Generative parking-trace pipeline
Segmentation, mask conditioning, and inpainting for inserting or removing complex objects while controlling temporal and geometric artifacts in downstream training data.
Systems, experiments, and applied machine learning
Start with the map. Open the system that interests you.
The atlas gives a fast architectural view. Each project now opens into its own case study with a dedicated diagram, implementation narrative, and evidence boundary.
Interactive system map
Seven different computational shapes.
Switch projects to compare their pipelines, then open the selected case study for the complete diagram and design story.
Graph AI / precision oncology
Mutations become candidate peptides, graph representations, binding estimates, and a ranked shortlist.
Tip: move across a node to focus its connections, or select a project to open its dedicated case study.
Case-study directory
Choose a system to go deeper
Every page includes a project-specific interactive diagram, the actual pipeline, key design choices, and a direct repository link.
01 / Graph AI
Personalized cancer vaccine design
Variants → sequence reconstruction → peptide graphs → MHC ranking.
Open case study 02 / Adaptive networksFORDE vision–language model
A fast training loop coupled to slower functional reorganization.
Open case study 03 / Deep RLDomain-shift quantification
Estimate policy suitability before a changed environment exposes failure.
Open case study 04 / Volumetric visionCardiac MRI segmentation
Hierarchical 3D representations for voxel-level anatomy masks.
Open case study 05 / Medical imagingMRNet MRI classification
Preserve slice evidence before aggregating a study-level decision.
Open case study 06 / Computational biologyLUAD survival modeling
Expression and clinical data converge on risk and research reports.
Open case study 07 / Developer systemsNixOS reproducible workstation
Pinned inputs, composable modules, encrypted secrets, checked activation.
Open case studyProduction work
Systems built under real constraints
These projects remain at the architectural level because their data and implementation are private.
Airflow and Kubernetes execution stack
Workflow boundaries and task routing across cloud and on-premise compute, designed around failure visibility, retry behaviour, and resource-aware scheduling.
JAX kernels for height maps and point clouds
Algorithmic restructuring and compilation-aware memory access for high-volume spatial workloads using JAX and Pallas-style kernels.
Smaller builds
Experiments beyond the main case studies
JAX training exercises, GPU experiments, forecasting studies, Kaggle work, and a terminal typing game in Rust remain available on GitHub.