Haozhe L.

Key Skills

Software

Label Studio

Other

Top Subject Matter

Artificial Intelligence / Machine Learning

Natural Language Processing / LLMs

Software / Technology

Top Data Types

Text

Image

Document

Top Task Types

Classification

Text Generation

Question Answering

Fine-tuning

Freelancer Overview

I have experience in AI training data through LLM and RAG projects. During my internship at a leading AI company in China, I helped build Q&A datasets, improve retrieval quality, and create evaluation workflows combining automated metrics and expert review. I also worked on agent systems and Text-to-SQL projects, which strengthened my skills in data structuring, annotation logic, prompt design, and AI system evaluation.

Labeling Experience

AI / RAG Intern — LLM-based data augmentation + expert feedback to data loop

TextPrompt + Response Writing (SFT)

Created an LLM-based data augmentation pipeline to expand training data for the dialogue assistant while controlling output format and reducing hallucinations. Converted expert feedback into weakly supervised data to form a feedback-to-data-to-optimization loop for improved end-to-end usability.• Expanded the training dataset from 10,000 to 30,000 samples via query rewriting and answer generation.• Used structured prompts to reduce hallucination and format deviation, improving stability.• Translated expert annotations into weak supervision signals for model optimization.• Ran multiple annotation/iteration rounds to reach ~95%+ end-to-end usable accuracy and reduce manual revision costs.

2023 - 2024

AI / RAG Intern — Tendering Dialogue Assistant (retrieval evaluation with expert-annotated QA/regression framework)

Text

Reconstructed and evaluated a RAG retrieval pipeline for a Tendering Dialogue Assistant, using annotated QA materials and expert review to guide iterative improvements. Performed embedding model selection and retrieval effectiveness measurement to ensure higher recall and reduced retrieval failures.• Recalled key document selection targets using a fixed-window Chinese chunking strategy.• Evaluated retrieval components including BM25, vector retrieval, cross-encoder reranking, and Aho-Corasick query expansion.• Selected embedding models by comparing cosine similarity on 300 manually constructed QA pairs.• Built a dual-track evaluation framework combining automatic metrics and expert review with multi-dimensional key point annotations.

2023 - 2024

Education

Work History