2026
Blindspot, Part 3: $0.0215 a Report, and the Things That Broke Budgets checked between stages, idempotent Instagram DM delivery, and two production failures: a gitignore comment that ignored nothing, and exit code 139 from an architecture mismatch. Blindspot, Part 2: Extraction Sets the Ceiling A weak extraction model found 1 of 7 checkable claims in a Reel. No amount of good research recovers a claim that never entered the request. Four rules about where the quality actually comes from. Blindspot, Part 1: Teaching a Pipeline to Watch the Whole Video The first version processed a three-minute protest Reel successfully and missed the police crackdown at the end. Every frame it looked at came from the opening. Four rules came out of fixing that. Parakhi: Where Does Your ₹5 Actually Go? Parakhi breaks an Indian product's MRP into what stays in India, what goes to tax, and what leaves the country. Every number is computed, sourced, and labelled with how much you should trust it. ScamDB: The App Took a Weekend, The Data Took Three Months I built a search engine for suspicious Indian phone numbers and UPI IDs. The hard part was never the app. It was that the number is always inside a screenshot. Brainrot: The Dumbest Video Format Taught Me the Most About Debugging I automated those Reddit-story-over-Minecraft-parkour videos. The format is junk food. The bugs underneath were genuinely interesting. ReelGen: Stop Scrolling Stock Footage, Start Shipping Reels I got tired of AI reels that look like AI reels. So I built a pipeline that actually cares about color, pacing, and footage quality. The Claude Code Tabs You Forget About I kept losing Claude Code sessions that were waiting on my reply. So I built a tiny local dashboard that surfaces them. Meta-Dev-CLI: Stop Clicking, Start Typing Why I built a CLI for Meta's WhatsApp API to keep developers (and AI agents) in their flow state. Live Kitchen Feeds: Transparency Without the Cloud Bill Why food delivery apps need live streams, and how to build them cheaply with WebRTC. Designing for Scale Scale is mostly about clarity — of data flow, ownership, failure handling, and observability. Event-Driven Systems Are Not Magic Event-driven systems are powerful, but they redistribute complexity rather than eliminating it. Debugging Distributed Systems Debugging distributed systems is fundamentally different. You're reading timelines, not stack traces. Production-Ready AI Adding AI to a system is easy. Making it reliable is not. Why Queues Matter More Than You Think Queues are the unsung heroes of reliable systems. They absorb traffic spikes, decouple services, and turn fragile sync calls into resilient pipelines.2025
Boring Systems Win The most reliable systems are boring. They use proven databases, avoid unnecessary abstraction, and log properly. Data Flow Is Everything Every backend problem eventually becomes a data flow problem.No writing tagged with that yet.