Apple Machine Learning Engineer Intern · MLE / MLSys · AI Agent
Carnegie Mellon University, School of Computer Science
I'm a graduate student at Carnegie Mellon University (School of Computer Science), pursuing an M.S. in Automated Science.
I received my B.S. in Computer Science (emphasis in Algorithms) from Santa Clara University, where I worked with Dr. Lang Chen. Dr. Nicholas Q. Tran served as my academic advisor.
My interests lie in on-device ML and LLM inference, ML systems, agentic ML workflows, and full-stack intelligent applications.
Looking for opportunities in On-device ML, MLSys, AI Infrastructure, and AI Agents. Please reach out at xthomaswang@gmail.com.
Selected projects I contribute to or build:
Proposed and driving a four-phase roadmap making embedding models first-class in mlc-ai/mlc-llm: OpenAI-compatible /v1/embeddings serving, a TVM-native runtime, and batched C++ encoder/decoder prefill — 14× embedding p50 speedup (36s → 2.5s) on Apple M1 Pro. 9 PRs (6 merged).
Core ML Tools — LightGBM to Core ML converter supporting Booster, LGBMClassifier, and LGBMRegressor models.
Machine Learning Engineer Intern at Apple (May 11 — Aug 21, 2026), building agentic end-to-end ML training workflows for cellular ML models and developing Call Context, a feature announced at WWDC26, using on-device LLM inference.
Article published in Communications Biology — E/I imbalance and internal noise cause weak neural representations and face recognition challenges in ASD.
Open source contributor to apache/tvm — upstreamed masked sequence prefill for the KV-cache attention kernels (encoder valid lengths + causal left-padding) and compiler/runtime fixes powering MLC-LLM device backends.
Building the Localized Agent Harness — a local-first agent harness with progressive long-term memory and Sleep Model embedding fine-tuning. Modular Melo / Support / Test architecture, supporting Ollama, MLC, and OpenAI-compatible backends.
Open source contributor to xthomaswang/OpenOT2 — lab automation framework for the OT-2 liquid handling robot.
Poster at GenAI4Health @NeurIPS 2025 — The Second Workshop on GenAI for Health: Potential, Trust, and Policy Compliance. San Diego, CA.
Research Engineer at CMU — built AI-powered biomedical literature analysis platform with RAG pipeline and embedding-based semantic search. Fine-tuned Qwen3 0.6B embedding model on multi-GPU, achieving strong benchmark on PMC-Patients PPR task.
Started graduate studies at Carnegie Mellon University in the MSAS program, School of Computer Science. Director's Scholarship recipient.
Poster at CogSci 2025 — 47th Annual Meeting of the Cognitive Science Society. San Francisco, CA.
ML Research Intern at LCCN Lab, SCU. Fine-tuned CNNs and Vision Transformers to test neuroscience hypotheses. Built fMRI data processing and representation analysis pipelines.
Scroll to browse earlier updates
Cupertino, CA
Built agentic developer workflows for cellular ML model training, evaluation, Core ML packaging, and iPhone performance validation. Developing Call Context, a feature announced at WWDC26, using on-device LLM inference; benchmarking different inference strategies and evaluating performance trade-offs on iPhone.
Remote
Proposed and drove a four-phase roadmap making embedding models first-class in MLC-LLM's on-device inference engine: Qwen3-Embedding conversion/quantization, OpenAI-compatible /v1/embeddings serving, a TVM-native runtime, and batched C++ encoder/decoder prefill. Cut embedding p50 latency 14× (36s → 2.5s) on Apple M1 Pro; upstreamed KV-cache masked prefill and compiler/runtime fixes to Apache TVM.
Pittsburgh, PA
Built an AI-powered biomedical literature analysis platform with Quick (<30s) and Deep (<=3min) modes using a RAG pipeline with parallel retrieval from PubMed and MedRxiv. Fine-tuned Qwen3 0.6B embedding model on multi-GPU, achieving strong benchmark results on PMC-Patients PPR task. Poster at NeurIPS 2025 Workshop on GenAI for Health.
Santa Clara, CA
Developed CNN models for neuroscience research and preprocessed fMRI data for HPC workflows. Article published in Communications Biology; poster accepted at CogSci 2025.
M.S. in Automated Science
School of Computer Science · Pittsburgh, PA
Director's ScholarshipB.S. in Computer Science, Minor in Economics
College of Arts and Sciences · Santa Clara, CA
REAL Program ScholarAI-powered literature analysis system using a RAG pipeline with PubMed & MedRxiv. Fine-tuned Qwen3 0.6B embedding model on multi-GPU, achieving strong benchmark on PMC-Patients PPR task. Poster at NeurIPS 2025 Workshop on GenAI for Health.
Open-source contributor to Apache TVM, an end-to-end deep learning compiler stack. Upstreamed masked sequence prefill for the KV-cache attention kernels — encoder valid-length masking and causal left-padding — plus compiler/runtime fixes, powering batched embedding inference in MLC-LLM's device backends (3 merged PRs).
A local-first agent harness for always-on agents that automate routine work and learn user habits. Modular architecture (Melo core / Support / Test), with progressive long-term memory and a Sleep Model that fine-tunes the embedding model offline. Hybrid local + API execution, supporting Ollama, MLC, and OpenAI-compatible backends.
Proposed and driving a four-phase roadmap (issue #3451) making embedding models first-class in MLC-LLM's on-device inference engine: Qwen3-Embedding conversion/quantization, OpenAI-compatible /v1/embeddings serving, a TVM-native runtime, and batched C++ encoder/decoder prefill — 9 PRs (6 merged). Cut embedding p50 latency 14× (36s → 2.5s) on Apple M1 Pro and fixed attention-scaling correctness on the Metal/CPU fallback path. Also used the framework to run quantized LLaMA 3 on-device in the Headache Note iOS app.
Open-source contribution to OpenOT2, a lab automation framework for the OT-2 liquid handling robot. Enabling reproducible scientific workflows through programmable protocols.
Fully offline iOS AI agent built by quantizing LLaMA 3 1.6B with MLC-LLM, optimized for iPhone. Swift-based health assistant with prompt engineering and SwiftData for private, on-device inference and personalized lifestyle recommendations.
Research on face recognition challenges in ASD using CNNs, ResNet50, and Vision Transformers as computational frameworks. Built fMRI data processing pipelines and Pearson correlation analysis. Article published in Communications Biology (Nature Portfolio, 2026); poster at CogSci 2025.
AI agent discovery platform with natural language semantic search using Sentence Transformers and Faiss for efficient similarity matching. FastAPI + SQLAlchemy backend with user ratings, recommendation logging, and SHA-256 data integrity.
SwiftUI iOS app with MVVM architecture, integrating a TensorFlow-trained CoreML model for food classification with 85% accuracy across 100+ categories. Optimized inference speed by 4x and reduced memory usage by 40%.
Full-stack iOS app with React Native and SQLite. Optimized data queries by 75% (200ms to 50ms). Integrated Apple HealthKit for real-time HRV metrics with 98% accuracy. 90% test coverage.
Full-stack web server using the Go Gin framework with RESTful API, JSON-based database for efficient data management, and a scalable commenting system.
Download the version most relevant to you.