Ocean Kun Hei OU | Data Science

Education

The Hong Kong University of Science and Technology (Guangzhou)

2023 – 2027 (Expected)

BSc in Data Science and Big Data Technology

Deep LearningNatural Language ProcessingMachine Learning Cloud ComputingDatabase Management SystemsDesign and Analysis of Algorithms Discrete MathematicsApplied StatisticsC++ Programming

Research Experience

DPC: Training-Free Text-to-SQL Candidate Selection via Dual-Paradigm Consistency

Nov 2025 – Jan 2026 · ACL 2026 Main

Developed DPC (Dual-Paradigm Consistency) to address the gap between LLMs' strong SQL generation and weak candidate-selection capabilities, reframing selection from probabilistic judgment as deterministic verification. The framework uses multi-agent collaboration and an adversarial feedback loop to synthesize Minimal Distinguishing Databases (MDDs), while exploiting the complementary execution logic of SQL and Python to construct model-independent reference anchors. Integrated as a plug-and-play module into state-of-the-art systems including DAIL-SQL and CHESS, DPC improves execution accuracy by an average of 1.8%. As second author, I conducted experiments across multiple models and systems, performed in-depth case analyses, and contributed to manuscript revision.

Read the paper View source on GitHub

Selected Projects

LLM Self-Alignment System · Instruction Inversion and Quality Filtering

Dec 2025

Reproduced a self-alignment pipeline by LoRA fine-tuning Qwen2.5-3B-Instruct on OpenAssistant-Guanaco as a reverse model (response → instruction), updating only 0.2% of model parameters. Sampled responses from LIMA, generated pseudo-instructions, and used the base model as a few-shot judge to score candidate pairs from 1 to 5. Fine-tuned again on the high-quality subset, creating a fully annotation-free loop that substantially improved response quality.

AnkiDroid Personalized Study Reports · End-to-End Data Pipeline

Sep 2025 – Dec 2025

Contributed multiple-choice and study-report modules to the open-source AnkiDroid project. Designed behavioral data collection for card ratings, review intervals, and time spent in each state, then serialized the data as JSON for backend processing. Built and deployed a Flask REST API for GBM inference, engineered features including memory_strength, review_interval, and cognitive_load, and connected the mobile client to remote predictions—completing the loop from data collection and feature engineering to cloud inference and personalized recommendations.

Multi-Tenant Kubernetes Platform

Apr 2026 – May 2026

Built a multi-tenant Kubernetes lab platform on K3s, providing dedicated namespaces for independent teams sharing cluster infrastructure. Implemented namespace-scoped RBAC using Roles, RoleBindings, and ServiceAccounts for tenant administrators, developers, and viewers. Applied least-privilege permissions to separate workload management from sensitive operations involving secrets, access policies, and cluster administration. Enforced resource governance through ResourceQuota and LimitRange. Added observability through Prometheus and Grafana configuration. Integrated Horizontal Pod Autoscaler (HPA) to adjust workload replicas according to CPU utilization, configuring replica bounds and resource requests to support elastic scaling under variable demand within tenant resource constraints.

Activities and Leadership

Chinese Orchestra Society, HKUST(GZ) · President

Aug 2025 – May 2026

Lead member recruitment and regular rehearsals; independently planned and delivered a dedicated concert, coordinating the venue, program, and end-to-end promotion while fostering exchange in Chinese traditional music on and beyond campus.

PLACE Summer Camp, HKUST(GZ) · Teaching Assistant

Summers 2024 & 2025

Co-designed icebreakers and team activities, helped international students adapt to campus culture, and handled on-site contingencies in a cross-cultural environment.

15th National Games of the People's Republic of China · Opening Ceremony Volunteer

Nov 2025

Supported opening-ceremony operations for a national sporting event, helping maintain smooth on-site coordination and execution.

Honors and Scholarships

National Second Prize, 19th Chinese Collegiate Computing Design Contest (2026)

Second-Class Scholarship for Hong Kong, Macao, Taiwan, and Overseas Chinese Students (2025)

Outstanding Award, HKUST(GZ) Undergraduate Research Opportunities Program (2024)

HKUST(GZ) First-Class Scholarship (2023)

Technical Skills

Python / SQL
LLM Agent Workflow
RAG Retrieval & Knowledge Augmentation
PyTorch
Transformer Architecture
Docker / Linux / Cloud Deployment
LLM Fine-tuning (SFT, LoRA, DPO)

Possess experience in large language model application development and data intelligence system construction; familiar with intelligent agent workflow design; capable of completing the full development process from data processing and model inference to system deployment.

Research interests: Data Intelligence · Text-to-SQL · LLM Agents