Dawson Xiong

Hey, I’m Dawson.

I’m studying computer science at the University of Waterloo and previously worked full-time at General Learning (YC F24).

I specialize in backend systems and UI/UX for web and mobile apps. Now, my free time goes to machine learning, mostly inference and reinforcement learning.

If I’m not at the keyboard, you might find me at the gym or the poker table.

Email GitHub LinkedIn X/Twitter

Beacon
Multi-phone person search · Hack the North 2026 finalist
Turns everyone’s phone into a camera network to find a missing person
Swift, Python, JavaScript, SwiftUI, FastAPI, Expo, ARKit, Three.js, YOLOE
More
- Searchers scan a QR code to join, then a Swift app streams each phone’s camera frames and ARKit position over WebSockets to a FastAPI hub
- A Baseten GPU service spots people with YOLOE and matches them to reference photos with OSNet in about 127 ms
- VGGT rebuilds the room in 3D from the phones’ frames, and the console guides searchers with a live heatmap, arrows, flashes and haptics
- In simulation, a PPO policy rescued 47.6% more people than greedy search
Stem Player
Split any song into stems and mix them on a 3D device
Use Kanye's Stem Player right in the browser. Splits any song into its stems.
TypeScript, Python, vinext, React Three Fiber, Web Audio, Cloudflare Workers, Modal, Demucs
More
- Upload a song or paste a link and the file goes straight to Cloudflare R2, then a Modal GPU job splits it into four stems with Demucs
- The stems come back through a webhook and play in sync with Web Audio, and the sliders on the 3D model turn each one up or down
- It also has echo, reverb and filter effects, a speed control and a mix recorder
- Cut a split from a minute to under 8 s, with sound in 1.5 s while it separates
ConvertKit
Local macOS file converter
A performance-first Rust macOS app with 22 file conversion tools, from video transcription to PDF splitting.
Rust, TypeScript, Tauri, React, Tokio, Tailwind, FFmpeg, whisper.cpp
Download for macOS
More
- The React UI sends typed jobs to Rust over Tauri, which runs FFmpeg, ImageMagick, Pandoc or whisper.cpp and streams progress back
- If only the container changes, like MP4 to MOV, ffprobe catches it and the file is repackaged instead of re-encoded, which is near-instant
- Each Tokio job has its own cancellation token, so encodes stop cleanly
- It handles 29 input formats and 202 conversions, all offline
Thock
Typing test for the terminal
MonkeyType in the terminal, with blazing fast performance built using Go.
Go, Bubble Tea, Cobra
go install github.com/dawsonxiong/thock@latest
More
- The whole app is one Bubble Tea model that runs every screen at 120 FPS
- It uses Monkeytype’s scoring, so fixed typos still count against accuracy
- Race rooms announce themselves on your network and show everyone’s cursor live, and each room’s two-word code also spells its address
- Each keystroke and redraw takes about 4 µs, 2,000× under a 120 FPS frame
react-native-latex-renderer
Native LaTeX rendering for React Native
An npm package that renders LaTeX in React Native as a native SVG, with no WebView. Peak 143 weekly downloads.
TypeScript, React Native, MathJax, react-native-svg
More
- MathJax turns LaTeX into SVG paths and react-native-svg draws them natively, so it works offline and renders instantly
- Inline and display math mix with plain text, with a KaTeX WebView fallback
Pulse
Developer news on every new tab
A Chrome extension that replaces the new tab with a high-signal developer news feed. Think daily.dev but cleaner and more relevant.
TypeScript, Next.js, React, WXT, Tailwind, Prisma, Postgres, Redis
More
- A daily Vercel cron job pulls about 40 RSS feeds from blogs and tech sites
- Headlines that overlap within 48 hours are merged into one card per story
- Stories are ranked by recency, your tags and how trusted the source is
LaTeX.ly
Handwritten math to LaTeX
Turns a photo of a handwritten or printed equation into LaTeX.
Python, JavaScript, Flask, Next.js, PyTorch, OpenCV, Tailwind
More
- OpenCV splits the photo into symbols, rejoining broken ones like = and i
- An 81-class PyTorch CNN reads each 64×64 symbol, left to right, into LaTeX
- I trained five versions of the model on 49k+ symbols: 24k rendered across 25 fonts and 25k handwritten from CROHME
- A Flask API serves the model to a Next.js front end
LinkedIt
Sponsor contact finder · Best Beginner Hack, GeeseHacks 2025
Finds the right people to contact at a company for sponsorship.
Python, JavaScript, Flask, React, Selenium, Vite, Tailwind
More
- Flask drives headless Chrome with Selenium to search LinkedIn
- Matches come back as JSON and show up as profile cards in React