Generative parking-trace pipeline
Pipeline for complex object insertion and removal over parking video traces using segmentation, masking, and inpainting to support downstream model development.
Research engineering, production systems, and applied ML
Work that sits between algorithm design and practical deployment.
Some projects are research-heavy, some are infrastructure-heavy, and some are purely applied. The interesting ones usually involve all three.
Production and Systems
Operational work
The work closest to deployment usually has the hardest constraints and the most useful feedback loops.
Airflow and Kubernetes orchestration stack
Reworked DAGs, improved dependency behavior, and designed scalable task routing across cloud and on-prem compute for data extraction and processing.
JAX optimization for height maps and point clouds
Ported processing paths to JAX and Pallas-style kernels to cut expensive workloads down dramatically and make throughput less hardware-specific.
Research and Manuscripts
Projects with a stronger theory component
Block-wise differentiable Sinkhorn attention
Designed a memory-efficient long-context attention mechanism with a custom multi-plan backward pass and linear-memory scaling target.
Intrinsic quantification of domain shift magnitude
Developed methods for measuring domain shift through reactive exploration and evaluated them across OpenAI Gym environments.
Copilot semantic assessment initiative
Combined NLP techniques, topic modeling, and semantic accuracy analysis to evaluate model usefulness and behavior in developer-facing contexts.
Applied ML
Bio, vision, and modeling projects
Personalized cancer vaccine design
End-to-end pipeline from TCGA mutations to neoantigen ranking using peptide-MHC prediction, graph methods, and biological validation signals.
Cardiac MRI segmentation with 3D SegFormer
Implemented a 3D MixTransformer-style encoder with overlapping patch embeddings and multi-scale decoding for volumetric segmentation tasks.
MRNet MRI classification with slice attention
Combined patch tokenization, self-attention, and inter-slice aggregation to build volume-aware diagnostic classifiers for MRI studies.
ODE modeling for drug response prediction
Analyzed regulatory network output and integrated ranked interactions into downstream survival and response modeling workflows.
Forecasting model benchmarking and SaaS transition
Benchmarked against strong public baselines and delivered alternative models while helping move forecasting capability into a more scalable product shape.
NixOS configuration and reproducible environments
Ongoing system configuration work around flakes, developer environments, container tooling, and maintaining a machine that stays pleasant to use.