13 of 13 projects
A whole-slide benchmark on 2,099 public TCGA slides showed that a frozen foundation encoder plus mean pooling and a linear head recovers receptor status and HRD from H&E. A more complex learned-attention aggregator did not beat that simple baseline: it piled nearly all of its weight onto ~10% of tiles and gained no accuracy for it.
A full-stack biomedical AI agent that routes scientific requests across 13 genomic model configurations, 23 governed tool packages, and 15 workflow skills while enforcing plan approval, provenance, evidence separation, and research-use boundaries.
A survey of 40 genomic models and a small hands-on comparison of three, on how training objective, context, outputs, and operational cost shape which model suits a given genomics workflow.
A dual-model workflow used AlphaGenome and Enigma to add functional evidence to variants missed by conventional high/moderate consequence labels, expanding the reviewable space while preserving clinical guardrails.
A structural-variant encoding study translated deletions, duplications, inversions, breakends, and gene fusions into matched REF/ALT DNA windows for AlphaGenome and Enigma.
A group-held-out study across public multi-caller data and a de-identified matched FF/FFPE cohort showed that read-level technical evidence—not DNA sequence alone—is the key signal for FFPE artifact-risk modeling.
Research & Analysis
Evaluating AlphaGenome's pre-computed variant effect scores against TWAS associations across 10 GTEx tissues reveals weak but real directional signal, chromatin-driven fine-mapping enrichment, and fundamental limits of sequence-based expression prediction
ML Engineering & Research
A hands-on journey through the complete LLM pipeline—tokenization, pretraining, distributed training, and fine-tuning—inspired by Andrej Karpathy's nanochat. Training a 561M parameter conversational model on 11.2B tokens to understand different aspects of modern language models.
Research & Analysis
Evaluating 8 SSL methods with 3 GNN architectures on large-scale mouse embryo spatial transcriptomics data reveals interesting performance of reconstruction-based approaches
Research & Analysis
A systematic evaluation of 8 SSL methods with 3 GNN architectures on mouse brain spatial transcriptomics data
Research Documentation
An exploration of modern self-supervised learning methods including SimCLR, BYOL, SimSiam, Barlow Twins, DINO, etc, with mathematical foundations and mutual information theory
Research & Implementation
A comprehensive guide to understanding and applying GNNs to spatial transcriptomics data, from foundational concepts to cutting-edge research implementations
Design and engineering
This site — a research portfolio with a grounded AI assistant, a local search index, and a motion system built to stay out of the reader's way