金弘扬Beijing
Hongyang Jin
I build AI products – apps, SaaS, models and open-source tools. This is where they all live.
Projects
15
Model downloads
1.7M
GitHub stars
3.9k
Apps2
Software you download and run on your own computer.
-
Removes background noise and room reverb from speech in audio and video files, and repairs damaged recordings. Everything runs on the Mac's GPU; nothing is uploaded.
-
An English pronunciation coach for Chinese speakers on the Mac. Read a sentence aloud and it shows which word and which sound was off, explains in Chinese how to fix it, and compares your mouth shape, rhythm and intonation. Scoring runs on your Mac; recordings aren't uploaded.
Browser extensions2
Free, open-source Chrome extensions.
-
Pipes a tab's network traffic into your own code: every request, with headers, bodies, cookies and timing, arrives at a server you run as JSON while you browse. No proxy, and no DevTools left open.
-
A web clipper: hides ads and popups, finds the article, and exports clean Markdown (code, tables, math, offline images) or a full-page PNG – all inside the browser.
Self-hosted2
Open-source products you run on your own server.
-
A non-custodial crypto payment gateway. Accept USDT, USDC, BTC, ETH and more across nine chains: every invoice gets a fresh address derived from your public key, payments go straight to your wallet, and signed webhooks tell your backend when they arrive.
-
License keys and file delivery for software sold through Stripe. Its API matches Lemon Squeezy's, so existing apps keep working after you change one base URL, and it imports past orders, keys and files in one go.
SaaS & platforms3
Built, not public yet. Links go to demo videos.
-
Product mockups for brand and design teams. Vision models work out the surface shape and lighting of a product photo, so a logo wraps around curves and folds with real shading – no diffusion model, so the logo stays pixel-exact.
-
Email templates for marketing teams. Upload a screenshot of an email and an AI agent turns it into an editable template: names, amounts and codes become variables, repeated blocks become loops, optional parts get conditions.
-
GPU autoscaling platform
Serves AI models such as speech recognition on pay-as-you-go GPU containers, scaling between zero and N instances behind one gateway, and forecasts GPU availability to start instances before capacity runs short.
Open source6
Code on GitHub and models on Hugging Face, free to use under each project's license.
-
1.7M
downloads on Hugging Face
A Chinese sentence-embedding model for semantic search and retrieval-augmented generation (RAG).
-
Enterprise knowledge-base Q&A you can trace: each answer cites the document, version and passage it comes from, and checks that the rule was in force on the date the question is about. Anything it can't verify is flagged or declined.
-
Turns SRT, WebVTT, ASS, TTML and STL subtitles into a Final Cut Pro project where every line is a Basic Title you can restyle, bilingual layouts included – and exports titles back to SRT. Also runs in the browser.
-
Agent skills: Douyin creator analysis, and Markdown-to-PDF typesetting with 12 style templates.
-
text2vec sentence embeddings running on ONNX Runtime.
-
Turns the Qwen3-TTS realtime WebSocket API into a plain HTTP API, with streaming.
-
Other repositories, experiments and older projects.
About
I'm Hongyang Jin (金弘扬), a developer in Beijing. I've spent about ten years building backend and AI systems – lately mostly agents, RAG and speech recognition. This page collects what I've made: apps, SaaS, models and open-source code.