<?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 Tejas Amle on Medium]]></title>
        <description><![CDATA[Stories by Tejas Amle on Medium]]></description>
        <link>https://medium.com/@tejas_amle?source=rss-e26a2f50b232------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*QmkwqnsYFKXxdivC.jpg</url>
            <title>Stories by Tejas Amle on Medium</title>
            <link>https://medium.com/@tejas_amle?source=rss-e26a2f50b232------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 11:06:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tejas_amle/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[Your Audit Log Is a Diary. Customers Want a Receipt.]]></title>
            <link>https://medium.com/@tejas_amle/your-audit-log-is-a-diary-customers-want-a-receipt-fc55fc8bafa6?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc55fc8bafa6</guid>
            <category><![CDATA[logs]]></category>
            <category><![CDATA[seal]]></category>
            <category><![CDATA[tamper-evident]]></category>
            <category><![CDATA[evidence]]></category>
            <category><![CDATA[tamper-proof]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 09:33:43 GMT</pubDate>
            <atom:updated>2026-11-19T09:33:43.278Z</atom:updated>
            <content:encoded><![CDATA[<h3>DocuSign seals your PDF. AWS signs its own logs. Oracle chains database rows like a blockchain. Nobody coordinated. Here is why SaaS keeps rebuilding the same primitive, and what it actually buys you.</h3><p>Every SaaS product keeps an audit log. Very few of those logs would survive a serious argument.</p><p>The reason is structural. The log is written by the same system whose behaviour is being questioned. It is stored where that system’s admins can edit it. So when a customer, an auditor or a court asks you to prove something happened, the honest answer is “trust us, our database says so.”</p><p>That is a diary. Diaries are useful. They are not evidence.</p><p>So I went looking at who has already solved it. A surprising number of serious platforms have, each in its own corner and with its own vocabulary. DocuSign calls it a seal. AWS calls it integrity validation. Microsoft and Oracle call it a ledger. The open source supply chain world calls it a transparency log. Police evidence systems call it chain of custody.</p><p>Same idea every time. Make tampering detectable, then let someone other than you check.</p><h3>Three words that are not synonyms</h3><p>Vendors use immutable, tamper-proof and tamper-evident as if they were interchangeable. They are not, and the difference decides what you can honestly promise a customer.</p><p><strong>Immutable</strong> means the system refuses to change or delete the record. It is a lock. Amazon S3 Object Lock in compliance mode is the cleanest example. A locked object version cannot be overwritten or deleted by anyone, including the account’s root user, until retention ends. AWS’s own docs say the only early way out is deleting the AWS account. (<a href="https://docs.aws.amazon.com/AmazonS3/latest/dev/object-lock.html">AWS docs</a>)</p><p><strong>Tamper-evident</strong> means the record could still change, but any change is detectable. It is a seal. Hashes and signatures do this work. Break the seal and verification fails.</p><p><strong>Tamper-proof</strong> means tampering is impossible. Almost nobody serious claims it. Read the actual documentation of the platforms below and you will mostly find the word “evident”, not “proof”.</p><p>A lock stops the change. A seal proves it. Real products usually need both, because a seal over records that can be quietly deleted only proves the records you still happen to have.</p><h3>Six ways the industry already builds it</h3><p>I grouped what I found by mechanism, not by vendor. The mechanism decides the guarantee.</p><p><strong>1. Seal the artifact.</strong> Sign the finished object so any later edit breaks the signature. DocuSign seals completed documents with a PKI certificate. It also attaches a Certificate of Completion, itself digitally signed, carrying the signing history: names, emails, IP addresses, timestamps. (<a href="https://www.docusign.com/trust/security/transaction-data-use">DocuSign Trust Center</a>) C2PA Content Credentials apply the same idea to photos and video, so a camera or an AI image tool can sign where a file came from. (<a href="https://spec.c2pa.org/specifications/specifications/2.3/explainer/Explainer.html">C2PA explainer</a>)</p><p><strong>2. Chain the log.</strong> Hash every log file, then sign a digest that points back to the previous digest. AWS CloudTrail does this every hour with SHA-256 and RSA. Edit a file or drop a digest and verification fails. The part I like most: the algorithms are open and standard, so customers can write their own validator instead of trusting AWS’s. (<a href="https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-log-file-custom-validation.html">CloudTrail docs</a>)</p><p><strong>3. Lock the storage.</strong> Write once, read many. S3 Object Lock and Google Cloud Storage Bucket Lock both let you set a retention policy that, once locked, cannot be shortened or removed. (<a href="https://docs.cloud.google.com/storage/docs/bucket-lock">Google Cloud docs</a>) This is what financial firms reach for under records rules like SEC 17a-4. It proves nothing about content. It only makes deletion impossible.</p><p><strong>4. Ledger the database.</strong> Chain rows inside the database itself. Azure SQL’s ledger hashes changes into blocks and stores a database digest in immutable storage, so you can recompute and compare later. (<a href="https://learn.microsoft.com/en-us/azure/azure-sql/database/ledger-overview">Microsoft Learn</a>) Oracle ships immutable tables that even DBAs cannot change, and blockchain tables that chain each row’s hash to the one before. (<a href="https://blogs.oracle.com/database/post/how-to-achieve-crypto-secure-data-management">Oracle blog</a>) Amazon had a dedicated product here called QLDB. More on what happened to it below.</p><p><strong>5. Publish to a log nobody owns.</strong> Put a fingerprint somewhere public, append-only and independently auditable. Sigstore’s Rekor does this for software signing, and you can staple an inclusion proof to an artifact to show it was logged. (<a href="https://www.chainguard.dev/supply-chain-security-101/what-is-sigstore-rekor">Chainguard explainer</a>) Estonia built Guardtime’s KSI into its government infrastructure so anyone can verify a record’s integrity without holding a private key. (<a href="https://interoperable-europe.ec.europa.eu/node/731634">EU case study</a>)</p><p><strong>6. Track custody end to end.</strong> Hash at intake, verify at every handoff, keep a log nobody can edit. Axon’s evidence platform, used by police forces, fingerprints every file, checks integrity before and after upload, and keeps custody records that even account admins cannot change. (<a href="https://wrps.ca/node/153">Waterloo Regional Police</a>)</p><p>Notice the ladder hiding in that list. Patterns one to four are still run by the vendor. Pattern five takes the vendor out of the trust path entirely. That ladder turns out to be the whole product question.</p><h3>The person who needs this is not your user</h3><p>This took me the longest to see. Tamper evidence is never for the person using your product. It is for the person who doubts them.</p><p>Your user already trusts their own data. The value appears when a third party has to believe it: an auditor, a regulator, a judge, a customer’s customer, the other side of a dispute. So the feature’s real job is to win an argument your user will have with someone else, usually months later, often after the people involved have moved on.</p><p>Seen that way, the use cases line up quickly.</p><p><strong>Proving what someone agreed to.</strong> Signed contracts, consent records, terms of service acceptance, a price change a customer approved. The dispute is always “that is not the version I saw.” A seal over the exact artifact settles it. This is the core of what DocuSign sells.</p><p><strong>Surviving an audit across a period, not a moment.</strong> A SOC 2 Type II or ISO 27001 auditor wants to know a control operated all year, not that a screenshot looks fine today. Evidence that carries its own capture time and integrity proof answers that without the auditor taking your word for it.</p><p><strong>Meeting retention rules that name the mechanism.</strong> Some regulations do not just say keep records. Broker-dealer rules require records that cannot be rewritten. AWS had S3 Object Lock independently assessed against SEC 17a-4, FINRA and CFTC rules, and lets customers download that report to hand to their regulators. (<a href="https://aws.amazon.com/pt/blogs/storage/protecting-data-with-amazon-s3-object-lock/">AWS Storage Blog</a>) That report is a sales asset, not a footnote.</p><p><strong>Defending against your own admins.</strong> The threat model is not only attackers. It is the insider with legitimate credentials, or the vendor’s own operators. That is why S3 compliance mode locks out even the root user, and why Oracle built tables a DBA cannot edit.</p><p><strong>Settling disputes between companies.</strong> When two businesses disagree about what happened, whose log wins? Neither, unless one side can verify the other’s record without trusting it. An independently checkable record is a neutral witness.</p><p><strong>Proving where software came from.</strong> Who built this binary, from which source, on which pipeline. Customers now ask, and signed provenance in a public log answers it.</p><p><strong>Accounting for what an AI agent did.</strong> The newest use case, and the one I think matters most. I will come back to it.</p><p>Commercially, this rarely arrives as a feature request. It arrives as a lost deal in a regulated vertical, a security questionnaire line you cannot tick, or a dispute that cost more to argue than it was worth. [Your example: a deal, RFP line or dispute where you saw this.]</p><h3>The hash is not the product. Who can check it is.</h3><p>Every system above hashes something. Hashing is cheap and solved. What separates them is who can verify the result, and that is a product decision, not an engineering one.</p><p>I think of it as three rungs.</p><p><strong>The vendor attests.</strong> “Our records are immutable.” You get an audit log in the UI, maybe a read-only auditor role, maybe a frozen snapshot of evidence. Useful, and a lot of software sits here today. But the proof is still the vendor’s word.</p><p><strong>The customer can verify.</strong> The vendor hands over the digest and the method, and you check it yourself. CloudTrail is the reference design. AWS publishes its public keys and the exact fields it signs, and documents how to build your own validator. Azure SQL’s ledger gives you a stored procedure that recomputes every hash against the stored digests and fails loudly on a mismatch. (<a href="https://learn.microsoft.com/sql/relational-databases/security/ledger/ledger-database-verification">Microsoft Learn</a>)</p><p><strong>Anyone can verify.</strong> The proof is anchored somewhere neither party controls. Rekor and KSI live here. The vendor could vanish tomorrow and the proof would still check out.</p><p>Each rung costs more to build and is worth more to the person doubting you. That also makes it a natural packaging ladder. Attestation in the base plan, customer-side verification for regulated tiers, public anchoring for deals where trust is the entire sale.</p><p>The test I would bring to a spec review is one question. If our own company were the party under suspicion, could the customer still prove what happened? If the answer is no, you have a diary with better marketing.</p><h3>Where the guarantee quietly ends</h3><p>The best vendors are refreshingly honest about the limits. Their docs are where I learned most of these.</p><p><strong>A seal proves nothing changed. It does not prove the thing was true.</strong> Oracle says it plainly: someone with stolen credentials can still insert false rows into a blockchain table, perfectly chained. (<a href="https://blogs.oracle.com/database/post/how-to-achieve-crypto-secure-data-management">Oracle blog</a>) The C2PA explainer makes the same point about media. It verifies that provenance is intact, not that it is true. Garbage in, sealed garbage out. The trust boundary is the moment of capture, and that is where most of the product design should go.</p><p><strong>The proof is only as safe as the place the digest lives.</strong> Microsoft’s ledger docs warn you to check that a privileged user has not quietly moved digest storage somewhere unprotected. (<a href="https://learn.microsoft.com/sql/relational-databases/security/ledger/ledger-verify-database">Microsoft Learn</a>) If your admin can edit the data and the fingerprint, you have built an expensive diary.</p><p><strong>The proof has to outlive your product.</strong> Amazon QLDB was the canonical managed ledger, with cryptographic verification built in. AWS ended support on 31 July 2025 and pointed customers to Aurora PostgreSQL. (<a href="https://docs.aws.amazon.com/qldb/latest/developerguide/verification.verify.html">AWS docs</a>) Microsoft used the moment to argue that Aurora offers audit logging but not cryptographic verifiability. (<a href="https://techcommunity.microsoft.com/t5/azure-sql-blog/moving-from-amazon-quantum-ledger-database-qldb-to-ledger-in/ba-p/4246237">Azure SQL Blog</a>) That is a competitor’s pitch, but the gap it names is real. Retention periods run for years. Product roadmaps do not. Proofs should be exportable and verifiable without your servers.</p><p><strong>Do not let regulation hype write your spec.</strong> A lot of 2026 marketing claims EU AI Act Article 12 requires tamper-evident logs from August 2026. The Article asks high-risk systems to log events automatically for traceability. One careful reading points out it does not prescribe tamper-proof mechanisms at all. (<a href="https://originstamp.com/en/blog/reader/eu-ai-act-article-12-logging-requirements-defensible-logs">OriginStamp</a>) And the Digital Omnibus moved the Annex III high-risk obligations to 2 December 2027. (<a href="https://www.euaiact.com/digital-omnibus-ai">EU AI Act tracker</a>) Build this because your customers’ counterparties will ask for it. Not because a vendor blog said a law already does.</p><h3>Agents turn the diary problem into a liability problem</h3><p>In my <a href="https://medium.com/@tejas_amle/i-won-a-flash-sale-without-being-in-the-room-de2fa75d7a72">last piece</a> I wrote about winning a flash sale through an AI agent while I was somewhere else in the building. The detail I keep returning to is the verifiable receipt the agent got back. It established what happened. It did not establish who pays when the agent gets it wrong.</p><p>That gap is about to show up in every SaaS product that lets software act on someone’s behalf.</p><p>When a person clicks a button, there are many witnesses. The person, their screen, their inbox, their memory. When an agent acts, the only witness is often a log written by the same software that acted. If the user, the agent builder and the platform later disagree, each of them is holding a diary written by an interested party.</p><p>Agents also change the question being asked. With a human, an auditor asks what happened. With an agent, they ask what happened and under whose authority. So the record has to bind the action to the permission that allowed it, at the moment it happened, at machine speed, with no human reviewing each step. Nobody reconstructs that after the fact.</p><p>A useful agent receipt, as I see it, carries five things. The agent and the human or org it acted for. The grant or scope it was acting under. A fingerprint of what it saw and did. A trusted timestamp. A link to the previous receipt, so gaps are detectable. The pattern already has a name in the agent tooling world: signed, hash-chained action receipts. (<a href="https://agentpatterns.ai/security/cryptographic-governance-audit-trail/">AgentPatterns</a>)</p><p>My bet is that the agent platforms enterprises trust will be the ones whose every action is provable by someone who does not trust them.</p><h3>If I were writing the spec</h3><ol><li><strong>Name the verifier before the mechanism.</strong> Auditor, regulator, court, customer or the public. That answer picks the rung, and the rung picks the architecture.</li><li><strong>Seal at capture, not at export.</strong> Evidence sealed on its way out only proves the export is intact.</li><li><strong>Keep the fingerprint outside your own blast radius.</strong> A separate account, separate keys, or a log nobody owns. If one admin can touch both the data and the digest, start over.</li><li><strong>Pair a lock with a seal.</strong> WORM storage handles deletion. Hashes handle edits. Either one alone leaves an obvious hole.</li><li><strong>Make verification boring.</strong> One click for the auditor, open algorithms for the engineer, an offline verifier for the lawyer. If checking a proof needs a support ticket, nobody will check.</li><li><strong>Design for your own sunset.</strong> Proofs should export in a standard format and verify without your servers. QLDB customers learned why.</li><li><strong>Say tamper-evident.</strong> Claiming tamper-proof is the quickest way to lose a security reviewer’s trust.</li></ol><p>DocuSign never set out to build a cryptography product. It set out to make a signature something two strangers could both rely on. That is the real job. Not storing records, but making them believable to someone with every reason not to believe you.</p><p><em>Tejas Amle, a PM who now asks who wrote the log before reading it.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc55fc8bafa6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Every App Got an Agent. Payments Got a Waiting Room.]]></title>
            <link>https://medium.com/@tejas_amle/every-app-got-an-agent-payments-got-a-waiting-room-3a5cf435c3f1?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a5cf435c3f1</guid>
            <category><![CDATA[payments]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Sat, 26 Sep 2026 13:48:25 GMT</pubDate>
            <atom:updated>2026-09-26T13:48:25.670Z</atom:updated>
            <content:encoded><![CDATA[<h3>A PM’s map of agentic fintech in India after GFF 2026: why most of it is meta features around the payment, whether Indians will trust an agent with their money, and whether the way we pay ever changes.</h3><p>I have been wondering about something for months.</p><p>Every tool I use has quietly turned agentic. I design in Figma and Google Stitch by describing what I want, and I write code the same way. Food went the same way: Zomato shipped an MCP server, and Swiggy connected Food, Instamart and Dineout to ChatGPT, Claude and Gemini. I can build a full order inside a chat window.</p><p>Then, on Zomato’s MCP, the agent hands me a UPI QR code and waits for me to scan it.</p><p>That one screen is the state of agentic fintech. The agent can do everything except the part that involves money.</p><p>So I went looking for every place an agent has actually shown up in Indian fintech. The map split cleanly in two, and it raised a question I couldn’t answer.</p><h3>Most agentic fintech is a meta feature</h3><p>Look at what’s actually live, and most of it sits around the transaction, not in it.</p><p><strong>The support executive.</strong> Bajaj Finance says its AI bots now handle 71% of DIY customer service volume, and that voice AI costs it about a third of a human agent. At GFF, NPCI launched a multilingual AI assistant for UPI: transaction support, complaints and mandate management.</p><p><strong>The sales and collections caller.</strong> Mahindra Finance and Sarvam announced at GFF that their voice agents run in 12 Indian languages across sales and collections, from pre-due reminders through to repossession-stage conversations. They say it has passed a crore calls.</p><p><strong>The stock tip.</strong> This one runs straight into regulation. Since SEBI’s December 2024 amendments, investment advisers and research analysts must disclose their use of AI tools and remain solely responsible for AI-based advice. An AI can recommend a stock in India. It can’t be the one accountable for it.</p><p><strong>The loan or insurance nudge.</strong> Bajaj Finance says AI analysis of customer interactions generated around 4 lakh additional loan offers in Q1 FY27. Policybazaar runs an AI assistant that suggests term, health and motor plans.</p><p><strong>The merchant’s back office.</strong> Razorpay’s Agent Studio launched in March with eight agents built on Anthropic’s Claude Agent SDK, for disputes, subscription recovery and similar work. PayU launched Agent HQ, a no-code agent store for SMBs, at GFF. Cashfree’s Relay agents went GA in August.</p><p>Don’t get me wrong. These are real improvements. Collections in your own language beat a robotic IVR, and a 2 AM dispute answered in seconds beats a ticket that sits for three days.</p><p>But look at the core act of paying in every one of them. A human sees the amount, taps, and enters a PIN. That part hasn’t moved.</p><h3>Then there are the agents inside the payment</h3><p>A much smaller set of launches is trying to change the act itself.</p><p><strong>Pine Labs’ P3P</strong> lets an agent pay against a UPI mandate you approve once. It went live in June, with Gullak as the anchor case: buy ₹500 of gold if the price drops below a threshold. At GFF, Pine Labs ran a flash sale where you could refresh for an hour or authorise an agent. I picked the agent and won. I wrote about <a href="https://medium.com/@tejas_amle/i-won-a-flash-sale-without-being-in-the-room-de2fa75d7a72">what was running underneath</a>, so I won’t repeat it here.</p><p><strong>Razorpay and NPCI</strong> have run agentic UPI pilots inside ChatGPT (October 2025, with BigBasket and Vi) and inside Claude (February 2026, with Zomato, Swiggy and Zepto). Both are built on UPI Reserve Pay.</p><p><strong>Setu</strong>, a Pine Labs company, launched agentic bill payments on ChatGPT and Claude in December 2025.</p><p><strong>Amazon Pay’s Smart Wallet</strong>, launched at GFF, lets you authorise specific agents with spending limits. It is live for flight booking only.</p><p>And then there’s the launch that didn’t happen. NPCI’s Unified Agent Protocol, which would let agents make small UPI payments without per-transaction approval, was due at GFF on 10 September. Business Standard reported that NPCI told industry the evening before that it was on hold pending RBI clearance, with liability, governance for “rogue” agents and user protection still open.</p><p>The same day, NPCI Chairman Ajay Kumar Choudhary said on stage: “AI may recommend, but authentication and final settlement must follow deterministic auditable rules.”</p><p>Read that twice. <em>AI may recommend.</em> The body that runs UPI just described the meta-feature layer as the approved zone, and the payment act as the protected one.</p><h3>Will Indians trust an agent with their money?</h3><p>This is the question underneath everything above, and the data points in two directions.</p><p><strong>Indians trust AI more than almost anyone.</strong> In KPMG and the University of Melbourne’s 2025 global study, 76% of Indian respondents said they were willing to trust AI, against a global average of 46%. Accenture’s 2026 Consumer Pulse found 86% of Indian consumers likely to follow gen-AI tools for half of their spending decisions.</p><p><strong>But trusting AI to recommend isn’t trusting it to pay.</strong> Visa’s February 2026 Asia Pacific survey found 42% of Indians open to using AI to make online purchases. India led the region, and it’s still under half. The same KPMG study that found 76% trust also found 78% of Indians worried about negative outcomes from AI. EY’s May 2026 survey of Indian banking customers found bank chatbots “underutilized and less trusted.”</p><p><strong>And the environment is hostile.</strong> Indians lost ₹22,495 crore to cyber fraud in 2025, per Home Ministry data, with investment scams making up most of it. For a decade, every bank SMS has said the same thing: never share your OTP, never let anyone transact for you.</p><p>Agentic payments ask people to delegate exactly what they were trained to guard.</p><p>My read is that Indians trust AI as an <strong>advisor</strong> today, not yet as an <strong>agent</strong>. The meta-feature layer is growing fast because it lives inside that trust. The payment layer is slow because it lives outside it.</p><h3>Meanwhile, in the US and China</h3><p><strong>The US built every rail and found that users didn’t come.</strong> Mastercard said it would enable agent payments for all US cardholders by November 2025. Visa, Google, Stripe and OpenAI all shipped agent payment protocols. OpenAI launched Instant Checkout inside ChatGPT in September 2025 and pulled back in March 2026, sending checkout back to merchants. On 8 September, Visa CEO Ryan McInerney said: “We are seeing adoption for shopping, but not yet for autonomous payments.” American Express went furthest on trust. Its Agent Purchase Protection covers eligible charges caused by a registered agent’s error.</p><p><strong>China did it inside the walls.</strong> Alipay reported 300 million AI Pay transactions by May 2026. Alibaba’s Qwen app drove over 120 million orders in six days over Spring Festival, heavily subsidised and still with a human confirmation step. Then regulators moved: on 24 August, a PBoC-guided industry body issued China’s first convention for agent-driven payments, built around “know your agent” checks.</p><p>So the US has open rails and hesitant users, and China has users inside closed loops. India is trying for both at once, on public rails, in a market with a scam problem.</p><h3>So, is there a future in the protocol?</h3><p>This is the question I set out to answer. I can make both cases, and I haven’t picked a side yet.</p><h3>The case for yes</h3><p><strong>The primitives already exist.</strong> UPI has delegation built in. Reserve Pay blocks funds for later debits. UPI Circle lets one person delegate authority to another within limits. One-Time Mandates authorise once for a future event. Every agentic payment design in India so far reuses these rather than inventing new ones. Pine Labs CEO Amrish Rau put it this way at the P3P launch: “UPI’s mandate framework was already architected for agentic commerce.”</p><p><strong>Bounded rules feel natural.</strong> “Pay my electricity bill on the due date.” “Buy gold if it drops below this price.” “Book the cheapest flight under ₹6,000 on Friday.” None of these needs an agent with judgment. They need a rule, a mandate and an audit trail. That’s closer to AutoPay than to a robot shopper, and Indians already use AutoPay.</p><p><strong>The authentication doesn’t vanish, it moves.</strong> In every live design, the human still authenticates, just earlier: once, when attention is available, with limits and an expiry. That’s a smaller behaviour change than it sounds.</p><h3>The case for no</h3><p><strong>The tap was never the pain point.</strong> UPI already takes seconds. Automating the last two seconds of a payment is a solution looking for a problem. The real friction in Indian commerce is choosing, comparing and trusting, which is exactly where the meta features already work.</p><p><strong>Even the US couldn’t make it stick.</strong> OpenAI had the distribution, Stripe’s rails and willing merchants, and still moved checkout back to the merchant. “Discover in the AI, pay where you always pay” may simply be the stable equilibrium.</p><p><strong>The regulator isn’t in a hurry.</strong> RBI’s April 2026 e-mandate framework requires additional authentication at registration and the first debit, a 24-hour pre-debit notice, and extra factors above ₹15,000. CERT-In’s July report proposed human-in-the-loop controls above financial thresholds that haven’t been set. The national protocol is on hold. None of that is a signal to move fast.</p><p><strong>Nobody owns the loss yet.</strong> Until someone in the chain says “if a registered agent gets this wrong, we cover it,” the trust gap above doesn’t close.</p><h3>If it does happen, who owns it?</h3><p>One more thing complicates the “yes” case: protocol soup. Five systems for agent payment authority were launched, demoed or reported in India this year:</p><ul><li>Pine Labs’ <strong>P3P</strong>, live on UPI, extending to cards.</li><li>NPCI’s <strong>Unified Agent Protocol</strong>, built but held.</li><li>A <strong>RuPay Agentic Protocol</strong>, which Business Standard reported NPCI has developed, with no detail public.</li><li><strong>Visa Intelligent Commerce</strong>, demoed at GFF with Razorpay, Pine Labs, Cashfree, Juspay, PayU and others.</li><li><strong>Mastercard Agent Pay</strong>, whose first authenticated agentic card transaction in India was demoed in February.</li></ul><p>UPI’s own history is interoperability beating closed wallets. If agentic payments do become real, it’s hard to see five parallel authority systems surviving. Something consolidates them, and whoever runs that layer shapes the market.</p><h3>What has not been answered yet</h3><p><strong>Who pays when it goes wrong.</strong> Amex has an answer for its own network. India has no industry answer.</p><p><strong>Where the human-in-the-loop threshold lands.</strong> Whatever number RBI or CERT-In eventually writes will quietly define the size of India’s agentic payments market.</p><p><strong>Consent at the agent layer.</strong> The DPDP Rules phase in consent obligations through 2027. An agent that shops for you is also profiling you. Which consent covers that, and who records it, is an open design problem that sits right next to payments.</p><h3>Where I’ve landed, for now</h3><p>Every other industry got agents by making the human optional. Design, code and food ordering all got faster because the human could step back.</p><p>Payments might be the exception. The human may never step back, just earlier: one careful yes, then a machine that is structurally unable to exceed it. Or the tap stays exactly where it is, and agents become the best meta feature fintech ever shipped, sitting around UPI rather than inside it.</p><p>I don’t know which one it is yet. I suspect the people building it at Razorpay, Pine Labs and NPCI don’t fully know either, which is what makes this the most interesting product problem in Indian fintech right now.</p><p>If you build payments in India, I’d like your read. What’s the first payment you’d let an agent make without you?</p><p><em>— Tejas Amle, a PM who can build a food order in a chat window and still has to scan the QR.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a5cf435c3f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Won a Flash Sale Without Being in the Room]]></title>
            <link>https://medium.com/@tejas_amle/i-won-a-flash-sale-without-being-in-the-room-de2fa75d7a72?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/de2fa75d7a72</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[payments]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 19:54:53 GMT</pubDate>
            <atom:updated>2026-09-13T19:54:53.881Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a8iTwJIZFD5-ojogkBhTpg.jpeg" /></figure><h3>Pine Labs put humans against AI agents at Global Fintech Fest. I picked the agent. Here is what was actually running underneath, and what a flash sale becomes when everyone picks the agent.</h3><p>Somewhere between 2 and 3 PM on 9 September, a five minute window opened on the bigbasket beta app. AirPods Pro 3 and iPhone 17 at 40 percent off, Mastercard only. The exact minute was never published, which meant the only way to win as a human was to sit with the app open for an hour and tap faster than everyone else in the building.</p><p>A lot of people in that hall did exactly that.</p><p>I did not. I had authorised an agent the day before, set my limits, and put my phone away. It bought the AirPods while I was somewhere in the event. I found out from a notification. I collected the box at Booth I2 between 4 and 5 PM with a photo ID and an OTP.</p><p>The AirPods are not the interesting part.</p><h3>Pine Labs built a demo you could lose</h3><p>The activation was called <strong>Humans vs Agents</strong>, run with bigbasket and Mastercard. Two doors to the same product page. Door one: refresh for an hour. Door two: authorise an agent once, walk away, let it buy when the price drops.</p><p>Most fintech launches are a press release and a slide. This was a live, adversarial thing with real money on the table and a publicly observable loser. If the agents had failed, a few hundred people would have watched them fail.</p><p>They did not fail. That is a much stronger claim than any deck makes.</p><h3>What was actually running</h3><p>The landing page carried one line that almost everyone scrolled past:</p><blockquote>Autonomous payments are powered by Pine Labs Payments Protocol and governed by Grantex.</blockquote><p>Two systems, two jobs, and people constantly mash them together.</p><p><strong>P3P is the rail.</strong> The Pine Labs Payments Protocol, launched in June 2026, lets an agent initiate and complete a payment with no human at the point of transaction.</p><p><strong>Grantex is the authority layer.</strong> It answers a completely different question: which agent is allowed to spend what, on whose behalf, within what bounds, for how long.</p><p>Keeping those separate is the good decision here. Moving money and delegating authority have different failure modes and different regulators looking at them. One system doing both would have shipped faster and aged badly.</p><h3>The trick is that the OTP still happens, just earlier</h3><p>This drop was Mastercard exclusive, so it ran on P3P’s card rail.</p><p>The card flow is a pre-authorisation. You enter your card once, your bank runs 3DS and sends you an OTP, you approve, and the issuer places a hold on the funds. That is your authority, established while you are actually sitting there paying attention.</p><p>Then the agent buys, later, against that hold. No second OTP, because there is nobody present to type one.</p><p>The human authentication step did not disappear. It got relocated to a moment when a human existed. That is the entire idea and it is cleaner than it sounds.</p><p>And if you lose the race, nothing is captured and the hold is released. The campaign rules said so explicitly, and a fresh authorisation was needed for the next day’s window. Authority that expires on its own is the correct default and it is rarer than it should be.</p><h3>The purchase itself is an HTTP status code from 1997</h3><p>This is my favourite part.</p><p>When the agent asks for something it has not paid for, the server does not send it to a checkout page. There is nobody to look at a checkout page.</p><p>HTTP 402 has been sitting in the spec since the nineties marked reserved for future use. Turns out the future was agents. The server is not describing a payment <em>page</em>, it is describing a payment <em>requirement</em>, in a form another program can read.</p><p>The agent retries with a scoped credential attached, and gets back a verifiable receipt. That credential is bound to one resource, one amount, one expiry. It cannot be replayed or pointed at something else.</p><p>Underneath, Grantex expresses my spending limit as a permission rather than a policy.</p><p>The agent is not trusted to respect a cap. It is structurally unable to exceed one, because the token it holds does not authorise more. As a PM, that distinction is the whole ballgame. The interesting engineering in agentic commerce is not the model. It is the permission system wrapped around it.</p><h3>Now imagine everyone has one</h3><p>At GFF, agents beat humans because almost nobody had an agent. That edge has a short shelf life.</p><p>So run it forward. Every buyer in the queue is a program with a pre-approved hold, sitting on a webhook, waiting. What is the sale a contest of then?</p><p>Not attention. Everyone is ready. The winner gets decided in the gap between inventory changing and a capture landing, and that gap is made of deeply unglamorous things. Whether you got a webhook push or you are polling, and how often. Your network path to the merchant. Whether your TLS session was already warm. Whether you already hold a valid grant or have to go get one. Whether the merchant decrements stock atomically.</p><p>None of that is consumer experience. All of it is infrastructure. Which means who gets the discount starts depending on whose backend is better provisioned, and that is a fairness problem merchants have not had to solve before.</p><p>Then there is the tension I cannot stop chewing on, and it comes straight out of Grantex’s own docs rather than from me speculating.</p><p>A grant token can be verified <strong>locally</strong>. Signature, expiry, scopes, all checkable against cached keys with zero network calls. Microseconds. In a contention race, fast is the entire game.</p><p>But local verification cannot see a revocation. Grantex says this plainly: verifying locally does not prove current authorisation state unless you make an online check.</p><p>So under load, in milliseconds, the merchant picks. Verify locally and accept a window where a revoked agent still transacts. Or verify online and add a round trip to every single request, at the exact moment every request in the system arrives at once.</p><p>That is not a bug. That is the permanent trade of delegated authority at machine speed, and every agentic commerce system is going to have to sit somewhere on it. My guess is the answer is neither, and it is short lived grants instead. Issue authority that lives for the length of the drop window and let expiry do the job revocation would have had to.</p><h3>What has not been answered yet</h3><p>I am bullish on this, and I want to be honest about the open edges, because the people building it are not pretending otherwise.</p><p><strong>Authentication above ₹15,000.</strong> RBI’s e-mandate framework lets banks skip additional factor authentication on mandate backed transactions up to ₹15,000 and requires it above that. An iPhone at 40 percent off clears that line comfortably. The card pre-auth path, where 3DS happens once at setup, is a reasonable answer. Whether it is the settled answer is not something anyone has publicly resolved.</p><p><strong>Liability.</strong> Nobody in the industry agrees on who pays when an agent buys wrong. A cryptographically verifiable receipt is a genuinely useful primitive here. It establishes what happened. It does not establish who eats it.</p><p><strong>Intent.</strong> Mandates were designed for recurring, scheduled payments to a known merchant. Subscriptions, SIPs, EMIs. One off purchases an agent decides on by itself is a real extension of that. Whether NPCI eventually writes a distinct framework for agent initiated buying is, I think, the most consequential item on this list.</p><p>None of these are reasons to wait. They are the frontier. The company that ships first is the one that gets to help decide where the lines land, and that is a position worth holding.</p><h3>The part I would steal</h3><p>Pine Labs could have announced a partner integration and moved on. Instead they put the protocol in a room, gave a few hundred fintech professionals a financial reason to beat it, and let the outcome be public.</p><p>Credit to Amrish Rau, to CTO Sanjeev Kumar and the team who built the rail, and to whoever in marketing decided a protocol spec should be something a person could lose at. Nobody understands a payments protocol from a press release. Everybody understands losing a queue.</p><p>I got AirPods out of it. What I actually got was a clear read on where the next few years of commerce infrastructure gets built. Not in the model, and not in the money movement. In the thin, boring layer that decides what a piece of software is allowed to do with your money while you are not looking.</p><p><em>— Tejas Amle, a PM who won a race he did not attend.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de2fa75d7a72" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My claude skills folder was making claude Dumber. I couldn’t see it for weeks]]></title>
            <link>https://medium.com/@tejas_amle/my-claude-skills-folder-was-making-claude-dumber-i-couldnt-see-it-for-weeks-cfcb1f774f3c?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/cfcb1f774f3c</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[context-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[skills]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 18:57:29 GMT</pubDate>
            <atom:updated>2026-08-31T18:57:29.466Z</atom:updated>
            <content:encoded><![CDATA[<h3>A PM’s notes on why your skills folder is a liability, not a superpower.</h3><p>There’s a specific kind of guy at every gym. He started with whey protein. Reasonable. Then creatine, because the research is genuinely good. Then pre-workout. Then BCAAs, ZMA, ashwagandha, a greens powder, and something a guy on Reddit swore by. Six months later he’s taking eleven things every morning, has no idea which one is working, and at least two of them are quietly cancelling each other out.</p><p>I did that. With AI skills.</p><h3>The logic seemed completely airtight</h3><p>Here was my reasoning, and I want to state it clearly because I still think it <em>sounds</em> correct.</p><p>Skills are packaged expertise. A skill makes Claude better at one specific thing. Therefore more skills = better at more things. And because they only load when relevant, there’s no downside to having them around. It’s a strictly dominant strategy. Free capability. Why would you <em>not</em>?</p><p>So I went shopping. Every time someone posted a skills repo, every time a “10 best Claude skills” listicle crossed my feed, I grabbed it and dropped it into my global skills folder — ~/.claude/skills/, the one that loads into every single session, on every project, forever.</p><p>By the time I stopped, I had:</p><ul><li>Four or five design skills — UI, UX, design systems, accessibility, one that was just “make things look good”</li><li>Three research skills, all of which claimed to do competitive research</li><li>A finance skill</li><li>Two product management skills</li><li>A software engineering skill and a QA skill</li><li>And a “strategic advisor” skill that promised to make Claude think like a partner at a consulting firm</li></ul><p>Twenty-five-ish, globally scoped. I felt like I’d assembled a personal Avengers team. I was going to write PRDs, ship websites, run competitive teardowns, and produce board-ready decks, all from one terminal.</p><h3>Then the bill showed up</h3><p>A few weeks in, my usage went vertical. I was hitting my limit in ~3 hours of work instead of a full day. Sessions that used to cost 3k tokens were costing 18k. And the outputs had gotten <em>weirder</em> — longer, more scaffolded, more full of headings I hadn’t asked for.</p><p>I asked for a one-paragraph answer about a pricing edge case and got a nine-section strategic memo with a stakeholder matrix.</p><p>That was the tell. That wasn’t a cost problem. That was a <em>behaviour</em> problem.</p><h3>I was wrong about how skills work, and being wrong made it worse</h3><p>My mental model was: Claude loads all my skills into context at the start of every conversation, so a fat skills folder means a fat context window.</p><p>That is not what happens. Skills use something called <strong>progressive disclosure</strong>, and it works in three levels:</p><ul><li><strong>Level 1 — always loaded.</strong> Only the skill’s name and description go into the system prompt at startup. Roughly <strong>100 tokens per skill</strong>.</li><li><strong>Level 2 — loaded on trigger.</strong> The full SKILL.md body only enters context once Claude decides the skill matches your request. Usually under 5k tokens.</li><li><strong>Level 3 — loaded on demand.</strong> Bundled reference files cost nothing until something reads them. Bundled scripts cost nothing but their output.</li></ul><p>Do the math on my hoard: 25 skills × ~100 tokens is about <strong>2,500 tokens</strong> sitting at rest. On a 200k context window, that’s noise. It’s a rounding error.</p><p>Which meant my entire theory of the crime was wrong. The shelf wasn’t expensive. So where were the tokens going?</p><h3>The tokens weren’t going in. They were coming out.</h3><p>Here’s the mechanism I’d completely failed to think about.</p><p>That Level 1 description isn’t decoration. It’s the <em>matcher</em>. It’s the only thing Claude sees when deciding whether a skill is relevant to what you just typed. And a description can run up to 1,024 characters of someone else’s marketing copy about how essential their skill is.</p><p>Now put five design skills next to each other, all written by different people, all describing themselves in overlapping language, all claiming to cover “UI, UX, visual design, and design decisions.” Ask a question that brushes against any of it.</p><p>There is no clean match. So a few things happen, all of them bad:</p><p><strong>One.</strong> More skills fire than should. Each trigger drags in a full SKILL.md, and if that file says “see REFERENCE.md for the full framework,” now you’re paying for that too. Two or three of those chained and you’ve burned 15k tokens before Claude has said a word to you.</p><p><strong>Two.</strong> The wrong one fires. My “strategic advisor” skill had a broad, ambitious description — the kind that sounds impressive on a README. It fired on things that were not remotely strategic. Ask about a copy tweak, receive a market-positioning analysis.</p><p><strong>Three, and this is the expensive one.</strong> A triggered skill doesn’t just add knowledge — it imposes a <em>process</em>. Most skills are written as workflows: first do discovery, then produce a comparison table, then define success metrics, then list risks. So a question with a two-sentence answer gets processed through a four-phase methodology and comes back as a document. The input cost was 5k tokens. The <strong>output</strong> cost was 4,000 words I didn’t want.</p><p>I wasn’t paying for a bigger library. I was paying for a bureaucracy.</p><h3>The part that actually cost me, though, wasn’t tokens</h3><p>Tokens are a number. You can watch a number. Here’s the thing I didn’t notice for weeks.</p><p>Claude is a frontier model. It has already read more PRDs than I will see in three careers. Its <em>default</em> judgment on “write a PRD for this” is, honestly, better than most of the PRD skills floating around on GitHub — which are usually one person’s template, hardened into a rule, published on a Tuesday.</p><p>So when I installed a generic PM skill, I wasn’t adding capability. I was <strong>replacing good general judgment with someone else’s specific mediocre opinion.</strong> And I was doing it silently, in every session, for months.</p><p>That’s the reframe I wish someone had handed me on day one:</p><blockquote><em>A skill is not knowledge injection. A skill is an </em><strong><em>override</em></strong><em>.</em></blockquote><p>The consultant version of this: the value of a framework is not what it includes, it’s what it <em>excludes</em>. Every framework you adopt is a set of things you’ve agreed to stop thinking about. That’s a great trade when the framework is better than your thinking. It’s a terrible trade when it isn’t. I’d adopted twenty-five of them without reading most of them properly.</p><p>Which brings up the other thing I did wrong, and I’d rather say it out loud than pretend I didn’t. A SKILL.md is a file of instructions that a model with filesystem and bash access will read and follow. I was downloading those from strangers on the internet and installing them globally. Anthropic’s own docs are blunt about this — treat a skill like installing software, and audit anything from an untrusted source. I was running npm install on my own judgment with no lockfile and no code review.</p><h3>The three questions I now ask before installing anything</h3><p>I deleted about 18 of the 25. Here’s the filter that survived.</p><p><strong>1. Does this encode something Claude genuinely cannot infer?</strong> Good answers: my company’s actual PRD format. Our design tokens. A regulatory checklist specific to my domain. A deterministic script. Our internal API’s quirks. Bad answer: “how to do good UX.” It knows. Delete it.</p><p><strong>2. Would it annoy me if this fired on an unrelated request?</strong> If yes, it does not belong in your global folder. It belongs in a project-scoped one — .claude/skills/ inside the repo where it&#39;s relevant, not ~/.claude/skills/ where it stalks every conversation you&#39;ll ever have. This single move fixed most of my problem.</p><p><strong>3. Can I predict exactly when it will fire?</strong> If I can’t read the description and say out loud when this triggers, the matcher can’t either. Either rewrite the description to be narrow and specific — what it does <em>and</em> precisely when to use it — or don’t install it.</p><p>And one hard rule on top: <strong>no two skills in the same scope may have overlapping descriptions.</strong> Pick one design skill. If two are good, merge them. Two skills competing for the same trigger is worse than either alone, because now you have a coin flip in your workflow.</p><h3>What my setup looks like now</h3><p>Global: 7 skills. All of them encode <em>my</em> taste or <em>my</em> company’s specifics — things that are true no matter what I’m working on.</p><p>Project-scoped: everything else, sitting in the repo where it’s relevant, invisible everywhere else.</p><p>Usage: back down to roughly 40% of where it peaked. Outputs are shorter. And critically — when a skill fires now, it’s because I could have predicted it would.</p><h3>The uncomfortable PM lesson</h3><p>I built a bloated product and shipped it to a user base of one, and the user was me.</p><p>I optimized for capability-on-paper instead of behaviour-in-practice. I added features without ever asking what they cost when they misfire. I never instrumented the thing. I never questioned the assumption that more was free — even though “more, and it’s free” is the single most reliable smell in this entire job.</p><p>Curation <em>is</em> the skill. The context window is a product surface, and every token in it is competing for attention with everything else in it. Treat your skills folder like a roadmap, not a shopping cart. Something has to get cut for something else to matter.</p><p>Now go look at your global skills folder. I’ll wait.</p><p><em>— Tejas Amle, a PM who spent four weeks paying for the privilege of being told what he already knew.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cfcb1f774f3c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kundli Matching Deserves Better Than a Blurry WhatsApp Forward]]></title>
            <link>https://medium.com/@tejas_amle/kundli-matching-deserves-better-than-a-blurry-whatsapp-forward-8c0de297c07e?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c0de297c07e</guid>
            <category><![CDATA[astrotalks]]></category>
            <category><![CDATA[nakshatra]]></category>
            <category><![CDATA[kundali]]></category>
            <category><![CDATA[relationships]]></category>
            <category><![CDATA[astrology]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Fri, 14 Aug 2026 12:21:29 GMT</pubDate>
            <atom:updated>2026-08-14T12:21:29.781Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yGMMPswAjigtcJIJvSrTOw.png" /></figure><h3>How 3,000 years of Vedic astrology, a 36-point scoring system, and a schoolyard name game ended up in the same free app</h3><p>Every Indian family group chat has run this exact play. Someone gets engaged. Within the hour — before the “congratulations” stickers even finish loading — an uncle has forwarded a scanned PDF titled something like <em>KUNDLI_MILAN_FINAL_v2.pdf</em>, shot at an angle, half a thumb visible in the corner. Somewhere in there is a number out of 36. Nobody in the group fully understands how that number got calculated. Everybody has an opinion about whether it’s good enough.</p><p>I’ve watched this happen at three weddings this year alone. Not once did anyone actually <em>read</em> the chart. They read the number, and then they read the room.</p><p>That gap — between a genuinely serious 3,000-year-old compatibility system and the fact that almost nobody under 40 can parse the document it produces — is the whole reason I built Naksha.</p><h3>Twenty-seven stars, eight scores, one document nobody reads</h3><p>Quick context, because most explainers either assume you already know this or explain it like you’re studying for an exam:</p><p>Vedic astrology splits the sky into 27 <strong>nakshatras</strong> — lunar mansions, essentially a finer-grained zodiac than the 12 signs you already know. Your nakshatra is decided by exactly where the Moon sat the second you were born, and it matters <em>a lot</em> more in Indian astrology than the Sun sign does in the Western version most of us grew up half-believing.</p><p>Before a wedding, the classical move is <strong>Guna Milan</strong> — matching two people’s charts across 8 dimensions (koots, if you want the real word) — Varna, Vashya, Tara, Yoni, Graha Maitri, Gana, Bhakoot, Nadi — for a total out of 36. Eighteen-plus is “fine.” Twenty-eight-plus and aunties start smiling. It’s a real, structured system with actual math behind it, computed from birth time and place with the kind of precision that would make a Swiss watchmaker nod approvingly.</p><p>And yet the artifact that reaches an actual human is a shaky photo of a printout in a language most people can half-read. The system is legitimate. The delivery is broken.</p><h3>The existing options were ugly, expensive, or both</h3><p>If you go looking for a “kundli matching” tool online today, you land in one of two universes. The first is a website that looks unchanged since 2011 — banner ads on every edge, a chart rendered in a font size clearly optimized for a CRT monitor, and copy that reads like it was translated four times in each direction. The second is worse: a slick app that gives you <em>just enough</em> free information to get anxious, then routes you into a paid call with an astrologer who is, coincidentally, always available right now.</p><p>Neither of those is actually respecting the person on the other end of the screen. One insults your eyes. The other monetizes your nervousness.</p><p>As a product person, that’s the part I couldn’t leave alone. The astrology isn’t the broken part of this system. The <em>interface</em> is.</p><h3>Then I added a schoolyard game next to a 3,000-year-old scoring system</h3><p>Here’s the part that made a few people on my team tilt their heads: right next to the serious 36-point Guna Milan engine, Naksha also ships <strong>FLAMES</strong> — the exact playground game you played in the back of a notebook in 8th grade. Friends, Lovers, Affection, Marriage, Enemies, Siblings. Two names, cross out the shared letters, count what’s left, spell it out. Zero astronomy involved. Zero seriousness intended.</p><p>Putting those two things in the same product sounds like a category error until you actually watch how compatibility curiosity works in real life. Not everyone opening this kind of app wants a rigorous 8-dimension breakdown of their relationship’s structural integrity. Sometimes two 19-year-olds just want to type their crush’s name into something and laugh at what comes out. That impulse is <em>just as real</em> as the aunty demanding a Nadi dosha check before the engagement is official — it’s just a different door into the same house.</p><p>FLAMES is the door with no lock on it. Guna Milan is the room you walk into once you’re actually curious about what’s underneath.</p><h3>Built for the person who’s never opened a horoscope app. Also built for the person who’ll correct my Lagna in the comments.</h3><p>There’s the person who’s never looked at a birth chart and finds Lagna/Rashi/houses/dashas mildly intimidating — for them, Naksha defaults to a <strong>Reading</strong> mode, plain English, friend-explaining-it energy. And there’s the person who grew up watching a grandparent draw a diamond chart by hand and has <em>opinions</em> about North vs. South Indian style — for them, a full <strong>Chart</strong> mode: real regional layouts, complete planet table, full Dasha timeline. Same math, two completely different front doors.</p><h3>Go break a name game, or an ancient one</h3><p>Naksha is free — no signup, no ads, no astrologer upsell waiting at the bottom of the funnel. For the serious version: <a href="https://getnaksha.com/love"><strong>getnaksha.com/love</strong></a> for the full 36-point Guna Milan. For four seconds of chaos: <a href="https://getnaksha.com/flames"><strong>getnaksha.com/flames</strong></a>.</p><p>Either way, faster than that uncle’s PDF ever loaded.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c0de297c07e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I built and shipped a game this weekend. That’s not the interesting part.]]></title>
            <link>https://medium.com/@tejas_amle/i-built-and-shipped-a-game-this-weekend-thats-not-the-interesting-part-11f4be3d7854?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/11f4be3d7854</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[pm]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Wed, 22 Apr 2026 14:03:36 GMT</pubDate>
            <atom:updated>2026-04-22T14:03:36.543Z</atom:updated>
            <content:encoded><![CDATA[<h3><em>A PM’s notes from the other side of the keyboard.</em></h3><p>I spent a weekend building claude-mario-runner — a Chrome-dino-style infinite runner that lives in your terminal, plus a global leaderboard at <a href="https://claude-mario-runner.vercel.app/">claude-mario-runner.vercel.app</a>. You install it with npx claude-mario-runner, press space to jump, and your best score ends up on a public board.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OailQaruDUMsL4It.png" /></figure><p>CLAUDE MARIO RUNNER leaderboard</p><p>Cute project. That’s not what this post is about.</p><h3>What actually happened</h3><p>I’m a product manager. I don’t ship production code for a living. But in one weekend, here’s what landed:</p><ul><li>A real-time terminal game with physics, sprites, collision detection, seeded RNG for reproducible runs, and a deterministic 30 Hz fixed timestep</li><li>A public Next.js leaderboard on Vercel’s Fluid Compute</li><li>An Upstash Redis backend with ZADD GT for atomic max-only writes, rate limiting, and reserved handles</li><li>GitHub device-flow OAuth for verified handles</li><li>An offline queue that retries submissions the next time you launch the game</li><li>A CI pipeline. Tests. A published npm package. Light/dark mode. Pixel cursors. A retro font stack modeled after Claude’s design system.</li></ul><p>The last time I tried to build something half this ambitious, I spent three weekends fighting TypeScript configs and gave up on the API route.</p><h3>The PM-to-SDE distance is collapsing</h3><p>Here’s the thing I can’t stop thinking about.</p><p>For a decade, the PM job has been about specifying intent — writing PRDs, drawing Figmas, arguing in reviews — and then handing that intent to engineers who translate it into code. The handoff was the job. The handoff was the <em>entire</em> job, really. Most of what a PM does is reduce the cost of a handoff that shouldn’t have needed to happen.</p><p>That cost is now approaching zero.</p><p>I’m not saying PMs will replace engineers. I’m saying the two jobs are going to merge into something new. The person who owns a feature will increasingly be the person who builds a first working version of it. Not because “everyone will code” — that tired prediction has been wrong for thirty years — but because the <em>description</em> of software and the <em>execution</em> of software are becoming the same artifact. When I said “add a light/dark toggle that persists, no FOUC on reload,” what came back was production-grade TypeScript with a blocking inline head script. I didn’t write TypeScript. I wrote a spec. The spec shipped.</p><p>The next few years, the title on my LinkedIn might not be PM or SDE. It’ll be something in between. People who can hold a product in their head and also type into a terminal. The boundary wasn’t natural — it was a function of how expensive the translation step used to be.</p><h3>Claude is shipping at a pace I’ve never seen</h3><p>Half the reason the above even works is that the tooling is sprinting. Every other day there’s another release — better tool use, better agents, new skills, background tasks, the IDE integration, prompt caching improvements. I’m not a fanboy. I’m just stating what I’m observing: the release velocity is compressing what used to be quarterly roadmap items into weekly drops.</p><p>For a PM, this is both thrilling and uncomfortable. Thrilling because leverage. Uncomfortable because if your moat was “I know how to scope a 6-week project,” that moat is being refilled with concrete.</p><p>My take: bet on the compression. Build things. Ship small. The gap between “I have an idea” and “someone in Bangalore just typed their name into my leaderboard” used to be a quarter. For me, it was Saturday afternoon.</p><h3>Try it</h3><pre>npx claude-mario-runner</pre><p>Leaderboard: <a href="https://claude-mario-runner.vercel.app/">claude-mario-runner.vercel.app</a><br>Source: <a href="https://github.com/TejasAmle/claude-mario-runner">github.com/TejasAmle/claude-mario-runner</a></p><p>If you beat my score, I’ll be annoyed, which is its own kind of feature.</p><p><em>— Tejas Amle, a PM who spent the weekend as an SDE and isn’t sure there’s a difference anymore.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=11f4be3d7854" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chrome for Testing vs Regular Chrome: Comprehensive Technical Analysis]]></title>
            <link>https://medium.com/@tejas_amle/chrome-for-testing-vs-regular-chrome-comprehensive-technical-analysis-56e053ac474d?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/56e053ac474d</guid>
            <category><![CDATA[chrome]]></category>
            <category><![CDATA[quality-assurance]]></category>
            <category><![CDATA[qa]]></category>
            <category><![CDATA[chrome-for-testing]]></category>
            <category><![CDATA[testing]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Sun, 15 Jun 2025 10:48:26 GMT</pubDate>
            <atom:updated>2025-06-15T10:48:26.701Z</atom:updated>
            <content:encoded><![CDATA[<p>Chrome for Testing represents Google’s dedicated solution for browser automation and testing, launched in June 2023. While built from the same codebase as regular Chrome, it incorporates targeted modifications that fundamentally change its behavior for testing environments. <strong>The primary difference is the complete removal of auto-update functionality, but the implications extend far beyond just version control — affecting security, performance, system integration, and developer workflows in significant ways.</strong></p><p>This specialized browser variant addresses critical pain points in web testing by providing versioned binaries synchronized with Chrome releases while maintaining near-identical web platform compatibility. However, these testing optimizations come with important trade-offs that developers must understand.</p><h3>Architecture and core design differences</h3><p>Chrome for Testing’s architecture prioritizes <strong>consistency and reproducibility over user experience optimizations</strong>. The browser disables auto-update mechanisms entirely, preventing the version inconsistencies that plague automated testing environments. This architectural decision cascades through every aspect of the browser’s operation.</p><p>The build and distribution infrastructure operates independently from regular Chrome, with versioned binaries published to Google Cloud Storage buckets across all release channels (Stable, Beta, Dev, Canary). <strong>ChromeDriver releases are synchronized with Chrome for Testing versions starting from Chrome 115</strong>, eliminating the notorious version compatibility issues that have plagued browser automation for years.</p><p>Unlike regular Chrome’s gradual rollout strategy, Chrome for Testing binaries are released simultaneously across all supported platforms (linux64, mac-arm64, mac-x64, win32, win64) with comprehensive JSON API endpoints for programmatic discovery and download automation.</p><h3>Security architecture and vulnerability management</h3><p>The security model differences between Chrome and Chrome for Testing are <strong>fundamental and represent the most significant risk factor</strong> for organizations considering adoption. Regular Chrome receives automatic security updates within 24 hours of critical vulnerability disclosure, while Chrome for Testing requires manual intervention for security patches.</p><p><strong>Chrome for Testing installations remain vulnerable to known exploits</strong> once security fixes are published in the Chromium repository. Recent critical vulnerabilities (CVE-2025–4664, CVE-2025–2783, CVE-2025–5419) demonstrate the real-world implications of this design choice. Google explicitly warns that Chrome for Testing is “unsafe to use for general browsing” and should only consume “trustworthy content.”</p><p>The certificate and SSL/TLS handling remains functionally identical to regular Chrome, but certificate store updates lag behind regular Chrome due to the disabled update mechanism. Safe Browsing protections may be reduced or modified to prevent interference with automated testing scenarios.</p><p>Security feature differences include disabled auto-update mechanisms, reduced Safe Browsing integration, simplified permission systems optimized for testing, and potential bypasses for security policies via command-line flags like --disable-web-security and --ignore-certificate-errors.</p><h3>User interface and experience modifications</h3><p>Chrome for Testing implements distinct visual indicators to differentiate itself from regular Chrome. The browser displays <strong>“Chrome for Testing” branding</strong> in window titles and shows persistent notification bars warning users that the browser is “reserved for automated testing” and recommending standard Chrome for normal browsing.</p><p>The user interface removes elements that could interfere with automated testing while maintaining identical rendering engines (Blink) and JavaScript execution environments (V8). Information bars can be disabled using the --disable-infobars flag introduced in November 2023.</p><p>Extension support represents a significant limitation — <strong>Chrome for Testing cannot install extensions from the Chrome Web Store</strong> through traditional methods. Extensions must be loaded manually as unpacked extensions or through automation frameworks. This limitation affects developer workflows significantly, as commonly used browser extensions are unavailable.</p><h3>Browser APIs and web platform compatibility</h3><p>Despite architectural differences, Chrome for Testing maintains <strong>identical web platform compatibility</strong> with regular Chrome. Both browsers support the same HTML5 features, CSS specifications, JavaScript APIs, and web standards. The WebDriver BiDi protocol implementation is enhanced in Chrome for Testing, providing better automation capabilities starting from Chrome 106.</p><p>Key API behavior differences focus on automation enhancement rather than web compatibility changes. The headless mode implementation uses unified code paths between headless and headful modes (Chrome 112+), making headless testing more accurate compared to legacy headless implementations.</p><p>Browser fingerprinting characteristics may differ due to testing-specific configurations and modified user agent strings. However, core web APIs like Web Storage, IndexedDB, Service Workers, and Web Assembly maintain identical functionality across both browsers.</p><h3>Extension ecosystem and developer tools</h3><p>Extension support limitations extend beyond just installation methods. While Chrome’s extension ecosystem provides thousands of productivity and development tools, <strong>Chrome for Testing requires alternative workflows</strong> for extension-dependent development processes. Testing extensions and automation tools function properly, but general-purpose extensions face compatibility challenges.</p><p>Developer Tools functionality remains nearly identical between both browsers, with full access to Network panels, Performance profiling, Console debugging, and Elements inspection. Remote debugging capabilities are enhanced in Chrome for Testing with better support for the --remote-debugging-port flag and automation frameworks.</p><p>The Chrome Developer Tools Protocol (CDP) receives optimization in Chrome for Testing with enhanced debugging features enabled by default and improved protocol monitoring capabilities for development workflows.</p><h3>Privacy features and data management</h3><p>Privacy architecture differs significantly between regular Chrome and Chrome for Testing. <strong>Google Account integration is limited or disabled</strong> in Chrome for Testing to prevent test interference, eliminating the comprehensive data synchronization capabilities that regular Chrome users expect.</p><p>Chrome Sync functionality is minimal, with user data stored locally rather than synchronized to Google servers. This design choice reduces privacy protections while also reducing data collection, creating a double-edged privacy impact.</p><p>Tracking and analytics integration is reduced in Chrome for Testing, with disabled or limited telemetry collection. However, this reduction also eliminates privacy protections that depend on Google’s backend services, potentially exposing users to different privacy risks during testing scenarios.</p><h3>Hardware acceleration and graphics performance</h3><p>Both browsers maintain <strong>identical hardware acceleration capabilities</strong> for GPU-accelerated graphics, WebGL rendering, and video decode acceleration. Chrome for Testing supports the same hardware-accelerated codecs (H.264, VP9, AV1) with identical performance characteristics.</p><p>The key difference lies in optimization philosophy — regular Chrome continuously adapts hardware acceleration based on usage patterns and system capabilities, while Chrome for Testing provides more predictable graphics performance optimized for testing consistency rather than adaptive enhancement.</p><p><strong>Hardware acceleration can be controlled more precisely</strong> in Chrome for Testing through command-line flags, enabling consistent performance testing scenarios. Graphics performance remains predictable across test runs, eliminating variability that could affect visual testing and performance benchmarking.</p><h3>Network and connectivity architecture</h3><p>Network stack implementation remains identical between both browsers, with full HTTP/2, HTTP/3, and QUIC protocol support. However, <strong>background networking activity is reduced</strong> in Chrome for Testing, providing more predictable network performance characteristics for testing scenarios.</p><p>Chrome for Testing supports enhanced network simulation capabilities with identical network throttling and offline simulation features. The browser can bypass certain network security restrictions for testing purposes using flags like --disable-web-security (requires --user-data-dir).</p><p>Connection handling patterns are more consistent in Chrome for Testing due to reduced background processes and simplified resource management, making network performance testing more reliable and reproducible.</p><h3>Performance optimization and resource management</h3><p>Performance characteristics differ significantly between the browsers due to architectural choices. <strong>Chrome for Testing eliminates background optimization processes</strong> that regular Chrome uses for memory management and CPU utilization optimization.</p><p>Regular Chrome implements Memory Saver features, dynamic tab management, and adaptive performance optimization based on usage patterns. Chrome for Testing provides <strong>consistent resource usage patterns</strong> that are more predictable for performance testing but may sacrifice some optimization benefits.</p><p>CPU resource management is simplified in Chrome for Testing with fewer background services competing for processor resources. Memory allocation patterns remain consistent across test runs, enabling reliable performance benchmarking and regression testing.</p><p><strong>Developer reports indicate mixed performance experiences</strong> — some teams observe slower navigation and increased resource usage in Chrome for Testing, while others benefit from the consistency and reduced background processing overhead.</p><h3>Cookie handling and storage mechanisms</h3><p>Cookie and storage handling maintains functional parity between both browsers, with identical support for HTTP cookies, Web Storage APIs, IndexedDB, and Cache APIs. However, <strong>third-party cookie policies may differ</strong> between the browsers.</p><p>Regular Chrome implements the ongoing phase-out of third-party cookies with Privacy Sandbox APIs, while Chrome for Testing may retain third-party cookies for testing compatibility. Local storage operates in isolation in Chrome for Testing, separated from regular Chrome profiles and data directories.</p><p>Storage management in Chrome for Testing supports custom data directories via --user-data-dir for isolated testing environments, enabling parallel test execution without data contamination between test runs.</p><h3>Media playbook and codec ecosystem</h3><p>Both browsers maintain <strong>identical codec support</strong> for audio and video playback. Supported video codecs include H.264 (AVC), VP8/VP9, and AV1 with equivalent hardware acceleration capabilities. Audio codec support encompasses AAC, Opus, MP3, and Vorbis with identical performance characteristics.</p><p>Autoplay policies represent a key difference — regular Chrome enforces strict autoplay restrictions based on user engagement, while Chrome for Testing can bypass these restrictions using --autoplay-policy=no-user-gesture-required for automated testing scenarios.</p><p>Media Session API support remains identical across both browsers, but Chrome for Testing provides enhanced programmatic control over playback states optimized for automated media testing scenarios.</p><h3>Certificate and SSL/TLS security handling</h3><p>Certificate validation mechanisms remain functionally identical between both browsers, with equivalent SSL/TLS protocol support and certificate chain validation. However, <strong>certificate store updates lag behind regular Chrome</strong> due to disabled automatic updates.</p><p>Chrome for Testing supports testing-specific certificate handling with command-line flags like --ignore-certificate-errors for development scenarios. Root certificate store updates require manual browser updates, potentially creating compatibility issues with newer certificate authorities.</p><p>Security warning systems maintain similar functionality, but may have reduced integration with Google’s Safe Browsing service for real-time threat detection during testing scenarios.</p><h3>Command-line interface and browser flags</h3><p>Chrome for Testing supports <strong>over 1,400 command-line switches</strong> identical to regular Chrome, with enhanced support for automation-specific flags. Testing-focused flags include --disable-hang-monitor, --disable-infobars, --disable-popup-blocking, --enable-automation, and --test-type=gpu.</p><p>Security bypass flags like --disable-web-security, --ignore-certificate-errors, and --disable-site-isolation-trials are commonly used in Chrome for Testing but represent security risks if used inappropriately.</p><p>Process management flags such as --disable-dev-shm-usage, --no-sandbox, --disable-background-networking, and --disable-renderer-backgrounding enable fine-tuned control over browser behavior for testing optimization.</p><h3>Update mechanisms and version lifecycle</h3><p><strong>Update philosophy represents the fundamental architectural difference</strong> between these browsers. Regular Chrome implements continuous automatic updates with rolling deployments across the user base over days or weeks, while Chrome for Testing requires explicit manual version management.</p><p>Chrome for Testing provides versioned binary archives synchronized with Chrome release cycles, enabling precise version control for regression testing and CI/CD pipeline consistency. The JSON API infrastructure supports programmatic version discovery and download automation.</p><p>Version compatibility matrices eliminate ChromeDriver compatibility issues that have historically plagued browser automation. However, <strong>security update responsibility shifts entirely to development teams</strong>, requiring active monitoring of security advisories and manual update processes.</p><h3>Operating system integration and system services</h3><p>System integration philosophy differs significantly between the browsers. <strong>Regular Chrome implements deep OS integration</strong> with comprehensive default browser capabilities, system notification integration, and OS-level performance optimizations.</p><p>Chrome for Testing provides minimal OS integration to avoid testing interference. File associations are simplified, system notifications are limited, and default browser functionality is disabled since the browser is not intended for regular use.</p><p>Background service architecture is streamlined in Chrome for Testing with reduced background processes, disabled sync services, and simplified service worker management. This design choice reduces system resource consumption while eliminating potential sources of test inconsistency.</p><h3>Profile and user data management</h3><p>Profile management represents a significant operational difference. <strong>Chrome for Testing cannot share profiles with regular Chrome installations</strong>, requiring separate user data directories and configuration management.</p><p>User data synchronization is minimal or disabled entirely, with local storage taking precedence over cloud-based data management. This isolation enables reliable test execution but eliminates the convenience features that regular Chrome users expect.</p><p>Profile creation and management must be handled programmatically in testing environments, with custom data directories specified via command-line arguments to prevent data contamination between test runs.</p><h3>Notification systems and permission handling</h3><p>Permission management systems are simplified in Chrome for Testing, optimized for automated testing scenarios rather than user privacy protection. <strong>Notification systems may be disabled or reduced</strong> to prevent interference with automated test execution.</p><p>Regular Chrome implements comprehensive permission systems with granular user privacy controls, while Chrome for Testing focuses on predictable permission behavior for testing consistency. Location services, camera access, and microphone permissions operate with simplified approval mechanisms.</p><p>The permission API maintains functional compatibility across both browsers, but permission persistence and user interface behavior may differ significantly in testing environments.</p><h3>Sync and Google account integration limitations</h3><p><strong>Google Account integration represents one of the most significant functional differences</strong> between these browsers. Regular Chrome provides comprehensive data synchronization including bookmarks, passwords, browsing history, and application settings across devices.</p><p>Chrome for Testing eliminates or severely limits Google Account integration to prevent testing interference and ensure data isolation. Bookmark synchronization, password management, and settings sync are unavailable, requiring alternative data management strategies for testing environments.</p><p>This limitation affects developer workflows significantly, as teams cannot rely on shared bookmarks, synchronized passwords, or consistent browser settings across testing environments without manual configuration management.</p><h3>Conclusion</h3><p>Chrome for Testing represents a fundamental architectural shift from user-focused optimization to testing-focused consistency. <strong>The browser successfully addresses critical testing pain points — version inconsistency, ChromeDriver compatibility, and update interference — while introducing significant security, privacy, and functionality trade-offs</strong> that organizations must carefully evaluate.</p><p>The most critical considerations for adoption include the security implications of disabled automatic updates, the operational overhead of manual version management, and the functional limitations around extensions and Google Account integration.</p><p>For web development and automated testing workflows, Chrome for Testing provides unprecedented reliability and reproducibility.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56e053ac474d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Simplifying Tax on Crypto 2022]]></title>
            <link>https://medium.com/@tejas_amle/simplifying-tax-on-crypto-2022-48e47f5ab578?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/48e47f5ab578</guid>
            <category><![CDATA[taxes]]></category>
            <category><![CDATA[crypto-taxes]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[crypto]]></category>
            <category><![CDATA[td]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Sun, 31 Jul 2022 08:46:16 GMT</pubDate>
            <atom:updated>2022-07-31T08:46:16.386Z</atom:updated>
            <content:encoded><![CDATA[<p>Our Attempt to Save You From Crypto Tax? Not Literally</p><blockquote><em>Your Profit is </em><strong><em>Our</em></strong><em> Profit; Your Loss is </em><strong><em>Your</em></strong><em> Loss.</em></blockquote><p><em>Directed by: The Government<br>Produced by: The Meme World</em></p><p>This is what precisely the meme world proclaimed as the government announced its taxing on digital assets for this fiscal year on February 1st, 2022.</p><p>As per section 194S of the Income Tax Act, the person making a payment to a resident on a purchase of Virtual Digital Assets (cryptocurrencies, NFTs, etc) shall be responsible for deducting TDS at the rate of 1% of the consideration period or cryptos effective from July 1st, 2022.</p><h3><strong>How does crypto TDS work?</strong></h3><p>From July 1st, 2022, if you have had crypto at your disposal and you decided to sell it off then you must be aware of the one percent Tax Deducted at the Source (TDS) of the transaction value. Let us dive deeper into section 194S &amp; see what’s in it for crypto enthusiasts like us.</p><p>If you are buying &amp; selling through an exchange, then no TDS will be deducted from the buyer on buying crypto using INR while the seller of the cryptocurrency will be liable to pay TDS. The crypto exchange will deduct this amount and deposit it with the Income Tax Department and the seller would be able to offset this 1% from their total tax liability as they file for tax returns.</p><p>A TDS certificate will also be issued in due course. However, in a P2P, direct buyer-to-seller transaction, the buyer is required to deduct tax under section 194S.</p><p>The next important thing is how you are selling your VDAs. If it is in INR then the crypto exchange will be deducted from the seller. However, if the transaction is done in crypto pairs then a 1% TDS deduction will take place from both the parties in respective coins which will then be converted to INR.<br>For instance, the sale of 1 Bitcoin from 10 Ethereum, will incur a deduction of 0.01 BTC &amp; 0.1 eth, being 1% of the sale value as TDS. The crypto deducted as TDS will be liquidated to INR at the end of the day and this amount will be paid as TDS to the government.</p><p>The tax will be deducted only if the consideration exceeds 10,000 INR for the fiscal year. In the case of individuals or HUFs (Hindu Undivided Families) who are not into business or profession or not subject to audit under the Income Tax, the limit would be 50,000 INR.</p><p>The TDS shall be charged on net consideration which is sale value after reducing charges &amp; GST.</p><h3><strong>A Deeper Dive into 194S</strong></h3><p>Going by the finance bill of 2022, no set off of any loans arising from the transfer of VDAs shall be allowed against any income computed under any other provision of the act, and such loss shall not be permitted to be carried forward to subsequent assessment years. Plus, a one percent Tax Deducted at Source (TDS) upon selling these VDAs is a mandate. Let us understand this with an example.</p><p>Say you invest 1 lakh rupees in bitcoin and (god forbid) incur a fifty thousand loss. You decide to book your losses &amp; exit. The catch is you’ll not receive 50,000 but 49,500 INR after a 1% tax deduction at source. Be it profit or loss, TDS will be deducted. However, if you choose to continue your crypto investment journey with the balance of 49,500 INR in Ethereum, let’s say Eher does well and your investment goes up to 80,000 INR. You book profit &amp; you obtain your 79,200 INR after a 1% deduction.</p><p>Summing up, 50,000 loss on bitcoin &amp; 30,500 profit on Ether, making your net loss 19,500 INR. Total TDS paid amounts to 13,500 INR. This TDS will be refunded as you show you have poked a net loss.</p><p>In the next fiscal, say you continue to invest in cryptos. You purchase Ether for another one lakh &amp; which amounts to 1.4 lakh INR. You book profit &amp; attain 1,38,600 INR, after 1% TDS.</p><p>When you file your ITR for this fiscal, the 19,500 loss from last year and 40,000 profit from this year, making your net profit 20,500 INR. You can pay 30% tax on this (6,150).<br><em>ERROR 404: Profit Not Found</em></p><p><strong>Why?</strong><br>If this was equity trading, you would have done this, but not with crypto. The government says that losses from Virtual Digital Assets (VDAs) transactions cannot be carried forward. The benefit is available for businesses, MFs &amp; stocks but not for VDAs.</p><p>Hence, you must pay 30% of 40,000 INR i.e. 12,000 INR. Since 1400 was paid as TDS, the tax to be paid amounts to 10,600 INR. This justifies the saying, your profit is our profit &amp; your loss is just your loss.</p><p>Zerodha’s founder Nithin Kamath highlighted that since 1% TDS will be deducted for each transaction, for any active trader, 50% of the capital will get blocked in TDS for 50 trades in a financial year, irrespective of your profit or loss. For any market to sustain itself, active traders are very important as they provide liquidity.</p><p>Furthermore, the government will not allow other costs like platform fees, broker fees, or internet charges to be deducted as expenses from the profit except for the cost of acquisition which is the purchase price. This is allowed in stocks &amp; derivatives trading.</p><p>Overall, the government did not legalize or ban crypto but it has made a move to discourage short-term trading at least.</p><p>This was my bit to simplify Taxation norms on crypto and save you from the tax anxiety (if not from the tax itself). If you liked this, don’t miss out on this tweet by Nischal Shetty, Founder of WazirX, where he reveals quick tips on Tax on Crypto.</p><h3>Nischal (Shardeum) ⚡️ on Twitter: &quot;1/Something important for Indian crypto traders to understand * Trading on foreign exchanges that don&#39;t deduct TDS would mean YOU have to pay TDS directly to Income Tax Department* You need to know PAN of the seller on international exchanges* May be asked to pay 20%... / Twitter&quot;</h3><p>1/Something important for Indian crypto traders to understand * Trading on foreign exchanges that don&#39;t deduct TDS would mean YOU have to pay TDS directly to Income Tax Department* You need to know PAN of the seller on international exchanges* May be asked to pay 20%...</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48e47f5ab578" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Here’s how AI can transform the lives of disabled]]></title>
            <link>https://medium.com/@tejas_amle/heres-how-ai-can-transform-the-lives-of-disabled-40c5d6a4225e?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/40c5d6a4225e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Wed, 01 Dec 2021 06:33:23 GMT</pubDate>
            <atom:updated>2021-12-01T06:33:23.023Z</atom:updated>
            <content:encoded><![CDATA[<p>Many believe that artificial intelligence is a futuristic concept that we only see in sci-fi movies with humanoid robots and holograms. However, it is becoming rooted in our reality, affecting various fields and groups, including persons with disabilities. Accessibility and inclusivity are genuinely revolutionized, thanks to artificial intelligence! People with disabilities can substantially enhance their daily life thanks to AI technology solutions.</p><p>We’ve already shown how smartphones can be tools for people with vision impairments. Many apps do indeed allow them to remain self-sufficient. Visually handicapped people, for example, can effortlessly read their mail using Seeing AI by placing documents beneath the smartphone camera. Any form of handicap profile can benefit from AI technology. People with limited mobility, for example, can control anything at home simply by speaking to a virtual personal assistant like Amazon Alexa.</p><h3>Human connection and communication</h3><p>AI-assisted voice-assist devices such as Echo, Google Home, and Alexa have opened up new avenues of accessibility for those with disabilities. Because Artificial Intelligence plays such a vital part in communication and engagement, it allows people with disabilities to access information much more quickly just by speaking to their gadgets.</p><p>Speech-to-text and text-to-speech technology advancements are helping those with speech difficulties. Voice systems, such as Voiceitt, may learn how to pronounce words over time and convert them into clear, standardized speech in the form of audio or text messages. The program could potentially assist persons with speech problems in communicating with one another face to face. Similarly, Google’s Parrotron is an AI tool that converts damaged speech patterns into fluent speech, making it easier for speech-challenged persons to be understood.</p><h3>Increasing accessibility</h3><p>Accessibility is aided by technology in general and artificial intelligence in particular. It’s not just about discovering the latest breakthroughs; it’s also about putting a solution in the hands of a specific group of people to better their lives.</p><p>It can do so by implementing a variety of solutions, including:</p><ul><li>People with visual impairments can use image recognition software.</li><li>People with vision impairments can use facial recognition software.</li><li>Lip-reading recognition for persons who are deafeningly deafeningly deafeningly deafeningly deaf</li><li>For those with mental disabilities, a text summary is helpful.</li><li>Real-time captioning or translations are available for persons with hearing impairments or who don’t speak the language.</li></ul><h3>Workplace Accessibility</h3><p>Many national and international companies lookout for ways to go beyond the conventional. Along with equal opportunity employment laws, they want to create more accessible and inclusive workplace settings.This usually includes more than just changing the physical workspace to accommodate people with impairments.</p><p>Instead, firms are embracing artificial intelligence benefits to increase job chances.Microsoft, for example, has launched a five-year, $25 million AI for Accessibility project to expand the number of assistive technologies and solutions available in the workplace.</p><p>People with disabilities will benefit immensely from these AI solutions because they will have more economic possibilities, job passions, and a stronger feeling of community and connection.</p><h3>Artificial Intelligence (AI) for Prosthetics</h3><p>Despite the numerous advancements that indirectly assist people with disabilities, parts of AI research are directly aimed at easing the disability in close collaboration with medical research. The development of prostheses is one of the fascinating areas of research in this area.</p><p>Compared to early prostheses that were fully static, mechanically flexible limbs have become ubiquitous. Recent breakthroughs in algorithm design and more powerful and compact chips have led to considerable gains in this area. Because of better signal processing algorithms, a modern prosthesis reacts faster to nervous system input and adapts to the environment, such as various floor conditions. This enables far more natural, intuitive mobility while also reducing the gap between artificial and natural limbs.</p><h3>Artificial Intelligence (AI) for People with Low Vision</h3><p>Although assistive technology for the blind has been around for a while, such as screen readers and braille displays, AI has recently led to substantial advancements. Text-to-speech (TTS) technology has grown considerably more versatile, allowing mobile ad-hoc use, such as listening to an e-book hands-free, thanks to more powerful voice synthesis algorithms, cellphones, and dedicated assistive devices. This is also linked to developments in image processing, particularly in optical character recognition (OCR), which allows text to be converted into audio. Google Lens is a fantastic example of this, as it can recognize and analyze text in a camera image of a traffic sign.</p><p>However, the user interface design, particularly error tolerance, is a critical feature that requires even more attention. For most users, incorrect command recognition or a failure to load is a minor annoyance. Still, a disabled person who cannot make an emergency call can be a severe impediment or even life-threatening.</p><h3>Conclusion</h3><p>Many AI technologies already exist that benefit disabled individuals, either as by-products of other technology or specially created for them. However, several technological hurdles remain in the way of AI having a significant impact on this vast and particularly vulnerable demographic. To achieve real progress toward inclusion, politicians and businesses must pay more attention to disabilities in product development and, ideally, include disabled people in the process.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40c5d6a4225e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Make yourself a cartoon character with Artificial Inteligence]]></title>
            <link>https://medium.com/@tejas_amle/make-yourself-a-cartoon-character-with-artificial-inteligence-eee5f2da76ad?source=rss-e26a2f50b232------2</link>
            <guid isPermaLink="false">https://medium.com/p/eee5f2da76ad</guid>
            <category><![CDATA[ｍedium]]></category>
            <category><![CDATA[generative-adversarial]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificail-intelligence]]></category>
            <dc:creator><![CDATA[Tejas Amle]]></dc:creator>
            <pubDate>Sun, 15 Nov 2020 11:01:53 GMT</pubDate>
            <atom:updated>2020-11-15T11:05:00.978Z</atom:updated>
            <content:encoded><![CDATA[<p>Did you ever in your childhood want to be part of any cartoon movie or series? Did you ever want to play a cartoon character role?<br>Now you can do all of that with the help Artificial Intelligence and Machine Learning. In this blog tutorial you will get to know everything about the technologies used for doing this and how can you actually code that.</p><h3>Content :</h3><ul><li>GAN (Generative Adversarial Networks)</li><li>StyleGAN</li><li>Programing Tutorial</li></ul><h3><strong>GAN (Generative Adversarial Networks)</strong></h3><p>GAN is a very exciting and very recent innovation and thus you can find many research paper on this topic these days. GANs are basically generative models which can create new data instances that resemble or are just like our training data. <br>For example, GANs can create images that look like photographs of human faces,which never existed or which are not real persons with the help of our training data.<br>One more example of GAN is changing a normal horse to zebra. The basic body structure of horse and zebra are the same, but with the help of GAN we can change the body color of the animal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/448/0*v_ta5agNLBxHUzki.gif" /><figcaption>Right side video created by CycleGAN</figcaption></figure><p><strong>Generator and discriminator</strong></p><p>The basic structure of GAN comprises two deep networks called Generator and Discriminator.<br>In simple layman terms:</p><ul><li>Generator: It generators images based on the given training dataset and noise.</li><li>Discriminator: It discriminates between the various kinds of data instances and then tries to change it according to its need.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_riZ88SX7g_oSQ7z.png" /></figure><h3>StyleGAN</h3><p>So in 2018 Nvidia announced its new paper called StyleGAN which generates images. There are many researches about GANs, but one of the most famous models is StyleGAN, which was proposed at the end of 2018. The figure below is an image generated by StyleGAN. You can see that a high-resolution image is generated that can be mis-recognized as a real one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/579/0*G5cZ1iVTnzjGOmLY.png" /></figure><p>One of the great feature of this is that it disentangles the latent space which allows us to control the attributes at a different level or we can say at differnet features. For example, the lower layer would be able to control the body shape while the higher layer control upper high level features such as the lighting or texture on the face.</p><p>The color in the beside images represents a cluster formed where you can think of it as the controllable part or lets say features of the image. Finally, in the first few layers, different parts of the head are represented as different clusters which means that it controls the pose and shape of the human. In the last layer, you can see that different parts of the lighting effects on the face are represented as different clusters. In the middle part of the layers, the facial features such as the mouth or eyes are represented as different clusters which means there is a variation in features.</p><h3>Now let’s start to code:</h3><p>Now that all of you have understood the basic concepts of GAN and StyleGAN and how it works, lets get on how to code yourself to be a cartoon character.</p><p>I will be also providing my colab link for this so that you can directly clone and run the file, but for those who also want to understand each and every code line as well, I will be also helping you for that.</p><p>Enough of talk, now let’s get started.</p><p>Here is the <a href="https://colab.research.google.com/drive/1QXu_b8Kq4Mnl5vtByaF3Dh4WjWeRHmHO?usp=sharing">Colab Notebook</a> link to help you along. <em>(Dont forget to make a copy in your drive and use GPU runtime)</em></p><p>This project runs in tensorflow verions 1, so make sure to use that.</p><pre>%tensorflow_version 1.x</pre><p>We will now clone this respository from Github.<br><em>Note: ! and % are special shell commands for command line.<br></em>cd will change your directory to the stylegan dir. <br>mkdir command make new dirctory or folder.</p><pre>!git clone <a href="https://github.com/justinpinkney/stylegan2">https://github.com/justinpinkney/stylegan2</a><br>%cd stylegan2<br>!nvcc test_nvcc.cu -o test_nvcc -run<br>!mkdir raw<br>!mkdir aligned<br>!mkdir generated</pre><p>Now upload your picture with your face in the raw directory. There is a script which which will automatically crop your face in 256x256 square image.</p><p>We will use here Elon Musk’s photo to animate or cartoonfy him.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*vyv43uQpwjlISPdo.jpeg" /></figure><h4>Upload your own photos</h4><p>Upload your photos with your face to raw/ folder. Nothing is to be done with the image, everything will be done by the script. You should get a pretty high-res picture of a face to get a sharp result.</p><p>We will now get an example image from net to work with.</p><p>The basic process of the script is:</p><ul><li>Extract faces and align the images according to the face</li><li>Project the images (i.e. find the latent code)</li><li>Toonify the images usinf given model(i.e. use the latent code with the toon model)</li></ul><p>Final results with the tonified image will be placed in the stylegan2/generated folder</p><pre>!wget https://drive.google.com/uc?id=1ZwjotR2QWSS8jaJ12Xj00tXfM0V_nD3c -O raw/example.jpg</pre><p>After this we will load the pre-trained model</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/605df72bbfe774be2cb26d0d268a6bc6/href">https://medium.com/media/605df72bbfe774be2cb26d0d268a6bc6/href</a></iframe><p>Then we will run the python scripts which will crop the image —&gt;</p><pre>!python align_images.py raw aligned</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/256/0*g-CEQ4sW1baPpC5_.png" /><figcaption>Croped image after running script</figcaption></figure><p>Finally, we will project the image to the latent space which will create the latent vector which will be saved as .npy file in the generated directory.</p><pre>!python project_images.py --num-steps 500 aligned generated</pre><p>Now that we have the latent vector with us, we can now generate the cartoon faces using our pre-build model.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/2b2fc3ec4744afa5e3218af01cc6c757/href">https://medium.com/media/2b2fc3ec4744afa5e3218af01cc6c757/href</a></iframe><p>Heyyy, congrats the cartoony face is ready which is in the genrated folder.<br>To see the genrated image run this code.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a274990520244f57d3c6180085e451fb/href">https://medium.com/media/a274990520244f57d3c6180085e451fb/href</a></iframe><p>Yayyyy! Now you have to genrated the tonified image you want.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/322/1*GWzkmfnYheBbuTNJ7w1o3w.png" /><figcaption>Image1: Input Image, Image2: Output Image</figcaption></figure><h3><strong>Lets now animate this your cartoon character</strong></h3><p>First of now, let’s clone the repo on the first-order model.</p><pre>!git clone <a href="https://github.com/AliaksandrSiarohin/first-order-model">https://github.com/AliaksandrSiarohin/first-order-model</a></pre><p>After that we will change the directory to first-order-model folder.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/bdc813d2ef20fb5134c595e1d975b04c/href">https://medium.com/media/bdc813d2ef20fb5134c595e1d975b04c/href</a></iframe><p>Then after this we will download the pre-trained weights for this. Note the file is quite big ~700Mb.</p><pre>!wget &quot;https://drive.google.com/uc?export=download&amp;id=1jmcn19-c3p8mf39aYNXUhdMqzqDYZhQ_&quot; -O vox-cpk.pth.tar</pre><p>Loding the model now</p><pre>from demo import load_checkpointsgenerator, kp_detector = load_checkpoints(config_path=&#39;first-order-model/config/vox-256.yaml&#39;,checkpoint_path=&#39;vox-cpk.pth.tar&#39;)</pre><p>Now what we will do is, download the video from where the our output image of cartoon character will will imiate the actions.</p><pre>!wget https://drive.google.com/uc?id=1LjDoFmeP0hZQSsUmnou0UbQJJzQ8rMLR -O src_video.mp4</pre><p>After all this finally we can genrate our output animated video of genrated cartoon character.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6b112f4d5d9317c867fd9f7a924eb72a/href">https://medium.com/media/6b112f4d5d9317c867fd9f7a924eb72a/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/502/0*RwVpvRinW13kCHDT.gif" /><figcaption>Final genrated output animation</figcaption></figure><p>Congratulations!!!<br>After all this learning and implimentation you have created a cartoon character of your own.</p><h3>What’s Next and what else can we do?</h3><p>There are still a lot of things we can still experiment on. What if we use some other other models such as the painting’s model to recreate a painting like Mona Lisa or we can also reverse blend the cartoon character to genrate a real face. We can also try Deepfake techniques and swap the faces with some cartoon characters.</p><p>Here I am lisitng some more project ideas which can be done using GAN:</p><ul><li>Emoyion Genration on Faces</li><li>Generate Photographs of Human Faces</li><li>Semantic-Image-to-Photo Translation</li><li>Face Frontal View Generation</li><li>Generate New Human Poses</li><li>Photos to Emojis</li><li>Face Aging</li><li>Super Resolution</li><li>Photo Inpainting</li><li>Clothing Translation</li><li>3D Object Generation</li></ul><p><strong>Thanks for reading my blog. Do give it a clap if you liked it!</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eee5f2da76ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>