Posts

A year ago, the conventional wisdom was that open source LLMs were 12-18 months behind frontier closed source models. The argument was structural: OpenAI, Anthropic, and Google had compute budgets, proprietary data, and research talent that academic labs and independent efforts could not match. That argument has eroded significantly. Not because the resource gap disappeared - it did not - but because the techniques that produce frontier models have been published, replicated, and improved upon faster than anyone predicted. ...

Most engineers wait for promotions to happen to them. Here is how to make them happen for you - without politics or brown-nosing.

The real-time web has two main options for server-to-client data delivery: WebSockets and Server-Sent Events. They are not interchangeable, and the choice between them shapes your infrastructure, your scaling model, and your operational complexity for years. The conventional wisdom - “use WebSockets for bidirectional, SSE for unidirectional” - is directionally correct but misses important nuance. Here is the complete picture. The Core Technical Difference WebSockets establish a full-duplex, persistent TCP connection. After the initial HTTP upgrade handshake, the connection is no longer HTTP. Both client and server can send messages independently at any time. It is a raw message channel. ...

Most companies running Kubernetes are overpaying. Not by a little - by 30-60% in many cases. The reason is almost never that Kubernetes is inherently expensive. It is that the default configurations prioritize availability over efficiency, and no one went back and tuned them. Here are the tactics that actually move the needle. The Root Cause: Overprovisioning at Every Layer Kubernetes costs compound because overprovisioning happens at multiple levels simultaneously: ...

The Stack Overflow Developer Survey data from 2024 and 2025 tells an interesting story. VS Code maintains its position as the most popular editor overall. But Neovim’s usage numbers have grown every year since 2020, and the growth is concentrated among engineers with more than five years of experience. This is not nostalgia for old tools. Senior engineers are switching from VS Code to Neovim because Neovim in 2025 is a genuinely different product than Vim in 2010. It is worth understanding what changed and why experienced engineers find it compelling. ...

Stripe launched in 2011 into a market where accepting payments online was genuinely miserable. PayPal had a SOAP API with XML response formats. Braintree required a phone call to get started. Banks had enterprise integration requirements that took months to navigate. Stripe shipped a REST API that a developer could integrate in an afternoon. That was the product. The payments were secondary. Fourteen years later, Stripe’s API is still the benchmark against which every developer-facing API is measured. Here is what they actually built and why it works. ...

The credit card bill arrives and you pay the minimum amount due. The bank is happy. You feel like you handled it. What you have actually done is signed up for one of the highest interest rates available in any legal financial product in India. This is how revolving credit destroys wealth, and it happens to financially literate people too. What Revolving Credit Actually Is A credit card allows you to borrow money every month. If you pay the full statement balance by the due date, you pay zero interest. If you pay anything less - even Rs. 1 less than the full amount - you are in revolving credit territory. ...

Most developers treat the PostgreSQL query planner as a black box. They write a query, it either runs fast or it does not, and if it does not they add an index and hope for the best. Senior engineers treat the planner as a collaborator. They understand what information it uses to make decisions, how to give it better information, and how to override it when it gets things wrong. ...

The most common mistake in microservice architecture is not the choice to use microservices. It is where you drew the lines. Teams decompose along the wrong axis, end up with services that are constantly chatting with each other, and then wonder why their “microservices” have worse latency and higher operational complexity than their old monolith with none of the benefits. The problem is almost always boundary placement. The Signs Your Boundaries Are Wrong You do not need to audit your entire architecture to find bad boundaries. The symptoms are obvious once you know what to look for: ...

When DeepSeek released R1 in early 2025, it was a signal that the compute-cost assumptions underlying the entire LLM industry were wrong. When R2 followed, it stopped being a signal and started being a structural fact. The question is not whether DeepSeek changed the competitive landscape. It clearly did. The question is what that change actually means for developers, enterprises, and the companies building frontier models. What DeepSeek R2 Actually Is R2 is a reasoning model - the same category as OpenAI’s o3 and Google’s Gemini Thinking series. These models do not just predict the next token. They run extended internal reasoning chains before producing an answer, which significantly improves performance on problems that require multi-step logic. ...

React Native has a reputation problem. Ask any iOS developer and they will tell you it will never match native performance. Ask a Flutter developer and they will say the same thing. Ask someone who last used React Native in 2019 and they will have horror stories about jank and bridge overhead. They are all describing a different product than what ships today. What Changed The old React Native architecture had a fundamental problem: every call between JavaScript and native code crossed an asynchronous bridge. Serialization, deserialization, async queue. That was the tax on every interaction. ...

Most people start a SIP, set it at Rs. 5,000 or Rs. 10,000, and forget about it for 20 years. This is better than nothing. But there is a simple modification to this plan that could add Rs. 50-80 lakh to your corpus over the same period: the step-up SIP. The math is compelling. Here is why it works and how to implement it. What Is a Step-Up SIP? A step-up SIP (also called top-up SIP or increasing SIP) is a systematic investment plan where you increase your monthly contribution by a fixed percentage every year. Most mutual fund platforms allow this - you set the initial amount, the annual step-up percentage, and the maximum amount. ...

Python’s async/await support has been production-ready since Python 3.7. Five years later, async code is common, but well-written async code is not. The patterns that look reasonable in tutorials often fall apart under production load. Here are the ones that actually work. The Event Loop Is Not a Thread Pool The first mistake most people make is treating async Python like it solves all concurrency problems. It does not. async/await is cooperative multitasking. Your code yields control at await points. If any await call blocks the event loop - a blocking database driver, a time.sleep, a CPU-heavy computation - everything else waits. ...

Every engineer who has cracked FAANG interviews will tell you the same thing: the problems are not random. There are patterns. Once you see them, you stop dreading interviews and start recognizing familiar shapes in unfamiliar problems. Here is the honest truth - you do not need to solve 800 problems. You need to deeply understand about 10 patterns and practice 5-8 problems per pattern. That is around 60-80 problems done right, not 500 done poorly. ...

Most iOS releases are incremental on the user side and frustrating on the developer side - new APIs that take a year to actually adopt, deprecated things that still require support, and Xcode bugs that somehow ship. iOS 19 is different. Several changes that the developer community has wanted for years landed together, and they compound. Swift 6 Concurrency Is Finally Usable Swift 6 shipped with strict concurrency checking - Sendable requirements, actor isolation, and no implicit data races. This was the right direction but made migration from older codebases genuinely painful. iOS 19 brings several improvements: ...

Paying the minimum due feels safe but it is one of the most expensive financial mistakes you can make. Here is the math that banks do not show you.

Senior engineers with 8-15 years of experience are learning Go as a second or third language in 2026. The reasons are pragmatic and the pattern is consistent.

Most home loan borrowers pay a higher interest rate than they need to. Banks count on inertia - the fact that you will not bother to renegotiate even when the math clearly favors it. Here is how to break that inertia and actually reduce your rate. Why Your Rate May Be Higher Than Current Offers Home loan rates in India are benchmarked to either MCLR (Marginal Cost of Funds Based Lending Rate) or RLLR/EBLR (External Benchmark Lending Rate, linked to RBI’s repo rate). Here is the problem: ...

ULIPs have been reformed, rebranded, and re-pitched. The math still does not work for most buyers. Here is why the product persists and why it keeps underperforming.

You have an education loan at 10-12% and Rs. 50,000 spare cash. Should you prepay the loan or invest in mutual funds? The answer is not as obvious as it seems.