Research Blog

A larger follow-up to the first post will run on a DGX Spark (aarch64), so I first tested whether its evaluation reproduces there, re-running three of its encoders on the original laptop (x86_64) and on the Spark. The cosine-geometry metrics (Layer 1) reproduced to four decimals across the two machines. The Mapper layer did not, and the check exposed an error of mine: the first post's anchor ρ is a single, badly sampled draw. A per-seed estimator with standard errors brings the machines into agreement across all 15 of the first post's evaluation rows and corrects one of its claims.

I ran 14 open-weights text encoders (15 evaluation rows; EmbeddingGemma was run twice, once with the same config as the other models and once with its prescribed task prefix). Here I use multi-layer evaluation crtieria. Specifically, cosine geometry plus a 25-seed UMAP bootstrap into KeplerMapper, applied to a biomedical corpus (SciCUEval) and a general-knowledge corpus (MMLU non-STEM). The cosine ranking rewards aggressive contrastive training. The topology layer reorders everything. The contrastive-biomedical encoder MedCPT wins continuous-anchor faithfulness outright. PubMed-pretrained MLMs (masked language models) that 'collapse' in cosine space recover clean topological structure. Architecture barely moves the needle; training regime does.

Academic principal investigators are often described as running 'small businesses' because they manage budgets, hire staff, and sell a vision. As someone who runs an actual small business, I find the analogy superficially appealing but structurally misleading. This post lays out where the comparison breaks down: who bears risk, who the customer is, what failure costs, and how time horizons differ. It is not a takedown of academia or of entrepreneurship, both of which are demanding in their own right, but an argument for retiring a comparison that flatters neither side when examined closely.

Self-hosting—running your own services on hardware you control—offers a DIY path to digital autonomy. From firewalls and home automation to NAS systems and git servers, the spectrum of possibilities is vast. This post explores what self-hosting means, shares personal experiences from academia to home labs, and honestly examines the trade-offs. Whether you want to fight platform enshittification, own your data, or just tinker with hardware for fun, self-hosting rewards curiosity and patience in equal measure.

The bioinformatics community faces a sustainability crisis: graduate students must publish novel tools to graduate, but no one funds long-term maintenance. This creates an ecosystem of abandonware that breaks microservices architectures and wastes researcher time. Without wholesale funding reform, we need to raise our engineering standards and lower the bar for what counts as 'publishable.' This post examines what bioinformatics can learn from mature open source communities and offers practical steps toward more sustainable software development.

In fields like microbial ecology, agriculture, and molecular biology, the practice of 'waiting to see what the data say' is often justified as embracing discovery. But without rigorous upfront design—particularly computational modeling to guide measurement strategies and statistical approaches—this flexibility masks p-hacking and arbitrary choices. This post argues that modeling, preregistration, and honest distinction between exploratory and confirmatory work aren't constraints on discovery; they're prerequisites for credible science. Targeted at computational professionals and data scientists, it offers practical guidance for designing rigorous studies, collaborating with wet-lab teams, and building institutional cultures around methodological transparency.

An exploration of language maximalism in scientific computing and the benefits of strategic polyglot programming approaches

This inaugural post introduces my research focus areas and the types of content you can expect from this blog, covering bioinformatics methodologies, data science writ large, and computational methods in biotechnology.