Archives

A custom Android Lint rule takes an afternoon. Rolling it out to ten teams takes an owner, tests, a stable message, a baseline ratchet and a kill switch.

Stop asking which Android linter is best. Ask what each team rule needs — types, a quick fix, resources, KMP — and pick the tool that can enforce it.

Compiling isn't enough to protect a Flutter/Kotlin/Swift/Rust E2EE module's architecture. Five CI gates that catch boundary breaks before they ship.

A headless two-device harness for E2EE chat: real keys, real store, real backend, no screen: reproducing exhausted one-time keys and a lost session on demand.

JUnit, Espresso, XCUITest, Maestro, Appium: a map of mobile test tools by the boundary they verify and the realism they need, with what Google and Apple say.

A green build and a rejected QA build answer different questions. One verification contract, invariants to impact scope, serves agent, reviewer, and QC alike.

Coding agents got strong; the bottleneck moved from the prompt to the workflow around it. Eight practices big tech converged on, mapped onto one 7-stage loop.

Three tool migrations, zero workflow changes. What METR, DORA 2025, and my E2EE work taught me about using AI vs. redesigning around it.

MethodChannel has a real cost model, a fake mythology, and a serious rival. What the ecosystem actually ships today, which calls deserve a Dart→FFI→JNI fast path — and the memory, threading, and lifetime rules you sign up for when you take it.

Realm crashes at init inside the iOS Notification Service Extension's 24 MB budget. The real causes — schema discovery, an encrypted page cache, cross-process sharing Realm only added in 2023 — and what Signal does instead.

A 45 MB video, a moving taxi, a connection that dies at 93% — media upload is five problems wearing one trench coat. The full map: resumable chunks, optimistic UX, dedup, async pipelines, and the resilience layer nobody blogs about.

Why log-grepping wastes your debugging time in Flutter — and the tools (DevTools, Crashlytics, the official MCP server) that answer each question better.

Grab runs 1,000+ modules in one app, Reddit's Android app ~800. The rules that make that survivable — contracts over imports, CI-enforced graphs — and why a Flutter module must hide behind a native boundary.

beginBackgroundTask, BGTaskScheduler, URLSession, silent push, PushKit, NSE — a decision guide to iOS background execution, tested against an E2EE messaging core.

A background queue is not background execution. Building the iOS lifecycle mental model — foreground, background, suspended — from an E2EE core that persists sessions to SQLCipher.

UniFFI, cbindgen, or hand-written FFI? Choosing the binding layer that lets Kotlin and Swift call a Rust library — with Firefox, Element X, and Signal as evidence.

E2EE and Feature X both need SQLCipher. The app links both, the notification extension links one. Counting the copies — in the graph, the bundle, and at runtime — is three different questions.

An iOS app and its Notification Service Extension are two executables. What sharing one Rust library really takes: ABI, packaging, linking, and shared state.

A FlutterEngine has one seat. Why one cached engine can't serve two Flutter segments split by native, what FlutterBoost really does, and what two engines cost.

Migrating Flutter to native screen by screen leaves two back stacks. A destination contract, a host router, one stack owner — and why back is the hard part.

A Flutter package needs the host app's auth token. Every obvious answer — a MethodChannel, importing AppCore — points the dependency the wrong way. The fix is to stop passing the token.

WebSocket, MQTT, SSE or push? The question sounds like a protocol benchmark. Messenger, Discord, Uber and Grab's own engineering posts say the channel is the smallest part of the answer.

Key-value, SQLite, Realm, files, secure storage — the choice looks like a benchmark question. The apps that store millions of messages say it's an architecture question.

Telegram clients don't fetch history to render a chat — they render what they already know. A source-verified tour of pts, getDifference, known ranges, and the sync engine behind instant-feeling chats.

One tap on Send, four very different journeys. What E2EE actually changes in WhatsApp, Signal, Messenger and Telegram — and what it can't protect.

My messages loaded in double-digit milliseconds and scrolling still stuttered. A frame-pipeline deep dive, with measured numbers from a real chat workload.

Messenger deleted 84% of its code to get fast. Telegram serves hundreds of millions from one module. Module graphs I extracted from three chat apps show which layers actually pay rent.

Snapchat was slow in native. Discord is fast on React Native. Evidence from a dozen engineering teams — plus my own A/B experiment — on what actually makes apps slow.

How to choose a tech stack for any website: a simple decision framework with tables, a decision matrix, and a decision tree — from blog to e-commerce.