WHO WRITES THIS
Computational biologist & ML engineer —
multi-omics, statistical genetics, LLM systems
Most of what is published here comes out of a working dry lab, not a content plan. The pipelines are real, the failures are real, and the numbers are measured on hardware I own. When something did not work, that is usually the post.
Things I have built
Open source, on GitHub. Each one exists because I needed it for actual work.
A private, fully-local RAG over your own PDFs: BGE-M3 + embedded Qdrant + a local LLM through Ollama. About 150 lines. Nothing leaves the machine.
A compact MOGONET-style multi-omics classifier — per-omics GCN with attention fusion. An honest reference implementation, including where it scores near-random.
Adaptive Reasoning for Integrated Analysis — an LLM-driven framework for transcriptome analysis.
A verification toolkit for AI-assisted analysis: held-out and determinism gates, plus a judge from a different model than the one being checked.
A free handbook on running local LLMs on modest hardware, building a private paper-RAG, and evaluating them honestly.
An open catalog of 88 reproducible bioinformatics methods. Each entry: input, pipeline, tool versions, output, limitations, QC, references.
More at github.com/shoo99.
If you are new here
These are the posts that show how the work actually goes, failures included.
It Almost Worked: Four Ways a Multi-Omics Pipeline Nearly Reported a False Success
Counting the same 110 patients twice, a replication guard that passed when nothing had replicated, and eight vascular genes that were one signal.
The Number That Survived a Full Day of Review
Two agents with deliberately different access caught six errors. A seventh survived both, because it arrived as prose.
DIA-NN Complete Tutorial
Raw files to a quantified proteome, including the parameters that matter and the mistakes I made first.
Getting in touch
I am glad to hear from people working on similar problems — multi-omics integration, proteomics pipelines, or running LLMs locally on private research data. Questions about a post, corrections, and collaboration are all welcome.
If you found an error in a post, say so plainly — corrections get published rather than quietly fixed.