• Drove post-training iteration loops (SFT → DPO/RLAIF → reward modeling) end-to-end, coordinating data curation, training runs, and deployment gating for models serving 1B+ devices.
• Designed and delivered synthetic data pipelines (scenario generation, preference pairs, reward signals) that raised alignment pass rates from ~60% to 92%+ on internal safety benchmarks.
• Shipped a fault-tolerant Natural-Language-to-Python platform spanning 2,000+ user intents, architecting a multi-stage system of semantic clustering, tool generation, and sandboxed code synthesis.
• Achieved ~95% tool-calling accuracy on held-out multi-step agentic workflows by developing a rigorous evaluation harness with zero-shot testing, adversarial perturbations, and argument fuzzing.
• Architected a meta-prompting self-improvement engine that turns failure patterns into automated remediation, reducing manual diagnostics and resolution effort by ~85%.
• Built a unified evaluation framework with dataset versioning and a metric registry to consolidate disparate data stacks, enabling reproducible, traceable comparisons across all models and agents.
• Led red-teaming sprints and built automated adversarial probing for jailbreak, prompt-injection, and policy-violation detection across agentic workflows.
• Mentored junior engineers on evaluation-driven development and contributed to cross-org alignment standards for LLM safety and quality.

• Improved tool retrieval and query-to-program quality (+10% recall, −30% hallucinations) through RAG, LoRA/SFT on large models, prompt optimization, and user-simulation evals — early contributions to Apple Intelligence.
• Built synthetic data and evaluation pipelines for intent understanding, enabling rapid iteration on LLM-driven features across Siri and system-level intelligence.
• Increased OS update success rate by +25% by developing time-series forecasting models to smooth peak server loads and reduce resource contention failures.
• Co-inventor on three U.S. patents covering sensor-driven display power control and predictive energy management.

• Built on-device ML models for intelligent charging and background scheduling using device-state signals under strict privacy constraints, improving battery-health outcomes without cloud data.
• Enabled Apple ProRes video on iPhone 13 Pro/Max by optimizing background task scheduling within power and thermal budgets.

• Unity editor tools and real-time inspectors; ~3x designer workflow speed-up; shipped to titles serving 2M+ DAU.

• Spring Boot/Node.js monitoring for millions of instruments; Oracle SQL optimized ~120s → ~3s; led a 6-person LSTM prototype for predictive monitoring.

Patents

• U.S. Patent No. 12,141,012 — Energy saving for battery powered devices (2024)
• U.S. Patent No. 12,436,593 B2 — Sensor-based display power control (2025)
• U.S. Patent No. 12,436,594 B2 — Predictive display power control (2025)

Academic Honors

• Dean's Honors List, The University of Hong Kong (2015-2019)
• CLP "Powering a Sustainable Generation" Scholarship (2017)
• Zhiyuan Scholarship (Soong Ching Ling Foundation) (2015)