Juncheng Zhu | CS @ UWaterloo

About

I'm a second-year Computer Science student at the University of Waterloo, with four years of competitive programming behind me (USACO Gold, CMO Qualifier). These days I build AI‑powered systems end to end, with a current focus on LLM infrastructure — hierarchical memory, retrieval, and agentic tool design.

Selected Projects

Dreamory: three-tier memory architecture with emotion engine and agentic layer

Python (FastAPI) PostgreSQL + pgvector Agentic Tools

Dreamory: Memory Infrastructure for AI Companions

Infrastructure that lets AI companions actually remember. A three-tier hierarchical memory system (context window, time-decay hot zone, cold storage with dual-axis content/emotion retrieval), a pure-code emotion engine grounded in attachment theory, and a bounded agentic tool loop with offline “dream” consolidation. Verified by 144 offline unit tests; now being re-architected into decoupled, agent-mediated tools.

Source Code Design Notes →

Onyx: AI-Enhanced Personal Analytics Platform

Python (Flask) Redis + SSE LLM Pipelines Docker/Nginx

Onyx: Personal Analytics Platform

An AI-augmented productivity platform self-hosted on AWS EC2 — a three-service Docker Compose stack (Flask/Gunicorn, PostgreSQL, Redis) behind Nginx, auto-deployed via GitHub Actions. State syncs across devices in ~150ms through Server-Sent Events and Redis Pub/Sub, and LLM pipelines generate persona-driven daily audits and auto-categorize sessions into live dashboards. Every AI response collects user preference ratings that are re-injected into later prompts as few-shot examples — style alignment without retraining.

Live Demo Data Logger →

Rust & PyO3 PyTorch (RL) Rayon Parallelism Flask/Nginx

Aetheris: Vectorized RL Engine

A high-performance Reinforcement Learning environment decoupling the game engine from the neural network. Features a custom Rust-based simulator utilizing Rayon for parallel execution, bypassing the Python GIL to dramatically accelerate Dueling Double DQN training throughput.

Source Code →

React 18 Vite Value Noise Canvas API

Sakura Rain Extension

(Published on Web Store) A high-performance Chrome Extension refactored with React & Vite. Replaced standard animations with a custom 1D Value Noise algorithm to simulate organic wind patterns, rendering 1200+ procedural petals at 60 FPS.

Chrome Store Source Code →

Vanilla JS HTML5 Canvas Physics Engine

Fireworks Vs Ango

A lightweight browser-based action game. Engineered a high-performance particle system from scratch without external engines, creating custom physics simulations for entity interactions.

Source Code →

Technical Skills

LanguagesPython, C/C++, JavaScript, SQL, Rust

Backend & CloudFastAPI, Flask, Nginx, AWS (EC2), Docker, Linux, Redis, Server-Sent Events (SSE)

FrontendHTML/CSS, React, Chart.js

AI/ML & DataPyTorch, RAG, pgvector, NumPy, Pandas

Tools & DatabasesGit, GitHub Actions (CI/CD), Pytest, PostgreSQL, SQLite

Education

University of Waterloo 2025 – present

Honours Computer Science. René Descartes National Scholarship ($25,000) · USACO Gold Division · CMO Qualifier (Top 50 in Canada).

Bayview Secondary School 2021 – 2025

Math Club President — mentored 30+ students and led the team to 13th place globally in HMMT Team Guts.