Tell Us What You're Building

Insights

Writing about the systems behind strong digital products.

Notes from building, designing, and shipping on the web, with a bias toward pragmatic decisions over trend-chasing.

The Product Behind the Viral Map

Featured article

Oct 6, 20268 min read

The Product Behind the Viral Map

What Unseen Bangladesh taught me about Product-Content Fit — why products don't always go viral, but the personal artifacts they create do.

Sep 22, 20265 min read

Before Building a Real Estate Voice Agent, I Needed to Understand the Calls

Building an AI voice agent for real estate looked simple until I asked what a good agent is actually trying to accomplish on the call. Here's why "call → AI → showing" is the wrong mental model, and what should replace it.

Rapid Software Development, AI, and the Rising UX Bar

Sep 17, 20264 min read

We Can Build Software Faster Than Ever. But Can Users Actually Use It?

AI and rapid development let teams build software faster than ever. But as competitors move faster and user expectations rise, usability and refinement matter more, not less.

Voice AI Latency Optimization Architecture Cover

Feb 15, 20264 min read

Voice AI Latency Optimization: What a 1,272ms Calendar Check Taught Us

Two different Flash-tier models across two live voice agents, n8n's built-in staticData cache replacing an external Redis layer, and a 5-question intake that was quietly adding two minutes to every call — here's what actually moved the latency numbers.

AI Content Pipeline Safety and Claims Gate Cover

Feb 10, 20264 min read

AI Content Pipeline Safety: The Claims Gate That Stops Fabricated Customer Results

A 4-stage Qwen pipeline generates every post for Minions.AI, and one safety mechanism does the real work: a gate that blocks any AI-generated claim it can't source — including a hard rule against fabricating customer results, because there aren't paying clients to attribute them to yet.

Building LLM Workforce Platforms Cover

Feb 2, 20264 min read

What It Actually Takes to Build a Workforce Management Platform for an AI Company

When GenMorphics AI Solutions needed a platform to coordinate their global team of domain experts, the hard part wasn't the AI — it was compliance, access control, and building tools flexible enough to match how the org actually runs.

Why Browser Agents Fail in Production Without Semantic Layers Cover

Aug 27, 20262 min read

Why Browser Agents Fail in Production Without Semantic Layers

Browser agents do not fail because the model is bad. They fail because modern web applications render complex DOM trees with opaque elements and unstable class names that are hostile to machine parsers.

Conversational Commerce Webhook Architecture Cover

Jan 28, 20264 min read

A Debouncer That Doesn't Scale Past One Worker (On Purpose, For Now)

SellerVai merges rapid-fire customer messages into one reply using an in-memory buffer that explicitly can't run on more than one process. Here's why that's the right call today, and what the code already says about the day it isn't.

$Rendering KaTeX Formulas at Scale Cover$

Jan 20, 20263 min read

Rendering Math Formulas When Your Content Has Two Different Pasts

MathPro Academy's course database holds years of plain-text content next to newly authored Lexical HTML. An earlier attempt at server-rendered LaTeX didn't survive that mix — here's the runtime renderer that replaced it, and why client-side was the right call, not a compromise.

Defensive Webhook Engineering Cover

Jan 14, 20264 min read

What SSLCommerz Actually Gives You for Webhook Security (It's Not a Signature)

Payment webhook handlers fail in ways you only discover in production. For MathPro Academy's SSLCommerz integration, the fix wasn't HMAC signature verification — SSLCommerz doesn't sign its IPN payloads that way. It was an active server-to-server validation query, an amount cross-check, and a fraud score the gateway hands you for free.

Self-Healing Chrome Extensions Cover

Jan 8, 20264 min read

Self-Healing Chrome Extension: Building a DOM Injector That Fights Back

LinkedIn periodically ships builds with every class name replaced by a hashed token, breaking any content script built against fixed selectors overnight. Here's the AI-driven strategy generator we built to regenerate selectors on the fly — and why it eventually lost anyway.

Badhan Donor Intake Parsing Architecture Cover

Jan 2, 20265 min read

AI-Assisted Data Entry: Teaching an LLM to Parse Telegram Donor Messages

Badhan's donor coordinators were already posting donor info into Telegram as free text. Instead of building a form nobody would use consistently, we built two parsers matched to two real workflows — one deterministic, one AI-backed.

Scaling Algorithmic Training Systems Cover

Dec 24, 20255 min read

Streaks Without Race Conditions: What Actually Runs Behind Codervai CP

The 'cohort pacing' story behind Codervai CP's module unlocking was cleaner than what's actually in production — a simple admin-publish flow. Here's what's real: the publish flow, the atomic streak upsert that survives a midnight submission storm, and why we didn't build a custom video pipeline.

Academic Certificate Verification Cover

Dec 18, 20255 min read

How We Made Academic Certificates Verifiable Without a Blockchain (or HMAC)

CPRBD at the University of Dhaka needed employers and embassies to verify that a professional certificate was real. The actual solution was less exotic than the platform's own docs once claimed: a structured certificate ID, a plain database lookup, and rate limiting.

Next.js 16 and Turbopack Deep Dive Cover

Dec 10, 20254 min read

What I Actually Had to Change When Moving to React Server Components

React Server Components aren't just a build optimization — they change how you think about where code runs and why. Here's what the migration looked like on a real project and where the component boundary decisions were non-obvious.

The Craft of High-Velocity Software Delivery Cover

Dec 1, 20255 min read

The Stack I Keep Coming Back To and Why I Stop Reconsidering It

Every few months a new framework or runtime promises to fix problems I don't have. Here's the case for PostgreSQL, TypeScript, and Next.js — and more importantly, the case for stopping the search.