PROJECT 01Public demo · maintained
J-Grad Admission RAG
Turns graduate-admission guidelines into applicant guidance that cites the official page
NLP / DOCUMENT UNDERSTANDING
From how a document looks,
to how a machine reads it.
Same text, different font, margins or headings — does a machine still recover the same structure? I study the link between document-structure understanding and RAG retrieval, and build systems that use it.
M1 at Science Tokyo · NLP · international student from China
Graduating 2028 / targeting AI/ML engineering and R&D roles / formerly back-end on banking core systems
The table, page and figure buttons change the lens on the same document.
FIG. 01 / DOCUMENT ATLASSELECT A LENS
One page, several structures.
Headings, tables, figures, notes. The relations between them decide how the document reads.
The atlas is a hand-made explanatory drawing with Japanese sample text, not model output or an experimental result.
01Works
Project details →PROJECT 02In development (phase 2: subtitles, replies, document answers and review working)
Japanese Speaking Coach
A Windows app for the moment you are mid-conversation in Japanese: furigana subtitles and replies you can say as they are; the recommended reply leads with its conclusion and can answer from your own course documents with a citation
02Research map
papers → directions → my questionsI study where necessary evidence can be lost when a system answers from documents. Retrieval-augmented generation, or RAG, finds relevant material and uses it to produce an answer. Documents mix current rules, obsolete clauses and exceptions. Reading the words is not enough if their validity or conditions are lost. My aim is to vary state representations or typography while keeping the wording fixed, and examine retrieval, reading and evidence use separately.
T1Semantic & Structural Evidence Fidelity in Text/Structure RAGDev. observationsT2Mechanistic Decomposition of Visual-Text FailureIn preparationT3Page-Level Visual Retrieval Stability under TypographyIn preparationT4Passive Typography and Evidence ArbitrationCandidate
No theme has reached a formal conclusion yet. Each theme carries one of the four states below. · 65 papers · 7 directions · 4 themes · 183 relations
03Working with AI
Read the records →What I delegated and where I re-decided. Not just the results — the judgement and the verification, kept.
Start with collaboration and verification practices you can use in other AI development, then move into project-specific design problems.
RECORD FORMAT
what happened → human correction → evidence
Last updated 2026-10-26