I build evidence-first machine-learning systems: reproducible evaluation, explicit data boundaries, deployable interfaces, and honest negative results.
专注算法与 ML Engineering，覆盖 LLM/RAG、检索排序、多模态、图学习、视觉异常检测、时间序列与语音音频。每个公开项目都提供可运行演示、自动化测试和冻结发布证据；不把 toy demo 包装成真实业务效果。
| Project | Algorithm / engineering focus | Public evidence |
|---|---|---|
| EchoForge-ASR | Chinese streaming ASR, PCM WebSocket protocol, VAD, transcript revision, robustness evidence | Demo · Release |
| ChronosGuard-TS | Leakage-aware forecasting, conformal intervals, anomaly detection, online replay | Demo · Release |
| GraphShield-Fraud | Temporal graph fraud detection, GraphSAGE, investigation workflow | Demo · Release |
| MOSAIC-Retrieval | Image/video-text retrieval, CLIP, FAISS, cold-start encoding | Demo · Release |
| TRACE-Rec | Retrieval, learning-to-rank, multi-interest recommendation, experimentation | Demo · Release |
| AlignScope-CS | SFT/DPO/LoRA alignment evaluation with evidence-first reporting | Demo · Release |
| FinSight-RAG | Financial RAG, retrieval evaluation, citation and evidence boundaries | Demo · Release |
| ForgeSight-AD | Industrial visual anomaly detection and pixel localization | Demo · Release |
| NexusMind-Inference | Guarded FastAPI inference gateway for LLM/RAG workloads | Demo · Release |
- Time-aware or source-aware splits before model comparison; no Test-driven model selection.
- Reproducible reports, machine-checkable artifacts, CI, packaged releases, and responsive demos.
- Explicit distinction between synthetic UI fixtures, public benchmark evidence, and production claims.
- Negative results and failed verification gates remain visible instead of being rewritten as wins.
Open to algorithm and ML engineering opportunities.