AI Agent Product Design & Development · Voice Interaction · Evaluation & Quality Tools
I focus on how AI agents turn user needs into verifiable task execution. Through Codex-assisted development, I explore real-time voice interaction, background task coordination, tool permissions, and execution-result verification. I also build tools for Chinese–Japanese language and speech quality.
| Project | What it explores | Start here |
|---|---|---|
| Lingbo Desk | A Windows desktop voice agent for continuous conversation, natural interruptions, and background work, with permission checks, progress controls, and execution receipts for file, Office, and browser tasks. | Architecture · Quick start |
| J-GameLQA | Chinese-to-Japanese game localization QA, combining rule checks with context-aware LLM evaluation. | Local demo · Evaluation |
| J-SpeechOps | Japanese speech data QC: generation, pronunciation evidence, human review, and release gates. | Architecture · Example flow |
| Japanese Speech X-Ray | Waveform, pitch, and intonation inspection, with experimental speech alignment and descriptive reference comparisons. | Product showcase · Reproduction guide |
These are prototypes and research-oriented tools. Each repository documents its validation scope, reproduction steps, and limitations.
- Use deterministic rules for technical constraints and model-assisted checks for context-dependent questions.
- Keep measured evidence, model predictions, and human approval separate.
- Make experiments reproducible, preserve failure cases, and state uncertainty explicitly.
Tools used across these projects: TypeScript · Node.js · Electron · Vue · Python · FastAPI · Pydantic · Streamlit · DeepSeek · Ollama / Qwen · librosa / pYIN · GitHub Actions
Snapshot captured 2026-11-11 18:41 (Asia/Shanghai).
| Project | Latest commit date | Commit |
|---|---|---|
| Lingbo Desk | 2026-11-07 | 389b349 |
| J-GameLQA | Not in this snapshot | — |
| J-SpeechOps | 2026-10-16 | b9dc139 |
| Japanese Speech X-Ray | 2026-10-28 | ccc19dc |