Projects
What I'm building
Most of my work centers around AI-powered learning tools, a focus that grew naturally out of the roadblocks I hit in my own studies. Whether I'm tackling Japanese or a new programming concept, I've noticed that traditional tools often become rigid right when you need them to be the most adaptive. These projects are my way of exploring how AI can fill that gap — not by handing out the answers, but by providing the flexible, personalized feedback that actually helps us grow.
Active
LoRA Adapter Reuse
Research line started as a summer 2026 internship at the Institute of Information Science, Academia Sinica (July to August) and continued as part-time research assistant work. The question is whether a trained LoRA adapter can be reused: moved to a different base model without data or retraining, or merged with other single-task adapters on the same host model. Everything runs on small open models against multiple-choice benchmarks, with controls and significance tests on every claim.
Milestones
MergeBench reproduction with LoRA specialists: plan and code
Base/instruct portability 2×2 and a merge-vs-multitask control
Merging single-task adapters: raw vs. aligned space, TIES, DARE, scale and norm
Data-free cross-model LoRA transfer: Cross-LoRA reproduction, LoRA-X
AntiCopilot
AntiCopilot flips the script on AI code assistants. Instead of writing code for you, it guides you through the learning process with personalized feedback, struggle signal detection, and concept-level spaced repetition, all integrated directly into VS Code. A two-person project: I built the learner-facing side, from the VS Code extension and web dashboard to struggle-signal capture and the FSRS review loop, on top of the LangGraph planning agent my teammate built.
Milestones
Extension: I'm-stuck hint flow, solution submit, sidebar redesign
Dashboard live on Cloudflare Pages; public hub at anticopilot.nhade.com
Memory-aware hint and code-correction routes
Expo phone companion for live struggle sessions
Shadow-mode struggle logger in the extension; standalone struggle sidecar service
FSRS review loop, per-user scoping, and skill-path status API merged
Frontend and extension re-wired to the new backend; practice and learn views
Backend REST layer ported onto the content-generation backend
Review and roadmap tracking integration
FSRS review loop & practice UI
Unified API & roadmap integration
Frontend subproject started
LangGraph agent prototype
Code highlight & webview provider (PoC end)
Struggle signal & webview v1
VS Code extension first PoC
栞 (Shiori)
栞 (Shiori) combines real-world Japanese content with hybrid AI evaluation to provide meaningful feedback — not just correctness scores. It ingests NHK news articles, evaluates grammar through both deterministic NLP and LLM reasoning, and adapts to each learner's proficiency and weak points.
Milestones
Final-exam MCQ deck with per-question answer-source labels
Personal RAG: grader surfaces similar past mistakes at review time
Off the free tier: backend on Hetzner with auto-deploy, frontend on Cloudflare Pages
Personalized learner profiles & analytics
Multi-language support (EN/JA/ZH-TW)
NHK news reading with TTS
Hybrid grammar evaluation (rule-based + LLM)
Initial project kickoff
Completed
AI Mini-Projects Collection
13 practical AI applications built during an AI development course — from chatbots to RAG pipelines to multi-agent systems.
Milestones
All 13 projects completed
Course started