Note on source code. Capy Files is a commercial product by Devinfi OOD. The source is private; this repository is a public case study of what was built and how. Code walkthrough available to recruiters on request: kevinrusev1@gmail.com or LinkedIn.
Capy Files is a toolbox for everyday file chores (shrinking a PDF for an upload form, turning iPhone photos into JPGs, cutting the background out of a product shot, transcribing a lecture), live at capyfiles.com.
Most "free online" tools upload your file to someone's server. Capy Files does all of the work inside your browser, with WebAssembly, WebCodecs and on-device AI models:
- Private by construction. Files, file names and results never leave the device. There is no upload step to secure because there is no upload.
- No waiting on a server. No upload or download time and no queue.
- Works offline. It installs as a PWA; after one visit the whole app, including its engines, works without a connection. (AI models download once, the first time an AI tool is used.)
- No ads, no trackers, no analytics.
- Free, with an optional Pro. Every free tool needs no account and adds no watermark. Pro (weekly, monthly, yearly or a capped lifetime launch deal) unlocks the best AI quality, big batches, the PDF editor and Workflows.
Real screenshots from production (
capyfiles.com), captured with Playwright.
 Home: pick any file, see which tools open it |
 Tool catalogue: 26 PDF tools |
 Compress PDF: preset or exact target size |
 Background removal with an on-device model |
 Whisper transcription in the browser |
 Image conversion: HEIC, JPG, PNG, WebP, AVIF, JPEG XL |
 Pricing: week pass, Pro monthly or yearly, and a capped lifetime launch deal |
|
 Home |
 AI tools |
 Compress PDF |
 Transcribe |
 Pricing |
61 released tools in seven categories (counted from the production tool registry):
| Category | Tools | Examples |
|---|---|---|
| 26 | Compress, merge, split, organise, edit, sign, redact, compare, protect / unlock, PDF/A, scan to PDF | |
| Images | 11 | Compress, convert, HEIC to JPG, resize, crop, remove metadata, image to SVG, photo editor, face blur |
| Video | 7 | Compress, convert, trim, video to GIF, extract audio, mute, screen recorder |
| Audio | 4 | Compress, convert, trim, voice recorder |
| AI | 5 | Background removal, 4× upscaling, OCR, transcription, hide personal details before pasting text into an AI chat |
| Design | 6 | Palette from image, contrast checker, QR codes, colour converter, gradients, device mockups |
| Files | 2 | Zip, unzip |
- AI on the device. Models run through Transformers.js and ONNX Runtime Web on WebGPU (half precision where the GPU
supports it), with an automatic WebAssembly CPU fallback:
- BiRefNet for background removal, in three quality levels (Lite at 512 and 1024 px, full model at 1024 px)
- RealPLKSR for 4× super-resolution
- YuNet for face detection (face blur)
- Whisper (base, small, large-v3 turbo) for transcription
- Tesseract for OCR
- WebAssembly engines. qpdf, MozJPEG and OxiPNG, pdf.js and pdf-lib, and WebCodecs-based video and audio processing. It also includes a custom PDF compressor written to replace AGPL-licensed Ghostscript.
- Edge hosting. Cloudflare Workers serve the app and the API, with D1 as the database and R2 for model files. Large model downloads are resumable: a dropped connection continues from the last byte instead of starting over.
- Offline, installable PWA. A service worker saves the app shell and the everyday engines (image codecs, qpdf, HEIC, AAC) on the first visit; an end-to-end suite runs every tool test with the network switched off.
- SEO without a server renderer. Every public page is pre-rendered to static HTML at build time (titles, canonical links, Open Graph, JSON-LD, sitemap). React takes over in the browser.
- Payments and accounts.
- Stripe Managed Payments (Link as merchant of record, handling VAT worldwide) with an embedded checkout: card, Apple Pay, Google Pay, Link and local methods, styled to the site.
- Signature-verified, idempotent webhooks; plans are delivered as signed entitlements the browser verifies.
- EU consumer law built in: two-box consent before the form exists, one-click withdrawal with a proportionate refund (price minus the days used), renewal reminders, and a Bulgarian-language contract for local buyers.
- Firebase Authentication: Google, Microsoft, GitHub and email link, with Sign in with Apple ready behind a flag.
- Security and quality.
- Strict Content Security Policy, HSTS and a locked-down Permissions-Policy.
- Cross-origin isolation only on pages that need it.
- TypeScript strict with no
any. - 900+ unit tests, and CI runs typecheck, lint, tests and build.
- Playwright end-to-end tests in Chromium and WebKit: every tool, checkout on 14 screen sizes from a 360 px phone to 4K, accessibility checks, and offline runs.
- Performance budget: every page loads under 200 KB of gzipped JavaScript up front; the rest loads on use.
- An automated licence compliance check fails the build on GPL, AGPL or non-commercial dependencies and models.
flowchart LR
subgraph Device["User's browser (all file processing)"]
UI["React 19 + Vite PWA"]
WASM["WebAssembly engines<br/>PDF, image, archive"]
GPU["On-device AI<br/>WebGPU, or WASM on the CPU"]
UI --- WASM
UI --- GPU
end
subgraph Edge["Cloudflare"]
W["Worker<br/>static pages, API, model delivery"]
D1[("D1<br/>accounts and plans")]
R2[("R2<br/>AI model files")]
W --- D1
W --- R2
end
UI <-- "pages and app code" --> W
GPU <-- "model weights, resumable" --> W
UI <-- "sign-in" --> FB["Firebase Auth"]
UI <-- "embedded checkout" --> ST["Stripe<br/>(Link, merchant of record)"]
ST -- "signed webhooks" --> W
Files never travel on any of these arrows. The network carries the app itself, the AI model weights (downloaded once and cached), and, for paying users, account and subscription state. More in docs/architecture.md and docs/on-device-ai.md.
| Area | Technology |
|---|---|
| Frontend | React 19, TypeScript (strict), Vite, Tailwind CSS 4, React Router |
| Processing | WebAssembly (qpdf, MozJPEG, OxiPNG, Tesseract), pdf.js, pdf-lib, WebCodecs, Web Workers |
| AI | Transformers.js, ONNX Runtime Web, WebGPU, WASM SIMD + threads |
| Hosting | Cloudflare Workers, D1, R2, rate limiting, cron triggers |
| Accounts and billing | Firebase Authentication, Stripe Managed Payments (embedded checkout), transactional email |
| Offline | Service worker (Workbox), installable PWA |
| Quality | Vitest, Playwright (Chromium + WebKit), oxlint, GitHub Actions, licence compliance check, JS budget |
I designed, built and launched Capy Files solo: product, visual identity and mascot, every tool, the processing engines, the AI pipeline, hosting and security headers, payments and accounts, legal pages, and SEO. It is sold through my company, Devinfi OOD.
- Running Whisper in a browser tab. The accurate model is hundreds of megabytes. It runs in a Web Worker so the page stays responsive, and the model size is picked for the device (large-v3 turbo on WebGPU, small on the CPU). Lectures are transcribed in 30-second windows using Whisper's long-form method, with silence skipped so the model doesn't invent text. Downloads resume from the last byte after a dropped connection, and the model is cached so the second run starts immediately.
- WebGPU when it's there, the CPU when it isn't. Every model runs on WebGPU (in half precision when the GPU supports it) and falls back to multi-threaded WebAssembly. The fallback is chosen per device and is invisible to the user; the same code path serves a gaming PC and a five-year-old laptop.
- Replacing Ghostscript for licensing reasons. Ghostscript is AGPL, which doesn't fit a closed-source product. I wrote a PDF compressor instead. It walks the document, recompresses images by their effective on-page resolution with MozJPEG, keeps transparency masks intact, removes duplicate images, and slims embedded fonts, then qpdf repacks the file. The result is never bigger than the input. For the same reason, a GPL image quantiser was replaced with my own, and the background-removal model was swapped for an MIT-licensed one.
- Cross-origin isolation only where it pays off. Multi-threaded WebAssembly needs
SharedArrayBuffer, which needs COOP/COEP headers. Those headers also break third-party sign-in and checkout popups. Tool pages are isolated; sign-in, checkout, account and legal pages are not. One shared header policy generates the static headers and the Worker's headers, and tests prove every path gets exactly the policy's headers. Navigating across the boundary is always a full page load. - SEO for a single-page app without a server renderer. All public pages are rendered to static HTML at build time. The app then renders over them, with a guard that hides a cached page shell if the service worker served it for a different URL.
Proprietary. All rights reserved. This repository contains documentation and screenshots only; see LICENSE.