Hiro N.

Key Skills

Software

Mighty AI

Other

Internal/Proprietary Tooling

Top Subject Matter

LLM + RAG knowledge Q&A over documents

Multilingual Labeling (EN/JP/CN)

AI Safety & Red Teaming

Top Data Types

Document

Audio

Text

Top Task Types

Question Answering

Transcription

Data Collection

Freelancer Overview

LLM + RAG 知识问答系统（PDF/Word/Markdown）. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Chroma, FAISS, and Other. Education includes Bachelor of Science, Wuxi Taihu University. AI-training focus includes data types such as Document and Audio and labeling workflows including Question Answering and Transcription.

Labeling Experience

多模态图像标注与质量控制项目

ImageData Collection

**项目范围：** 覆盖电商产品图、自然场景图及AI生成图像三类数据源，针对物体检测、语义分割及图像描述生成任务进行高质量标注，服务于计算机视觉模型的训练与评估。 **具体数据标记任务：** - 边界框标注（Bounding Box）：对目标物进行精确矩形框选，覆盖率达100% - 多边形分割（Polygon Segmentation）：对不规则物体边缘进行逐点标注，支持像素级语义理解 - 属性标签分类（Attribute Tagging）：标注颜色、尺寸、材质、动作等结构化属性字段 - 图像描述撰写（Image Captioning）：为单图及多图场景编写自然语言描述，遵循风格统一指南 **项目规模：** - 累计处理图像 15,000+ 张 - 单日最高产能 500+ 张 - 参与标注团队 5–8 人，承担审核及抽检职责 - 覆盖中英双语标签体系 **质量指标：** - 标注准确率 ≥ 98%（以预置测试集抽样验证） - 边界框 IoU（Intersection over Union）≥ 0.85 - 二次通过率（Second-pass agreement rate）≥ 95% - 每批次抽检比例 ≥ 20%，不合格批次退回重标 - 标注响应时间：单张复杂场景 ≤ 3 分钟

2023 - 2024

IndexTTS/edge-tts/CosyVoice 音频生成与流水线

OtherAudioTranscription

Developed an audio generation pipeline using TTS components and orchestrated generation workflows via automated CI/CD. Focused on producing speech outputs from prompts and managing model execution through GitHub Actions pipelines. • IndexTTS / edge-tts / CosyVoice TTS workflow • Automated pipeline execution (GitHub Actions) • Prompt-driven audio synthesis • Integration and deployment of audio services

Not specified

LLM + RAG 知识问答系统（PDF/Word/Markdown）

DocumentQuestion Answering

Built an LLM + RAG knowledge Q&A system using document sources (e.g., PDF/Word/Markdown) and implemented retrieval over knowledge bases for answering user questions. Designed the workflow to ingest documents, split/index content, and support function-calling style interactions for more controlled responses. • Document-to-knowledge ingestion and indexing • Retrieval-Augmented Generation (RAG) for Q&A • Function calling / tool invocation integration • Prompt engineering for consistent answer quality

Not specified

Education

Work History