Occasional notes on things I’ve built or had to work through.
agent
lakehouse
The agent's abilities are written down as skills, and every project registers its research plan before any query runs.
Features computed from pairs of genome-scale metabolic models add predictive signal beyond growth and SMETANA baselines, and triple the positive yield of the top screening decile.
Which parts of simulating a cell are hard in a way a quantum computer might actually help with, and which are not.
Cross-dataset prediction of carbon utilisation is limited by disagreement between the experimental label and the annotated mechanism, not by taxonomy or training-set size.
Running a stack of genome annotation tools and getting tables back that actually join.
one store
Modelling genomics, phenotyping and photosynthesis into one star schema, with LinkML as the source of truth.
Simulating reaction systems small enough that randomness decides the outcome, and benchmarking the result against the packages that already existed.
Every step of a 16S workflow has several defensible tool choices, and the inferred network changes with them.
Why networks from different microbiome studies refuse to line up, and what a shared schema plus taxonomy projection makes possible.
A reinforcement learning agent sets the feed of a simulated chemostat community. It held the community in place and did not reach new targets.
