://adaptive-vision-loops.lovable.app/</a></figcaption></figure><h3>Instrumentation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/546/1*LTJwci4z65yI1L_o9ZcDdg.png" /></figure><h3>Hybrid Orchestration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*aFmwXKt2wh1NVx25gIY0jw.png" /></figure><h3>Autonomous Loop</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/571/1*FVAIC4jhhVQUhHf4VkOe8w.png" /></figure><h3>The Ultimate Paradigm Shift</h3><p>The winners of the next fintech epoch won’t be the ones who support the most payment methods or sign the most gateway partnerships.</p><p>The winners will be the ones who build the most efficient decision loops.</p><p>In a world where margins are squeezed and friction is fatal, the lowest <strong>Cost-Per-Decision</strong> wins. The rule engine is dead. Long live the loop</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v3W-m-AFdZkGt31gvHfa7A.png" /></figure><p>If you liked the article and want to try out an interactive dashboard to on CDP try this one out — <a href="https://agentic-flow-hub-89.lovable.app/">https://agentic-flow-hub-89.lovable.app/</a> It give a direct metric comparion between static rules and Agentic workflows.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07f3eebe15d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Prompt Monkey to Engineering Director: The 1% Claude Code Playbook]]></title>
            <link>https://medium.com/@amritasarkar/from-prompt-monkey-to-engineering-director-the-1-claude-code-playbook-4dc08a9a02b0?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4dc08a9a02b0</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[engineering-director]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Thu, 14 May 2026 06:57:05 GMT</pubDate>
            <atom:updated>2026-05-14T06:57:05.602Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Most people use AI coding tools as a fast typist. Here’s how the top 1% use Claude Code as a parallel engineering org they direct</em></p><p>I’ve spent the last several weeks obsessing over how the best Claude Code users work. I pulled apart Anthropic’s own internal power-user guide — the one written by engineers on the Claude Code team for themselves — and cross-referenced it with patterns from the community.</p><p>The conclusion is uncomfortable: <strong>most people are using Claude Code at maybe 10% of its leverage.</strong></p><p>This isn’t about knowing more commands. It’s a fundamentally different mental model. And once you see it, you can’t unsee it.</p><h3><strong>Layer 1: The Foundation — CLAUDE.md Is Your Most Compounding Asset</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oZiXiwp_0If6Z2VwDMHN0A.png" /></figure><p>Before any commands, flags, or tricks: <strong>invest in your CLAUDE.md.</strong></p><p>This is a markdown file at your repo root, checked into git. Every time Claude does something wrong, you add a rule to it. You end every correction with: <em>“Update your CLAUDE.md so you don’t make that mistake again.”</em></p><p>Claude is remarkably good at writing rules for itself.</p><p>The compounding math: a developer 3 months in with a maintained CLAUDE.md is operating in a fundamentally different environment than someone who started last week. The tool has been trained — on their specific codebase, their conventions, their mistakes — in a way that is not transferable.</p><blockquote>This is what the Claude Code team calls <strong>Compounding Engineering.</strong> They’ve even built it into PR workflows — install the GitHub Action, tag @claude in PR comments, and learnings get added to CLAUDE.md automatically during code review.</blockquote><p>Every correction makes every future session better. Stack these corrections for 90 days and you have something that looks like institutional knowledge.</p><p><strong>Paired with this:</strong> run /memory to configure auto-memory and auto-dream. Auto-dream is a subagent that periodically reviews your accumulated memory, removes outdated assumptions, merges overlapping notes, and keeps only high-signal entries. The naming is deliberate — it works like REM sleep consolidating short-term memory into long-term storage.</p><h3><strong>Layer 2: The #1 Leverage Point — Verification</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TOCWce0uQk77m9XSdnnPWQ.png" /></figure><p>This is the single most impactful practice in the entire guide, according to Anthropic’s own team. I’ll quote the principle directly: <strong>give Claude a way to check its own output.</strong></p><p>If Claude can close the feedback loop on its own, it will iterate until the output is right. If it can’t, it will hand you something plausible-looking and wait.</p><p>In practice:</p><blockquote><strong>For frontend work:</strong> Install the Claude Code Chrome extension. The framing from the team is blunt — if you ask a human engineer to build a website but don’t let them use a browser, will it look good? Probably not. With a browser, they’ll iterate until it does. Same logic applies.</blockquote><blockquote><strong>For code quality:</strong> After any non-trivial change, run /simplify. This fires parallel agents that review your changed code across four dimensions simultaneously: reuse, quality, efficiency, and CLAUDE.md compliance. One command, four reviewers, runs in parallel.</blockquote><blockquote><strong>For logic:</strong> Give Claude a test suite it can execute. No test suite means Claude is flying blind on whether its changes actually work. Invest in runnable tests and Claude’s output quality changes dramatically.</blockquote><p>The pattern is the same across domains: create a feedback loop Claude can close without you. Then let it run.</p><h3><strong>Layer 3: Parallel Execution — The Actual Force Multiplier</strong></h3><p>Here’s where most people leave the most performance on the table.</p><p>The biggest productivity unlock in Claude Code is running <strong>3–5 sessions in parallel, each in its own git worktree.</strong> Run claude — worktree to start a session in an isolated environment. Add — tmux to launch each in its own Tmux pane.</p><p>This isn’t multitasking. This is task decomposition — breaking a project into independent work streams and running them simultaneously. While one session is refactoring authentication, another is writing tests, and a third is updating documentation. You’re reviewing outputs, not waiting on them.</p><p>For large-scale migrations, the /batch command takes this further. It interviews you about the migration scope, then fans the work out to as many worktree agents as needed — potentially dozens. Each agent works in isolation, tests its own changes, and creates a PR independently. You merge what passes.</p><p>For custom automation, you can define agents in .claude/agents/ with specific models, tool sets, and isolation modes. Example:</p><p># .claude/agents/worktree-worker.md</p><p>— -</p><p>name: worktree-worker</p><p>model: haiku</p><p>isolation: worktree</p><p>— -</p><p>Then prompt naturally: <em>“Migrate all sync IO to async. Batch the changes and launch 10 parallel agents with worktree isolation. Each agent should test its changes end to end, then put up a PR.”</em></p><p>That instruction, to a properly configured system, will actually execute.</p><h3><strong>Layer 4: Plan Before You Build</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tgg72hArxFpj11y2TSzDgg.png" /></figure><p>Press <strong>Shift+Tab</strong> to cycle into plan mode. The discipline here is counterintuitive: pour your effort into the plan so Claude can one-shot the implementation.</p><p>The typical flow: enter plan mode → refine the plan aggressively → switch to auto-accept edits → Claude executes.</p><p>Two patterns from the Claude Code team worth copying:</p><p><strong>Dual-Claude review.</strong> Have one Claude write the plan, spin up a second Claude to review it in the role of a staff engineer. You get adversarial review before a single line of code is written.</p><p><strong>The abort signal.</strong> The moment something goes sideways mid-execution, go back to plan mode and re-plan rather than trying to course-correct mid-stream. Attempting to fix a bad trajectory with more prompts is usually slower than starting the plan over with better constraints.</p><p>On model choice: the team uses Opus with thinking for everything, even though it’s larger and slower. Their logic — less steering required plus better tool use equals faster overall results. The time you save on fewer correction loops outweighs the latency cost.</p><h3><strong>Layer 5: Prompting Like a 1%er</strong></h3><p>The best practitioners don’t accept first solutions. Three prompts that separate top users:</p><p><strong>“Grill me on these changes and don’t make a PR until I pass your test.”</strong> Forces Claude to validate your understanding before anything ships. Useful when you’re working in a domain where you want to actually understand what was built.</p><p><strong>“Prove to me this works.”</strong> Have Claude diff behavior between main and your feature branch. Not “does this look right” — prove it.</p><p><strong>“Knowing everything you know now, scrap this and implement the elegant solution.”</strong> Use after a mediocre first attempt. Claude will often produce a significantly cleaner implementation when given permission to start over versus being asked to improve the existing one.</p><p>The other technique worth building into your workflow: /btw. While Claude is actively working on something, /btw lets you ask a quick question without interrupting its flow. It’s single-turn with no tool calls, but has full context of the conversation. Ask about the retry logic while the migration is running. Get the answer. Let Claude keep going.</p><h3><strong>Layer 6: Automation Infrastructure</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OsCrK6vcMYB0PwJpdcTvXA.png" /></figure><p>If you do something more than once a day, turn it into a skill or command. Drop a markdown file in .claude/commands/ and it becomes a slash command. Check these into git so your whole team benefits.</p><p>Ideas the Claude Code team actually runs:</p><ul><li>A /techdebt command that executes at the end of every session, scanning for duplicated code</li><li>A command that syncs 7 days of Slack, Google Drive, Asana, and GitHub into a single context dump before planning</li><li>A /loop that runs every 5 minutes to babysit PRs — auto-addressing review comments, rebasing, shepherding to merge</li></ul><p>For recurring cloud jobs that run even when your laptop is closed, use /schedule. Example: <em>“Schedule a daily job that looks at all PRs shipped since yesterday and updates our docs based on the changes. Use the Slack MCP to message #docs-update with the changes.”</em></p><p><strong>Hooks</strong> give you deterministic control at lifecycle events:</p><ul><li>PostToolUse → auto-format code after every write/edit (prevents CI failures)</li><li>Stop → run checks on long tasks</li><li>PermissionRequest → route risky operations to Slack or an Opus-powered review agent before they execute</li><li>PostCompact → re-inject critical instructions after context compression</li></ul><h3><strong>Layer 7: The Operational Stack — Quick Reference</strong></h3><p><strong>Area</strong></p><p><strong>Command / Pattern</strong></p><p>Parallel work</p><p>claude — worktree, /batch, isolation: worktree</p><p>Planning</p><p>Shift+Tab for plan mode, /effort max for hard problems</p><p>Memory</p><p>CLAUDE.md, /memory, /dream</p><p>Verification</p><p>Chrome extension, /simplify, test suites</p><p>Scheduling</p><p>/loop (local), /schedule (cloud)</p><p>SDK speed</p><p>— bare flag (~10x faster startup for non-interactive use)</p><p>Multi-repo</p><p>— add-dir</p><p>Sessions</p><p>— name “task-name”, /branch to fork, — teleport to move between devices</p><p><strong>The Compound Effect</strong></p><p>Here’s the thing about this system: it’s not a set of tricks. It’s an infrastructure that gets better with use.</p><p>The CLAUDE.md accumulates corrections. The memory system consolidates patterns. The agents get tuned to your specific workflows. The verification loops get more sophisticated.</p><p>A developer who builds this infrastructure for 90 days is operating in a different category than one who treats every session as stateless.</p><p>Most people using Claude Code today are treating every session as stateless.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O_7C9VRPWPEH5I7vTUKwCg.png" /></figure><p><strong>Where to start tomorrow:</strong></p><ol><li>Create a CLAUDE.md and commit to updating it every time Claude makes a mistake</li><li>Install the Chrome extension if you do any frontend work</li><li>Run one session in plan mode before writing code</li></ol><p>Those three changes alone put you ahead of roughly 80% of current Claude Code users. The rest is compounding.</p><p><em>This article is part of my ongoing exploration of AI-native workflows for product and engineering leaders. I write about building at the intersection of technical credibility and strategic leverage.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RCiZThIJB-qxuXBjd1lv4w.png" /></figure><p><strong><em>Follow me for more: The Product Scientist.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4dc08a9a02b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Margin Reckoning]]></title>
            <link>https://medium.com/@amritasarkar/the-margin-reckoning-fca5fc097eaa?source=rss-96212bbd166e------2</link>
            <guid isPermaLink="false">https://medium.com/p/fca5fc097eaa</guid>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[model-routing]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Amrita Sarkar]]></dc:creator>
            <pubDate>Sun, 03 May 2026 14:35:17 GMT</pubDate>
            <atom:updated>2026-05-03T14:35:17.920Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A Field Guide to Profitable AI Products</strong></p><p><em>Why the next decade of product leadership will be won on the cost-per-query line, not the feature backlog.</em></p><p>There is a particular silence that descends on a boardroom when a CFO asks a question the product team cannot answer. I have watched it play out more times than I would like to admit, and it usually involves some version of the same question: <em>“How much does each query actually cost us?”</em></p><blockquote>In 2023 and 2024, that question was easy to dodge. AI was new, the models were a marvel, and growth bought forgiveness. By late 2025, the dodge stopped working. Compute lines on AI-product P&amp;Ls began growing faster than ARR. Investors started asking pointed questions about gross margin. Audit committees began treating inference cost as a board-level risk. The question had teeth now.</blockquote><p>We are living through a structural break in how software economics work, and most product leaders have not adjusted their operating system. This piece is for the ones who want to.</p><h3><strong>The shift no one warned us about</strong></h3><p>The SaaS playbook we all internalized had a cheerful assumption baked into it: marginal cost is approximately zero. Once you built the product, every additional user was almost pure margin. That assumption was not a footnote — it was the whole reason 80% gross margins became a normal benchmark. The cloud bill was a rounding error. The product was the price.</p><p>AI products broke that assumption.</p><blockquote>When a user types into your AI feature, you are not serving a static page. You are running an inference. That inference has a cost — sometimes pennies, sometimes dollars, occasionally more if context windows balloon. Multiply by power-user behavior, by free-tier abuse, by features the marketing team rolled out without consulting the platform team, and you have a COGS line that scales with usage. Not with infrastructure step-changes. With every single query.</blockquote><p>This is not a SaaS company anymore. This is closer to a utility, a logistics business, an airline. The marginal cost is real, and the discipline that real marginal cost demands is also real.</p><p>The leaders who will win the next product cycle are the ones who internalize this before their CFO does.</p><h3><strong>Why most teams are flying blind</strong></h3><p>Walk through enough AI product orgs and you notice a pattern: the people building features and the people paying the bills are usually two functions apart. Product looks at engagement and retention. Finance looks at AWS spend. Engineering looks at latency and uptime. Almost no one is looking at the line that actually matters — <em>cost per successful outcome.</em></p><blockquote>Cost per outcome is the right unit, not cost per query. A query that returns an unsatisfying answer is a cost without revenue. A query that triggers three clarifying follow-ups is four costs for one outcome. A free-tier user who burns fifty expensive queries before churning is a cost with no revenue attached at all.</blockquote><p>Most dashboards do not show this. Most product reviews do not surface it. Most quarterly planning ignores it. The scoreboard is wrong, and the team is playing the wrong game.</p><p>The first move in margin thinking is not a new tactic. It is a new dashboard.</p><h3><strong>The Five Levers</strong></h3><p>Once you start measuring cost per outcome, the levers reveal themselves. There are roughly five, and every AI product team should be able to articulate where it stands on each.</p><p><strong>Lever one: model routing.</strong> Frontier models are the Ferraris of inference — extraordinary, expensive, and absurd overkill for delivering pizza. A surprising fraction of real production queries can be answered by smaller, cheaper, faster models, often 70 to 90 percent cheaper, with no quality hit users can detect. The product question is not <em>which model is best</em>. It is <em>which model is sufficient</em>, and how to route confidently between them. Teams that build a routing layer early get a structural cost advantage that compounds for years.</p><p><strong>Lever two: caching.</strong> A meaningful share of production queries — 30 to 60 percent in many systems — are functional repeats of queries that came before. Semantic caching turns those into near-zero-cost lookups. This is the single highest-ROI engineering investment most AI teams have not yet made, and the math is rarely close.</p><p><strong>Lever three: context discipline.</strong> Every token in the prompt costs money on the way in and shapes what the model has to generate on the way out. Bloated system prompts, redundant instruction blocks, and lazy context stuffing are not just hygiene problems — they are a recurring tax on every call. Teams that treat context as a budget, not a buffer, run materially leaner.</p><p><strong>Lever four: pricing architecture.</strong> This is the lever product teams underuse the most. If your price is flat and your cost is variable, you have invited adverse selection: heavy users subsidize the rest, and over time, only heavy users stay. Consumption-based pricing, tiered fair-use thresholds, and explicit premium tiers for expensive workloads are not anti-customer. They are pro-business-model. The most generous thing you can do for your average user is make sure your service still exists in three years.</p><p><strong>Lever five: build versus buy.</strong> API providers offer speed-to-market, frontier capability, and zero ops burden. Self-hosted open models offer control, customization, and — at sufficient scale — meaningful unit-cost advantages. Most teams should buy until volume justifies building. Some should build sooner because of data residency, regulatory posture, or specific latency profiles. Almost no team should build first and figure out scale later. The build-versus-buy decision is one of the highest-stakes calls a product leader will make this decade, and it deserves to be made on a model, not on instinct.</p><p>These five levers are not exhaustive. They are the minimum viable vocabulary. A product leader who cannot speak fluently about routing, caching, context, pricing, and infra posture is not yet ready for the room where AI margin decisions get made.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xMKF1idHvw6kl45OkkqWqA.png" /><figcaption>Image Created with GPT</figcaption></figure><h3><strong>From feature factory to margin engineering</strong></h3><p>The harder shift is cultural.</p><p>For a decade, product management was optimized to ship — fast cycles, ruthless prioritization, the next feature, the next experiment. Some of that still applies. But layered onto it, there is a new discipline that does not yet have a clean name. Some call it FinOps for AI. Some call it cost-aware product design. I think of it as <em>margin engineering</em> — the deliberate, ongoing work of understanding what every product decision does to gross margin, and making the trade-offs explicit.</p><p>Margin engineering does not mean being cheap. It means being honest. The free-tier feature that delights users and burns four times its expected cost is not a winning experiment until you have a path to unit economics. The shiny agent workflow that calls the model thirty times in a chain is not a clever feature until you have benchmarked cost-per-outcome against value-per-outcome. The decision to default users to the most capable model is not a generosity until you have priced what that generosity costs.</p><p>The product manager of 2027 will read inference dashboards the way the product manager of 2017 read funnel dashboards. The discipline is the same. The dashboard is different.</p><h3><strong>What boards actually want to hear</strong></h3><p>Boards are not, in my experience, hostile to AI investment. They are hostile to AI investment that cannot articulate its own unit economics. The conversation they want to have is not <em>“are we spending too much on inference?”</em> It is <em>“do you understand what you are spending it on, and what it returns?”</em></p><p>The product leaders who answer that question with a structured, evidence-based view earn license to keep building. The ones who answer with platitudes about innovation get budget cuts and a quieter seat at the next review.</p><p>It is, in the end, the oldest game in business: know your numbers, defend your margins, and do not confuse motion with progress. AI did not change that. It just raised the stakes — and shortened the runway for product leaders who try to operate without the new instruments.</p><p>The margin reckoning is here. The good news is that the playbook is learnable, the levers are real, and the leaders who get there first will define the next decade of profitable AI.</p><blockquote>If you liked the article will be amazing if you can clap :) also excited to connect and chat more on how you are engineering margins in your organization.</blockquote><p>Till the next time- Amrita Sarkar (The Product Scientist)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fca5fc097eaa" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>