Teacher
A local AI learning companion for desktop and Android that connects private document retrieval, knowledge tracing, on-device analysis, and LAN sync.
HAN XIZE / XUZETOP1
Exploring the space between local AI, hybrid retrieval, and agent tooling. This is where I keep my open-source projects — and the code and materials that make research checkable.
01 / SELECTED PROJECTS
Ongoing work across AI applications, developer tools, and research reproduction.
A local AI learning companion for desktop and Android that connects private document retrieval, knowledge tracing, on-device analysis, and LAN sync.
An agent cache runtime for LLM applications that brings cache awareness into the application's execution flow and provides groundwork for cache-related development.
An orchestration bridge across coding agent tools — Codex, Claude Code, OpenCode, and Qwen — with isolation, audit, and approval mechanisms.
AI research assistance for topic selection, experiment design, paper outlining, citation auditing, and iterative review — helping researchers organise work and check their arguments.
Public reproduction materials for offline hybrid retrieval evaluation, comparing BM25, dense retrieval, and RRF within a unified candidate space, with frozen rankings, labels, analysis code, and figures.
Showing all 5 projects
02 / ACADEMIC & RESEARCH
Focused on retrieval methods, evaluation process, and the boundaries of reproduction. Research materials and publication details are listed separately.
Combining SQLite FTS5 BM25, BGE-M3 dense retrieval, and RRF, with frozen inputs and the analysis pipeline published so that offline evaluation results can be verified.
The manuscript has been accepted; the public materials keep the frozen inputs, analysis pipeline, and figures, but not the raw corpus — see the repository notes for the scope of reproduction.
Read the research materials and reproduction notesThe International Conference on Machine Learning, Embodied Intelligence and Large Models (MLEIL 2026) · Accepted for the conference proceedings; the proceedings will be submitted to EI for potential indexation, subject to official publication
Within a unified candidate space of 1,696 educational entities, we compare SQLite FTS5, BGE-M3 dense retrieval, and RRF fusion. Over 93 questions and 186 query instances, RRF reaches NDCG@5 0.8916 and MRR@5 0.9330 — 0.1124 and 0.1781 above Dense and BM25 respectively. A fixed-list quota ablation further shows how deployment source policy changes top-five coverage and graded ranking.
© 2026 Xize Han, Xin Xie. This is the author's version of a paper accepted for publication in the Proceedings of MLEIL 2026.
03 / ABOUT
My public projects revolve around local AI applications, retrieval evaluation, and agent tooling: tools that serve a real problem, and records that let an experiment be revisited and checked.
Implementations, research materials, and repository documentation live on GitHub.