Version-aware RAG system for Pydantic 1.x→2.x migration — hybrid retrieval (dense + BM25 + RRF) with version filtering, reaching 0.958 Recall@5 on change retrieval and 0.948 on supporting evidence.
Leakage-safe fraud modeling framework with TreeSHAP + temporal holdout evaluation; TabICLv2 hit 0.951 ROC-AUC, used to quantify the marginal value of third-party fraud signals.
M.S. student in Data Analytics Engineering at Northeastern University (Seattle), building on a B.S. in Computer Science & Environmental and Sustainability Sciences from Northeastern (Boston). I've spent my internships engineering LLM-based agents and RAG systems — from multi-agent orchestration layers to retrieval pipelines — and I like turning research-grade AI into things people can actually use.
Programming
PythonSQLJavaRAI & LLM Systems
RAGMulti-Agent SystemsAgentScope OpenAI APIFunction CallingMachine Learning
XGBoostCatBoostScikit-learn TensorFlow/KerasResNet50OptunaSHAPRetrieval & Data
BM25RRFPandasNumPy PostgreSQLMySQLMilvusRedisChromaDBEngineering & Visualization
GitpytestStreamlitTableau PlotlyMatplotlibCurrently an AI Engineer Intern at Wuhan Survey & Design Institute (remote), building a Plan-and-Execute multi-agent orchestration layer — while working through my M.S. in Data Analytics Engineering at Northeastern.
Apr 2026 – Sep 2026 · Remote. LLM-based intent router + Plan-and-Execute multi-agent orchestration; pushed routing accuracy from 65% to 90%+ and cut workflow latency 50%.
Apr 2025 – Jul 2025 · Remote. LLM-driven document-to-presentation pipeline with ViT-based layout understanding; improved content/design/structure quality 13–29%+ over baselines.
Jun 2023 – Aug 2023 · Beijing. Cross-dataset facial analytics pipeline (FER2013, KDEF) with ResNet50 transfer learning — 82.3% emotion accuracy, 99.0% gender accuracy.
Add your build/study playlist here — track names, links, whatever fits your vibe.
Email: shi.jiaqi1@northeastern.edu
Phone: (857) 654-7075
LinkedIn: linkedin.com/in/jiaqis
GitHub: github.com/KianaShi
👋 Hey — thanks for stopping by. Wire this up to a real form (or Google Form) if you want visitors to leave a note.
Empty. I don't throw away good ideas.