Local by Default: What "Local-First" Actually Means at ZumiLabs
What runs on your computer, what doesn't, and why "by default" matters: a plain account of where your data goes in every ZumiLabs tool.
Deep dives into local-first system design, on-device machine learning, sandboxed components, and high-performance browser tools.
What runs on your computer, what doesn't, and why "by default" matters: a plain account of where your data goes in every ZumiLabs tool.
A guide for product teams: the costs of cloud-only tools that never appear on the invoice, and the questions to ask before your team's files go into one.
Search Console said our site had 0 videos. How we fixed it with VideoObject structured data and video sitemap tags for 243 videos, built from one source.
Two course series, 42 files and 18.8 hours of audio transcribed on one Mac with ZumiScribe's local engine, with speakers, in 13.6 minutes. No upload.
Real timings from 2,444 ZumiForge jobs: how fast local image models run on Apple silicon, where they match the cloud, and where they still fall short.
Most real image work is a set, not a single picture. The design decisions behind ZumiForge's queue: one row per prompt, parallel queues, exact retries and a job record you can trust.
Three of our course thumbnails came back covered in fake letters, even though every prompt said "no text". Why it happens, and the prompting changes that stopped it.
A new homepage button had white text on bright blue at 3.07:1, below WCAG's 4.5:1. How we caught it, how the ratio is worked out, and the one-line fix.
Hick's law is quoted to justify short menus and fewer options. What the 1952 experiments measured, where the formula applies, and what to do about choice instead.
Our header hid every link behind a burger below 1,100px. How we kept Tools and Learn visible on phones, quietened the call to action, and tested every width.
Why live captions and transcripts matter in meetings and lectures, not just recordings, and the design choices behind captions people can actually read.
One meeting recording through ten speech-to-text services: the words were similar, but speakers, short replies, punctuation and timing were not.
Capturing slides automatically during a meeting and placing each picture next to what was being said, plus phone photos, exports and privacy.
Speech engines hear words well but lose track of who is speaking. How we designed speaker labels that people can fix while recording and afterwards.
What we learned running speech models in the browser so audio never leaves the computer: GPU checks, workers, memory, languages and long files.
Why a browser-based transcriber needed a small helper app on the Mac and Android, and how one local port, setup checks and permissions made it work.
Why we stopped calling a usage report anonymous, and how we turned privacy claims into plain facts: what is sent, to whom, how often, and where.
Why prompt-and-pray AI produces fragile UI spaghetti, and how engineering a mathematical token compiler transforms autonomous AI into a reliable, manageable, and editable design partner.
How we built a high-density, real-time news monitoring wall in pure React/Vite using non-destructive HLS virtualization, freeze-frame preservation, back-off retries, and serverless edge scrapers.
Creating a photo-real, 60 FPS interactive hero character that tracks cursor gaze across CTAs with 0ms latency, optical interpolation, and mobile gyroscope support — inspired by @alaa.alaff.
Turning a date, time and place of birth into deterministic visual art in the browser — without AI models, without servers, and why sixty-five iterations were spent learning what variation actually means to a human observer.
I built a camera check that reliably tells a real face from a photo, a screen and a mask - then defeated it with 28 lines of arithmetic and no camera at all. The gap between those two facts is the whole story.
A browser-only heart rate monitor that never sends a pixel anywhere. Checking it against a real ECG meant deleting three of its four metrics - and the most useful lesson was that the wrong numbers looked the most convincing.
Rendering Microsoft Visio (.vsdx) diagrams client-side without servers or desktop applications is a local-first dream. Here is our engineering journey from flat XML fallbacks and WebAssembly filesystem crashes to a clean programmatic iframe integration.
Adding RAW photo support to a browser file manager should have taken an afternoon. It took most of a day, six independently reasonable failures deep. Here is each one, and why the symptoms all looked identical.
Snapping what is visible on screen is easy. Stitching a dynamic, infinite-scrolling dashboard down to the pixel-without crashing the browser or losing absolute header positions-is where the real engineering starts.
How do you screenshot 50 open tabs in the background without reloading pages or forcing active focus? Inside OmniCapt's extension-level DevTools Protocol bridge.
Screenshots of web designs are useless if covered in GDPR popups, sticky alerts, and chat bubbles. Learn how OmniCapt sweeps layers out of the frame using dynamic CSS injection rules.
Recording video clips is only half the battle. Bringing multiple recordings together side-by-side or sequentially, entirely locally in Chrome, represents the ultimate developer video utility.
Cloud LLM APIs are fast, but they cost money, add latency, and compromise data privacy. Here is how OmniBrief orchestrates Chrome's built-in Summarizer and LanguageModel APIs (Gemini Nano) for offline research.
Keyword search fails when you forget specific phrasing. Discover how we built client-side dense vector math using ONNX runtime and MV3 offscreen documents for private research indexing.
Web pages are more than prose. Learn how OmniBrief captures Schema.org metadata blocks to render native tabs for TV series cast, recipe details, events, and shopping lists without LLM cost.
Using custom cloud LLM keys shouldn't compromise your local environment. Here is how we implemented PIN-derived PBKDF2 and AES-GCM local storage encryption.
The web browser is no longer a sandboxed viewport with no hard disk options. Explore how ZumiFiles recursively loads, creates, and mutates directory structures using the File System Access API (a WICG standard in Chromium-based browsers).
Need to drop a multi-pane file manager into any web project without breaking style layouts? How ZumiFiles packages React, state channels, and Tailwind CSS styles inside a single custom element.
How do you save custom file descriptions, tags, and workspace notes without a database or backend API? Explore ZumiFiles' atomic sidecar JSON metadata synchronization mechanics.
Processing binary video shouldn't require third-party servers. Discover how ZumiFiles uses the browser's built-in WebCodecs API (via mediabunny) to trim and split videos instantly, entirely on-device.
Arrows, boxes, and blur-outs turn a raw screenshot into a usable report. OmniCapt overlays a canvas drawing surface and flattens the annotations into the final image.
Designers need the same page at desktop, tablet, and mobile widths - captured consistently, not by hand-resizing. OmniCapt drives the DevTools Protocol Emulation domain to do it.
A bug-repro recording is far more useful when viewers can see exactly where you clicked. OmniCapt paints an animated highlight layer synced to real pointer events.
Every summary in OmniBrief can be listened to, not just read - with zero audio files and no cloud TTS bill. The Web Speech API makes it possible, quirks and all.
Coaxing clean, parseable JSON out of a small on-device model is the hard part of any local-AI feature. OmniBrief uses the Prompt API's responseConstraint to make it deterministic.
Manifest V3 service workers can't touch the DOM, run long, or reliably host WebAssembly. OmniBrief runs its heavy work in an offscreen document instead - here's how.
OmniBrief can summarize with on-device Gemini Nano, a local Ollama server, or five cloud APIs - all behind one call. Here is the adapter pattern that makes models hot-swappable.
Comparing two directory trees - what was added, removed, or changed - is a classic desktop task ZumiFiles does entirely client-side, without reading every byte upfront.
Rendering a poster image for every local video usually means an upload or a bundled media engine. ZumiFiles decodes a single frame with the browser's own pipeline instead.
ZumiFiles can send a file straight from one browser to another with no server in the middle. WebRTC data channels make the transfer direct, chunked, and private.
CABE is a Chrome extension whose bookmark set changes automatically based on the site you're on - Jira links in Jira, docs in GitHub. Here's how it tracks context and resolves dynamic links.
Shipping third-party API keys to the browser is a liability. CloudSend is a small Node.js service that holds the credentials and exposes one authenticated endpoint to many targets.
No articles found matching your criteria.