Generative parking-trace pipeline
Use segmentation, mask conditioning, and inpainting to insert or remove complex objects. The pipeline controls temporal and geometric artifacts in the resulting training data.
Systems, experiments, and applied machine learning
The diagrams show how each project works. Open a case study for the implementation, design choices, and limits of what it demonstrates.
Interactive system map
Switch between projects to compare the pipelines. Open a case study to see the full diagram and why I built it that way.
Graph AI / precision oncology
The pipeline starts with mutations and produces candidate peptides, graph representations, binding estimates, and a ranked shortlist.
Move over a node to highlight its connections. Choose a project, then open its case study for more detail.
Case-study directory
Each page has an interactive diagram, an explanation of the pipeline and design choices, and a link to the repository.
Reconstruct sequences from variants, build peptide graphs, and rank candidates by MHC binding.
Open case study 02 / Adaptive networksA model that trains in a fast loop and reorganizes its functional structure in a slower one.
Open case study 03 / Deep RLEstimate whether an RL policy suits a changed environment before it fails there.
Open case study 04 / Volumetric visionBuild hierarchical 3D representations of an MRI volume to label anatomy at each voxel.
Open case study 05 / Medical imagingKeep the information from individual MRI slices, then combine it to classify the whole study.
Open case study 06 / Computational biologyUse expression and clinical data to estimate risk and produce research reports.
Open case study 07 / Developer systemsBuild a workstation from pinned inputs and reusable modules, with encrypted secrets and checks before activation.
Open case studyProduction work
I describe the architecture here because the data and implementation are private.
Use segmentation, mask conditioning, and inpainting to insert or remove complex objects. The pipeline controls temporal and geometric artifacts in the resulting training data.
Define workflows and route tasks across cloud and on-premise compute, with visible failures, clear retry behaviour, and scheduling that accounts for available resources.
Restructure algorithms and memory access for large spatial workloads, taking compiler behaviour into account with JAX and Pallas-style kernels.
Smaller builds
I've also put JAX training exercises, GPU experiments, forecasting studies, Kaggle work, and a terminal typing game in Rust on GitHub.