David Lee
Software engineer focused on large-scale distributed systems and machine learning.
Currently working on Google Cloud Pub/Sub. Previously worked on infra at Amazon.
Experience
Google
- Machine Learning Engineer
Present
- Working broadly on the intersection of ML systems, agentic AI, and reinforcement learning
Google
- Software Engineer
Jul 2025 - Present
- Worked on Google Cloud Pub/Sub, a globally distributed messaging service at ~4TB/s throughput with high availability
- Worked on AI inference SMT end-to-end, including implementation, testing, metrics, and SLOs
- On-call for production systems, including debugging and customer support for companies like Anthropic
Amazon
- Software Engineer Intern
May 2024 - Aug 2024
- Engineered an AWS Serverless automation pipeline (CDK, Lambda, SQS) that reduced weekly on-call burden by 30% and eliminated a class of high-severity tickets across multiple service teams
- Contributed to Ruby on Rails & React frontend, designed new frontend POC with CI/CD
- Delivered core intern project one month early, enabling successful execution of two stretch goals focusing on infrastructure and frontend CI/CD improvements
L
2024
- Multi-agent RAG system that automates data analysis, code execution, and database queries
- Created built-in Excel frontend for seamless user interaction
- Selected for Y Combinator interview (top 5% of applicants)
UC Berkeley
- BA Computer Science
May 2025
- Graduated with BA in Computer Science
- UPE Honor Society (top 1/3 of CS students)
Projects
- Implemented a transformer language model from scratch using a modern LLAMA-style architecture with SwiGLU activation, multi-head self-attention, RoPE positional embeddings, and FlashAttention
- Built a training pipeline with pretraining, supervised fine-tuning, and DPO-based preference optimization
- Implemented multi-GPU training using PyTorch DDP and distributed data loading
- Evaluated KL regularization, reward model design, and optimization stability in small-model preference learning
- Implemented an AlphaZero-style reinforcement learning system with ResNet-based policy/value networks and MCTS using PUCT selection
- Built a self-play data generation pipeline with asynchronous workers, experience replay, and temperature-based action sampling
- Analyzed policy improvement dynamics, value bootstrapping behavior, and the interaction between neural priors and search depth
Anime Recommendation System
- Built a multi-stage recommendation system with two-tower retrieval, ANN-based candidate generation, cross-encoder ranking, and diversity-aware reranking
- Implemented bi-encoder embeddings trained with contrastive learning and HNSW-based approximate nearest neighbor search
- Optimized for sub-100ms latency using model distillation, quantization, and batched inference under production-style serving constraints
Lightweight Pub/Sub
- Built a distributed publish-subscribe system using Raft consensus for leader election and replicated log-based fault tolerance
- Implemented configurable delivery semantics (at-least-once, exactly-once), message ordering guarantees, and dead-letter queues
- Designed partition-based scaling, consumer group coordination, and offset management for high-throughput workloads
- Implemented core OS subsystems including thread scheduling, synchronization primitives, virtual memory, and filesystem support
- Designed and debugged preemptive scheduler behavior, race-condition avoidance, and priority inversion mitigation under concurrent workloads
- Built process loading, user-kernel transitions, and page fault handling with careful memory allocation tradeoffs
- Built a secure, end-to-end encrypted file storage client in Go applying cryptographic primitives for confidentiality, integrity, and access control
- Designed authentication, file upload/download, efficient appending, secure sharing, and controlled revocation under adversarial threat assumptions
- Reasoned about key management, data organization, and tamper-resistant storage across untrusted server APIs
Voice Language Learning Agent
- Built a voice-based language learning agent focused on personalized, interactive feedback rather than static content delivery
- Designed agent control flow, explicit state tracking, and response adaptation based on user proficiency and interaction history
- Addressed reliability challenges including turn-taking errors, speech recognition failures, and conversational drift
AI Trading System
- Implemented an automated trading system combining fundamental signal extraction with agent-driven decision logic and position management
- Designed clear separation between signal generation, execution engine, and risk controls to prevent feedback loops and overfitting
- Analyzed failure modes including noisy data pipelines, delayed signals, regime shifts, and agent behavior leading to unstable outcomes