HBV & liver cancer microenvironment
Data integration and evidence synthesis around hepatitis B, liver cancer progression, immune microenvironment, and treatment response.
INDEPENDENT DEVELOPER · 2026
Evidence,
made checkable.
I’m 陈微 (Chenwei), an independent developer working at the intersection of HBV, liver cancer microenvironment, bioinformatics, and scientific agents — building tools that make evidence findable, traceable, and reusable.
01 / RESEARCH
Questions I keep returning to
How do we connect biological questions, multi-omics data, and literature evidence — precisely enough to cite a sentence, and honestly enough to say when the evidence is missing?
Bioinformatics & knowledge systems
Multi-omics analysis, literature evidence, and reproducible pipelines that connect biological questions to inspectable data.
Agent development
Tool-using scientific agents: retrieval, verification, citation, and workflows that know when the evidence is missing.
02 / SELECTED WORK
Things I’m building
2026
SciPal Open source
A scientific knowledge workspace: hybrid retrieval, evidence search, review workflows, and a two-database architecture that separates business data from the literature corpus.
2026
Sentence-window retrieval Research
A sentence-ID-native retrieval unit: 500-word word windows snapped to sentence boundaries, with context expanded by sentence ordinal rather than character windows.
In progress
2026
Evidence benchmark study Research
Comparing sentence-window retrieval with parent-document, auto-merging, PaperQA2, GraphRAG, LightRAG, RAPTOR, and HippoRAG on public scientific QA datasets.
In progress
03 / PUBLICATIONS
Selected publications
Papers are currently under review. This section will list accepted papers and preprints after publication, with DOI, code, and evidence.
04 / NOTES
Working notes
FIELD NOTE · 01
Why sentence IDs matter
Chunk IDs make retrieval easy; sentence IDs make evidence checkable. The gap between the two is where citations go to die.
FIELD NOTE · 02
Benchmarks are data cards
Before comparing methods, write down what the evidence actually is: a sentence, a paragraph, or a document.
FIELD NOTE · 03
Fail closed, then fail loudly
Silent fallbacks between data layers are worse than a visible error. A wrong answer with a citation is still a wrong answer.
05 / CONTACT
Let’s compare notes.
Open to collaborations on HBV/HCC data, bioinformatics, scientific agents, and evidence-grounded retrieval.