L

Leon Liu

SWE Intern @ TikTok • Prev MLE @ Jingo Jingo

📗 CS at UCLA building ML systems.

🏢 Shipped recommendation infrastructure at Jingo (AI startup) serving 5M products with RAG and vector search.

🔬 Diffusion models for autonomous driving world modelling and AR Video generation at UCLA VAIL (Lab).

🌱 Side projects:

  • 🐍 RL agents for Atari Surround
  • ⌨️ Transformer EMG→QWERTY decoder for keyboardless typing from biosignals
  • 📝 SaaS resume builder accessible editor → LaTeX/ATS-friendly PDF
  • ... and more!

Experience

Software Engineer Intern (Incoming)

Amazon (AWS Transform) • Software Engineer Intern

AI Agents
LLM
Code Migration
AWS
Java
AWS
  • Building AI agents that automate Java migration of mainframe systems in finance and government

Software Engineer Intern

TikTok (Recommendation Infrastructure) • Software Engineer Intern

Recommendation Systems
AI Agents
Distributed Systems
Performance
Stream Processing
Developer Tooling
C++
Kafka
Flink
  • Migrated the C++ recommendation ranking service to a new framework representing feature retrieval as DAGs, allowing parallelization and pruning based on dynamic model demand, reducing cores per request and p99 latency
  • Built a correctness pipeline with Kafka and Flink over online production traffic, verifying feature and model-score parity across service migrations and ranking stages, and feature consistency between serving and training
  • Built AI agents to diagnose feature issues such as training-serving skew, querying consistency metrics and analyzing source code and live configuration

Machine Learning Engineer Intern

Jingo (AI-Native E-commerce Startup) • Machine Learning Engineer Intern

Machine Learning
Backend Engineering
LLM
Vector Search
FastAPI
Redis
PostgreSQL
SQLAlchemy
  • Built the LLM-powered product search + recommendation service over a 5M+ product catalog, backed by FastAPI, Redis, PostgreSQL, RAG, and vector search
  • Cut search p95 latency ~50% by optimizing async worker orchestration, and achieved near-instant infinite scroll by prefetching paginated results into a Redis buffer
  • Lifted rec relevance in A/B tests with a contextual rec engine combining graph-based user memory and an LLM query preprocessor/reranker
  • Improved search recall with an ETL pipeline cleaning raw product data to enrich embeddings and metadata

AI Researcher

UCLA Vision And Autonomy Intelligence Lab (VAIL) • AI Researcher

World Models
Gen AI
Diffusion
Autonomous Driving
PyTorch
DeepSpeed
  • Trained image and video models conditioned on digital twin sim-to-real data to create a controllable photo-realistic simulator-grounded generative world model for downstream self-driving simulation
  • Scaled training of Flux and Wan 2.1 generative models across 8 L40S GPUs using a PyTorch training pipeline, with DeepSpeed ZeRO to optimize throughput of distributed training jobs

Check out the project at Dreamland.

🚧 No full page available yet 🚧

Software Engineer Intern

UCLA Bionics Lab • Software Engineer Intern

Machine Learning
Computer Vision
Healthcare
PyTorch
MediaPipe
  • Trained LSTM achieving >90% accuracy for automating stroke rehabilitation assessments
  • Developed data pipelines integrating MediaPipe-based 3D triangulation, doubling clinical data processing throughput

Projects

Dreamland — Controllable Driving World Model

World Models
Gen AI
Autonomous Driving
Research
PyTorch
DeepSpeed
  • Built hybrid driving world model pairing generative video models with ground truth driving simulator for controllability and realism, achieving 12.8% lower FID (better image quality) against prior baselines
Demo

Prolio — Resume Versioning App

Full Stack
SaaS
CI/CD
LaTeX
Next.js
React
Flask
PostgreSQL
Redis
GCP
Docker
OAuth
  • Designed and deployed a full-stack SaaS resume platform with modular editing, multiple versions, Gemini LLM-powered reviews, and PWA client (Next.js, React, Tailwind, shadcn)
  • Engineered microservice backend (Flask, PostgreSQL, Redis, Celery, GCP) with async LaTeX/PDF compilation and secure OAuth authentication
  • Productionized infra with Docker, GitHub Actions, and GCP Cloud Build, enabling CI/CD and scaling
Source

emg2qwerty — sEMG-to-QWERTY AI Text Decoding

Machine Learning
Signal Processing
PyTorch
PyTorch Lightning
Transformers
Python
  • Trained Transformer sequence-to-sequence models to decode wrist-worn EMG signals into typed QWERTY text, enabling hands-free typing from biosignals
  • Outperformed Meta's published ConvNet baseline by ~10% lower character error rate through Transformer and RNN fine-tuning
  • Developed full training pipeline with preprocessing (log spectrograms, augmentation), CTCLoss optimization, and benchmark evaluation across GRU, LSTM, and Transformer architectures
  • Increased training throughput by 2.5x via spectrogram down-sampling (125Hz→50Hz) with negligible accuracy loss, enabling faster experimentation
Source

Lingsim — AI-Powered Language Learning App

Full Stack
AI/ML
EdTech
Chat
Speech Recognition
TTS
Next.js
React
Firebase
TypeScript
GPT-4
Whisper
  • Built multilingual AI tutoring full-stack app integrating speech recognition (Whisper) + GPT-4 dialogue + ElevenLabs TTS to enable real-time conversational practice
  • Demonstrated sub-second (~500 ms) speech→text→response→speech loop across ~10 languages, showcasing feasibility of real-time interactive AI tutors
  • Designed end-to-end system with Next.js, React, Firebase, and REST APIs to support secure authentication and optimized low-latency chat
Demo

Reliable Transport Layer (TCP over UDP)

Networking
Systems Programming
C
  • Built a custom transport protocol in C providing reliable, ordered delivery on top of UDP by adding packet retransmission, acknowledgments, and error detection
  • Validated correctness with stress tests simulating dropped/out-of-order packets, demonstrating consistent end-to-end data integrity

Basilisk — Multi-Agent Social Deduction Simulator

LLM
Multi-Agent Systems
Game AI
Python
LangChain
Pydantic
Gemini
Ollama
  • Developed framework to simulate and manage social deduction games with both human players and LLM-based agents
  • Integrated LangChain and Ollama/Gemini APIs for multi-agent orchestration, dialogue, and strategy execution
  • Implemented structured input/response validation with Pydantic, ensuring reliable agent interactions and reproducibility
  • Built orchestration layer for game-state tracking, rule enforcement, and cross-agent coordination in Python
Source

Finduo — Full-Stack Gaming Social Media Web App

Full Stack
Social Media
Gaming
MongoDB
Express
React
Node.js
TypeScript
  • Designed and implemented full-stack social media platform for users to find gaming partners
  • Implemented user profiles, dynamic feed, search capabilities, and viewing friend invites with React
  • Implemented RESTful API for auth, user profiles, ratings, and friends list in the backend with MongoDB, Express, Node.js
Source

Dynamic Analysis of Cryptographic API Misuse in Node.js

Security
Dynamic Analysis
Research
Node.js
TypeScript
Bash
  • Paper accepted at IEEE MIT URTC
  • Implemented multipart system to perform dynamic analysis on Node.js codebases to detect vulnerabilities from API misuse
  • Implemented custom module loader that instrumented code to log API usage
  • Developed rule enforcer to detect API misuse defined by a modular set of declarative rules
DemoSource

Unity Real-Time Strategy PVP Game

Game Development
Real-time Systems
C#
Unity
  • Engineered robust game systems and mechanics, leveraging data structures to optimize performance
  • Architected and implemented modular scripts and game systems with Object-Oriented Programming
  • Utilized Unity Profiler for game performance tuning and software testing
  • Enhanced user experience by identifying and resolving performance bottlenecks
DemoSource

Brewin Interpreter

Programming Languages
Interpreters
Compilers
Python
  • Implemented an interpreter in Python for a custom language "Brewin"
  • Designed interpreter for several variations of the language with common features
  • Handled function calls, recursion, scoping, variables, control flow (conditionals and loops)
  • Built systems for dynamic and statically-typed variations, structs, error handling, lazy evaluation
Source

Reverse Proxy for ACM Cyber

Networking
Backend
Security
TypeScript
Deno
  • Implemented a reverse proxy with HTTP request parsing, authentication (bcrypt), and structured logging to securely route and monitor backend traffic
  • Designed and tested routing, auth, and logging subsystems to improve reliability and observability in multi-service environments

Atari Reinforcement Learning

Reinforcement Learning
Python
PyTorch
PyTorch Lightning
  • Implemented policy gradient methods and deep Q learning for a reinforcement learning agent in Atari
  • Trained agent to solve and beat the Atari Surround environment
  • Spearheaded and taught a quarter-long workshop for ACM AI for 60+ members
Source