Portrait of Rick Liu

Hi, I'm Rick.

>

From idea to users, I ship AI products end to end.

“A hundred blows may leave no mark. The hundred and first splits the stone—and every blow mattered.”

My experience

My projects

TermPocket

    Made with:

A finance app for Canadian university students that budgets by school term, not calendar month. Free to use, with 50+ users in production, and I'm building it into a company.

  • Pocky, an AI assistant (EN/ZH) that drafts transactions from chat for one-tap confirmation, answers questions about your spending, and plans savings with deterministic math. Model chosen by an 82-case eval in October; prompt caching makes each turn ~70% cheaper.
  • On-device screenshot import (OCR in the browser, zero silent amount errors on 215 benchmark rows) and Whisper dictation, nudged toward your merchant names (90% to 97% on a 60-clip synthetic set); plus a read-only MCP server for your AI client.

PreBell

    Made with:

1k+ average daily visits over 14 days, including crawler traffic. Automated US market briefs and visual earnings checks in English and Chinese at prebell.site.

  • A Rust + WebAssembly verification pipeline rejects unsupported numbers and explanations before publication, backed by 1,100+ tests.
  • Tracks earnings for 562 companies, produces detailed checks for 140, and exposes filing-backed results through a read-only MCP server.
  • Early users show more interest in earnings than daily briefs, guiding the next iteration toward clearer quarter comparisons and reasons to return for the next report.

SlopFish

    Made with:

A fork of Stockfish for a chess variant where capturing is mandatory. Built for Waterloo's CO 456 Game Theory course and placed top 4 in class.

  • The rule must hold at every search node, so it reaches into move generation, ordering, check evasions and quiescence, where stand-pat pruning is disabled because a forced capture breaks alpha-beta's lower bound.
  • Trained a variant-specific NNUE on 17.2M self-play positions and shipped it as the default. An independent self-play arbiter caught rule violations in 2.5% of games (promotions counted as captures); fixed and verified at zero in 400 games.

DearByte

    Made with:

A self-hosted personal AI agent that turns Apple Watch health data, calendar and memory into morning briefs over Telegram. Open source.

  • TypeScript agent loop on the Anthropic SDK with a validated tool registry and two model tiers: a cheap worker for bulk work and a stronger brain for judgment calls, swappable between DeepSeek and Claude. Every call is cost-logged under a weekly cap.
  • Apple HealthKit bridged over remote MCP on Cloudflare Workers; spending context from TermPocket over read-only MCP; x402 purchases on Base Sepolia with explicit approval, seller allowlists and caps enforced in code.
DearByte running in Telegram, showing a contextual morning brief and a testnet purchase approval request

openfilm (open source)

    Made with:

Contributor to openfilm, the open-source AI video agent (8k+ stars, MIT). Some of what I've shipped across its agent layer, render pipeline and Electron studio:

  • Built resumable rendering: films render in 1–10 s chunks, so a stopped render resumes where it left off. A 60 s 1080p film resumed frame-identical in half the time.
  • Agent layer: conversations remember their agent and model across restarts, and the OS pings you when a turn finishes or needs an answer. Plus a 0.1× to 100× clip speed past Chromium's 16× media cap; the rest is under My PRs.

AGE Game Engine

    Made with:

A C++20 module-based MVC game engine for 2D ASCII games, running Space Invaders and Snake.

  • Entity-Component System with rendering, input, game loop and collision built in; new games subclass GameEngine and override virtual hooks.

My Education

University of Waterloo logo

Bachelor of Mathematics

University of Waterloo

2023 - 2028

  • Major in Computational Mathematics
  • Major in Statistics
  • Minor in Combinatorics and Optimization
  • Minor in Computing (Computer Science)
  • Activities: Founding Member at Sustainable Physical AI Society (SPAIS); ML Lead at WAT.ai (Sentiment Pulse); Core Member at Waterloo Blockchain Club (Tech & Event Team); Basketball Intramural Champion
  • Available for 4- or 8-month co-op placements, May – December 2027 (Summer and Fall 2027).

Blog

10 min read

TermPocket: a budget that runs on school terms, not months

A finance app for Canadian university students that budgets by term, reads screenshots on your device, and lets you record spending and plan goals by talking to Pocky.

  • project
22 min read

DearByte: Agent That Knows You

An archived experiment in a self-hosted personal agent: Apple Watch, calendar, budget and approval-based payments. What I built, what broke, and why I stopped development to focus on specialized tools and integrations.

  • project
17 min read

A Waterloo Student's Honest Recruiting Recap 2025/2026

From zero interviews to 30+ rounds including startups, unicorns, and FAANG. The unfiltered version.

  • career
  • personal
16 min read

PreBell: market news where code checks every number

An automated US market site in English and Chinese. A model writes the briefs and earnings checks; code decides whether each number and each reason is true enough to publish.

  • project
View all posts →

Contact me

Open to Summer & Fall 2027 co-op

4- or 8-month placements, May – December 2027. I'm also open to part-time work and interesting side projects. Email is the fastest way to reach me.

m69liu@uwaterloo.ca