Ankur Gupta

Artificial Intelligence, H1B Visa & The Test Of Accountability  I have been thinking about the brouhaha over the H-1B visa and believe a deeper look at Indian IT is in order. The sector’s under fire on two fronts:  →Due to the hike in H1B visa fees to $100,000. →From the disruption by artificial intelligence. Indians make up 70% of H1B visa recipients, with China trailing behind at 12%. The new fees surpass 60% of Indian workers’ annual salaries in the US, as they draw lower median wages, hitting us more than global peers. Indian IT companies could also see a drop in profit margin.  Besides that, AI is threatening to eliminate entry-level jobs. OpenAI CEO Sam Altman said that he's confident customer service jobs will be among the first to go, along with those who do routine coding tasks. It’s disheartening that highly skilled professionals — from developers to creators — who help build the system are under threat. They are stuck in bloated teams and face delays as they await endless approvals. While they bring value, creativity, and problem-solving skills — essential to enable differentiation in products and services — they are struggling to prove their worth. The service-model mindset is partly to blame. The role of the 'knowledge worker' has been diluted, from someone who applied expertise in solving complex problems and innovating to merely executing tasks. By including people in the decision-making chain who add no value but facilitate communication, we end up hindering agility. Corporate organisations are filled with pointless mid-management roles that reduce accountability. It’s possibly the result of an experiment gone wrong or an inability to fix the basics. Anthropologist David Graeber used the term "bullshit jobs" to describe such redundant positions in his 2018 book. In the era of AI, when information can be aggregated and distributed without intervention, it’s a wonder why we retain such roles. The lack of accountability that drives the current top-down system has to be fixed. So far, the skilled went abroad to become system builders, while India was left with mid-level staff. Now, we need to reframe India as a launchpad. Educational institutions, startups, and R&D labs must turn co-creators and lead skilled talent to use AI as a catalyst for creation. We need to set clear expectations for tech workers, provide resources and autonomy, give feedback, and establish a culture of ownership. The NEP 2020, which aims to integrate application-based learning within mainstream education, could prove to be a game-changer for early years, but companies also need to create pathways and opportunities for innovation at home. It’s only through such interventions that we can change the mindset that the only path to success lies abroad. So, yes, we do need to rethink some roles. But it’s not that of a highly skilled tech professional. #ArtificialIntelligence #H1BVisa #Software #IT #Technology

49

2 Comments

Keeping candidates on hold without closure is unfair. If someone invests time in your process, they deserve at least a clear yes or no. Basic communication is not optional — it’s professional courtesy.

20

1 Comment

Prakhar's HLD for Dropbox's files and folders sharing Prakhar sharing a folder 1. Prakhar right-clicks his "Project Alpha" folder and invites his colleague, Simran, granting her "edit" access     2. His client sends a request to the Share Service to define rules     3. Creating the Share 1. Share Service creates a unique, shared record with File/folder identifier, Access level (view/edit), Recipient email (Simran), Expiration settings (if any) 2. It attaches an Access Control List (ACL) to shared record, marking Prakhar as the owner and adding Simran as an editor 4. Creating a "Link" 1. Share Service contacts with Metadata Service to place a special pointer, like a shortcut or a symbolic link, inside Simran's file directory 2. Pointer points to the shared record     5. Invitation: A notification is sent to Simran. She clicks "Accept" Simran seeing the Folder 1. No Data Duplication: When Simran accepts, we don't copy all of Prakhar's files to her account     2. Instant Appearance 1. The next time Simran's device syncs, it fetches her file list 2. Metadata Service sees the pointer and seamlessly includes the contents of "Project Alpha" 3. The folder appears instantly on her device, with zero file data downloaded yet 🤔 over to you, what are benefits of "mount point" approach? Let me know in the comments! 👇 📣 Join 7-day email course to effectively optimize 60-minute system design interview - https://lnkd.in/gDBzCnE3 

48

16 Comments

Urgency invented today becomes catastrophe tomorrow. 🚨 Most pressure in modern IT companies is synthetic. • Deadlines pulled from thin air. • Red banners. • Slack pings marked “URGENT”. Here is the part nobody likes to admit 👇 Fake urgency trains teams to: • cut corners ✂️ • silence doubt 🤐 • skip thinking 🧠 • push risk downstream 🔥 Nothing explodes immediately. That is why it feels safe. Until one day it does. Real urgency looks boring: • clear priorities • small batches • production feedback • calm, repeatable execution Panic is not speed. Noise is not progress. Pressure is not leadership. If everything is urgent, then nothing is under control. ⏱️ Build systems that absorb pressure. 🏗️ Remove noise before adding speed. 🔥 Respect reality, or reality will collect interest. #Leadership #Delivery #SystemsThinking #EngineeringReality #Prioritisation #NoDrama

6

2 Comments

Go Performance Tuning at Scale: Zepto’s pprof Journey (Part 1) At Zepto, performance isn’t just a priority — it’s an obsession. The best part about making systems more efficient is that it not only boosts speed but also drives real cost savings. In this blog, Akash Agarwal shares our journey of performance tuning — the challenges we tackled, the wins we unlocked, and how the process didn’t just improve our systems, but also shaped the way our engineers think about building scalable, efficient systems. Blog link: https://lnkd.in/gtzdtmGg #ZeptoTechXpress #PerformanceTuning Shivanshu Kumar Tiwari Ashish Bhatt Sumit Gupta Aditya Haridas Menon Harshal Khodifad Abhinav Nimesh Uday Solanki Devansh Chaudhary Devansh Goswami Sumit Ojha Vivek Gover Prashant Giri Balram Vishwakarma Shubham Prashar Anoop Singh Mayank Agarwal Nikhil Mittal

124

https://lnkd.in/gxbvpkgt The HIRE Act might initially disrupt India's IT sector, but could it actually strengthen India's position in the long run? 🤔 Short-term: Yes, the 25% excise tax will create challenges for outsourcing arrangements and may reduce immediate demand. Long-term potential benefits: • US companies may convert subsidiaries to branch offices, creating direct employment opportunities • Higher-value work could shift to India as companies justify their investments beyond cost arbitrage • Indian operations might evolve from cost centers to innovation hubs and regional headquarters • Direct employment by US companies could mean better compensation standards and career advancement paths • Companies may consolidate operations into fewer, larger centers of excellence in India Instead of just competing on cost, this could push the India-US partnership toward more strategic, high-value collaboration. Sometimes regulatory pressure creates the push needed for transformation. What's your take? Could this be the catalyst that elevates India's role in global business operations? #HIREACT #IndiaIT #GlobalBusiness #FutureOfWork

2

🚀 Kafka 4.2 Just Broke Its Own Golden Rule ❌ One Partition = One Active Consumer (in a Group ID) - Those days are gone. For over a decade, we've designed systems around one immutable truth:  one partition = one consumer. Until now. 👇 Kafka 4.2 changes everything with new enhancement: ✅ Multiple consumers can now read from the SAME partition simultaneously ✅ Yes, you read that right — one partition, multiple consumers ✅ No more complex partitioning strategies just for workload distribution I created a deep-dive YouTube video covering: 🔹 Explained Share Groups vs Classic Consumers 🔹 Practical live demos 🔹 DLQ In-build functionality 🔹 System design implications you need to know 🎥 Watch here: https://lnkd.in/g7fyF8fs Question for you: Do you think this reduces the importance of partition-based scaling? Or does it open new design possibilities? Drop your thoughts below! 👇 #Kafka #ApacheKafka #Kafka42 #SystemDesign #SoftwareEngineering #EventDrivenArchitecture #DeadLetterQueue #Microservices #Java #DevOps

32

1 Comment

System Design Part 2 Continuing my System Design Blog Series, this post is a deep dive into the Low-Level Design of WhatsApp. 📱 From handling 2 billion daily messages to ensuring high availability, we’ll walk step by step through how WhatsApp is architected at scale. 🚀 High-level estimate Here is our high-level estimate: TypeEstimateDaily active users: (DAU)50 million Requests per second (RPS) : 24K/s Storage (per day) : ~10.2 TB Storage (10 years) : ~38 PB Bandwidth : ~120 MB/s

6

Idempotency: the most underrated concept in distributed systems Retries are unavoidable in distributed systems. Timeouts and network failures mean the same request can be executed more than once. Any system that assumes exactly once execution is fragile. Consider a payment API. A client charges ₹1,000. The payment succeeds, but the response is lost. The client retries. Without idempotency, the user is charged twice. State transitions must be replay-safe. To prevent this, idempotency must be enforced before the state change. Each request includes a client-generated idempotency key that represents intent (for example, order_id + attempt_id). The service stores this key with a unique constraint and processes the request in a transaction. On retries, it returns the stored result instead of re-executing the charge. A payment should move from CREATED to SUCCESS only once; retries should read the existing state and return it. Most production issues do not come from failures, but from retries without idempotency. #tech #engineering #idempotency #software #transaction

8

🏠 It’s 2025 — and buying a home with white money still feels like a challenge. Over the past few weeks, I’ve been exploring home options in Gurgaon — and what’s truly disturbing is how openly many sellers, including small developers and investors, ask for 30–40% payment in cash. We’re not talking small amounts — often over ₹1 crore — discussed casually, without hesitation or fear. Unaccounted cash doesn’t just hurt individual buyers; it distorts the entire housing ecosystem. It inflates prices, reduces formal credit flow, and erodes trust in developers. Even transparent builders find it hard to compete with cash-driven pricing. A recent report highlighted how even professionals with 100% white income struggle to buy homes without facing pressure for off-record payments. I personally don’t want to support this practice. I earn my salary in white, taxes are deducted at source, and every rupee is accounted for. Converting that into untraceable cash just doesn’t feel right — financially or ethically. This isn’t a Gurgaon problem; it’s a national issue. To fix it, we need policy and market alignment: ✅ Revise circle rates regularly to match actual market prices ✅ Align capital gains taxation to realistic valuations ✅ Incentivize full-disclosure transactions (lower stamp duties or simplified digital paperwork) ✅ Mandate digital payments beyond certain thresholds More than a decade ago, the Ministry of Finance identified real estate as a high-risk sector for unaccounted money. Today, with Aadhaar, PAN, and digital registries in place, the infrastructure for transparency already exists. What’s missing is consistent enforcement and the right incentives. Transparency shouldn’t be the exception — it should be the baseline. #RealEstate #Transparency #Housing #Gurgaon #India #Policy #BlackMoney #Governance #FinancialIntegrity #UrbanIndia

109

Built my first MCP server. It finds offline businesses using Serper and sends the data to Clay via a webhook. After 6 months of deliberation. And multiple discussions with Anirudh and Romit, where I asked them 10000 questions, I finally built it. Why now? Because I feel it has become important enough that businesses can’t afford to ignore it anymore. Especially automation agencies like us. Rapid adoption from all three big LLM providers Anthropic, OpenAI and Google. Plus platforms like Zapier, Postman, Pipedream have made it possible to (i) easily create servers and/or (ii) integrate it into your workflows for a growing number of use cases everyday. What does my MCP server do? 1) It automatically creates the right queries to search on google using Serper. Selects the cities I want to target based on the states I mention. 2) Allows for batch creation and modularity for the number of pages I want to search. If I am looking for hotels I'll search for 20 pages vs searching for physical therapists where 5 pages should suffice. 3) Stores the API responses into a local folder as a csv file. Each run has a dedicated folder so I can always go back and access the data if I want. 4) Extracts the business details from the JSON responses and creates the final csv I can import to Clay. 5) Then it gets interesting because it also asks me if I'd like to directly send this data to a webhook table in Clay. This means even the manual import step is not necessary. For now at-least it seems like there are a few gaps: 1) Context window is a challenge, with Claude’s desktop client at-least. 2) Tool selection is not perfect. Both these things will drastically improve in the coming months. My assumption also is that using a client like Cursor will make things better, need to test. The big takeaway I had was the benefits of building your own server. I thought earlier, if out of the box MCP servers will become more readily available then users won’t need to spend as much time building, configuring or implementing their own servers. However, it turns out building your own server gives you the advantage of capturing your business logic more reliably. For example even if there are out of the box servers available for Serper (tongue twister much), I wouldn’t be able to achieve some parts of my workflow with it. Like my querying, batching logic, or the ability to transfer data via webhook to Clay. I could have potentially integrated other MCPs or tools to achieve the same but as I mentioned tool selection is currently a problem. There may be even better approaches to solve for a multi step workflow which captures your business context and needs - than just directly building that into your MCP server. Need to explore and build more.

75

20 Comments

If ONDC made sense, Agentic commerce absolutely will kill it. Stripe launched Agentic commerce protocol for US, while the infra and market is ready in India. An opportunity for Indian Fintech space. The whole value ecosystem of marketplaces to connect buyer and Seller via Search, Algorithms and recommendation has become a commodity now. Dehlivery can deliver and so can multiple other delivery startups. Businesses can get easy storage. Agents(chatgpt) can create customised checkout experience for buyers on the go, with integration with Stripe. Discovery is easy with Gemini, Gpt, perplexity etc, as a user I never have to visit Amazon again. As there is not much Amazon would add value to a user. ONDC if can create a meta-agent for businesses, Clients like GPt and other can simply place order and checkout in seconds without any API or marketplace. The missing part is trust for Returns, Quality, but that was still always the case before eCommerce, if small businesses can create solid brands from customer experience perspective, they don't need to spend 15% commission to marketplace. Local commerce and quick commerce can bridge the gap, having phyiscal presence at convenient distance brings trust. https://lnkd.in/gnUXPjzw

10

Rapido – Please Explain This Logic This screenshot clearly shows “Bad Performance – 1 Order Cancelled.” But the real question is: What if the customer cancelled the ride? If a customer books a ride and then: • Does not arrive at the pickup point • Switches off their phone • Or cancels the ride themselves Why is the captain still getting “Bad Performance”? The captain spends time, fuel, and effort to reach the pickup location. If the customer cancels or becomes unreachable, how is that the captain’s mistake? Yet the system still shows it as captain cancellation, which directly affects: • Order allocation • Earnings • Captain performance rating • Motivation of captains who are honestly working This is not fair for captains. Simple logic: If the customer cancels → it should be counted as customer cancellation. If the captain cancels → then it should affect captain performance. We hope Rapido will review this system and fix it, because fairness is important for the people who keep the platform running. #Rapido #FairSystem #CaptainVoice #RespectCaptains #GigWorkers #uberdrivers #uber

1

Mistake I see often: directly porting REST APIs into MCP. REST APIs were built for humans, flexible, verbose, and atomic. But for AI agents, that richness quickly turns into noise. Too many tools → context overload Atomic calls → extra token cost & latency Complex params → confusion & hallucination The better approach? Curate, don’t copy. Auto-generate only for prototyping. Rename, simplify & hide irrelevant params Start with the agent’s goal, not the API spec One design principle I keep coming back to: “Design agent-first tools. It makes agents faster, cheaper, and more reliable.” The blog from Jeremiah Lowin was such a refreshing one on this topic. Highly recommended read. #API #AI #Agents #LLM #MCP

34

4 Comments

I came across a fascinating article written by famous Gaurav Sen that breaks down how Netflix tackled one of the toughest problems in large-scale systems — connection churn and uniform traffic distribution using an approach called Backend Subsetting, built on top of consistent hashing. At hyperscale, traditional consistent hashing and random assignment aren’t enough. If you hash every request to all backend servers, adding or removing a server can force millions of connection reassignments — leading to instability and performance spikes. Netflix’s solution was elegant and practical: 👉 Subsetting: Instead of spreading every request over the entire pool of servers, traffic is divided into multiple subsets (or shards). Each client is mapped to a specific subset, and load balancing happens within that smaller group. Why this matters: - Reduced connection churn by ~10× → far fewer open connections and less resource waste on the API gateway. LinkedIn - When a server is added or removed, only that subset’s traffic is affected — not the whole system. LinkedIn - Upgrades and failures become predictable and boring events, not outages. Bottom line: Load balancing in distributed systems isn’t just about fairness — it’s about predictable scaling, minimizing churn, and isolating failure blast radius. The “backend subsetting” idea helps isolate impact and keeps traffic stable even during dynamic changes. If you’re building highly available systems, this is a neat pattern to understand — and an even nicer reminder that sometimes the best algorithm is the one that plays well with real-world failure modes. Here is the link of the article for interested folks ! https://lnkd.in/gxJTvdYP #DistributedSystems #SystemDesign #LoadBalancing #Scalability #NetflixTech #Architecture #TechLeadership

5