LLM & SLM training
Post-training Llama and Qwen models, including 3B-parameter models, using supervised fine-tuning and LoRA. Developing small language models for domain-specific healthcare applications.
MACHINE LEARNING ENGINEER & RESEARCHER
HaoqiuSongPh.D.
Language models.
Intelligent systems. Healthcare AI.
I lead the technical development of machine learning systems—from LLM/SLM training and reinforcement learning to evaluation and production deployment.
PROFESSIONAL PROFILE
Technical leadership.
End-to-end ML engineering.
I am a Machine Learning Engineer at Welldoc and hold a Ph.D. in Computer Science & Applications from Virginia Tech. I specialize in language-model post-training, reinforcement learning, agentic systems, and model evaluation for healthcare applications.
As the primary ML engineer across clinical, product, design, data, and engineering teams, I own technical decisions across the model lifecycle: architecture, data and training, evaluation, CI/CD, and deployment. My research background spans clinical time-series modeling, explainable deep learning, and genomic foundation models.
Reinforcement learning
Reward and ranking models, preference optimization, and contextual policies for personalized decision-making and next-best-action selection.
Agentic AI systems
Retrieval-augmented generation and multi-agent workflows for query routing, patient-state estimation, planning, and grounded reasoning.
Evaluation & deployment
Reusable evaluation frameworks for safety, factuality, relevance, and actionability, with regression testing and reproducible deployment pipelines.
PROFESSIONAL EXPERIENCE
From technical direction
to deployed systems.
MAR 2026 — PRESENT
Welldoc
Columbia, Maryland
Machine Learning Engineer
Primary ML engineer with ownership across architecture, training, evaluation, CI/CD, and deployment.
- Technical leadership. Drive ML implementation across cross-functional clinical, product, design, data, and engineering teams, translating application requirements into model architectures and evaluation workflows.
- LLM/SLM post-training. Fine-tune 3B-parameter Llama and Qwen models with LoRA-based supervised training to align healthcare content with clinical guidelines for readability, tone, safety, and appropriateness.
- Personalized decision systems. Develop longitudinal patient-state modeling and care planning, with ranking models, DPO-style preference optimization, and contextual reinforcement learning.
- Agentic systems and evaluation. Build RAG and multi-agent workflows, alongside reusable LLM-as-a-Judge evaluations and regression tests. Develop a small language model for virtual dietitian support and escalation to human experts.
MAY 2025 — AUG 2025
Children’s National Hospital
Washington, DC
Research Scientist Intern
- Built an LLM-orchestrated pediatric ICU decision-support system with protocol-grounded retrieval, modular query routing, and patient-context extraction.
- Developed an interpretable LSTM using 100K+ multivariate ICU samples and 100+ clinical features, with temporal and feature-level explanations.
- Engineered automated EHR and medication pipelines for computable phenotyping, reducing manual chart-review effort by more than 80%.
DEC 2022 — FEB 2026
Virginia Tech
Alexandria, Virginia
Graduate Research Assistant
- Designed scalable EHR time-series pipelines for clinical forecasting and decision-support research across millions of longitudinal data points.
- Fine-tuned genome foundation models and optimized deep classifiers for pathogen identification.
- Built reproducible Python, Shell, and Nextflow pipelines for metagenomic surveillance and integrated language models into grounded diagnostic workflows.
RESEARCH
Selected publications
FIR-LSTM: An Explainable Deep Learning Framework for Predicting Iatrogenic Withdrawal Syndrome in Pediatric Intensive Care Units
Explainable deep learning for clinical risk prediction in pediatric intensive care.
Virseqimprover: an integrated pipeline for viral contig error correction, extension, and annotation
A computational pipeline for improving and annotating viral genome assemblies.
PathoLM: Identifying pathogenicity from the DNA sequence through the Genome Foundation Model
Adapting genomic foundation models to identify pathogenicity from DNA sequences.
EDUCATION
Academic background
Ph.D.
Computer Science & Applications
Virginia TechM.S.
Computer Science & Applications
Virginia TechB.E.
Computer Science & Technology
Hubei University of Technology