Aditya Pandey
I work on the performance of numerical and machine learning software: GPU kernels, compiled training loops, automatic differentiation, and the I/O that feeds them.
I'm currently consulting for Evovest on their open-source machine learning stack, and I'm open to contract work on GPU and performance problems.
Work
- EvoTrees.jl Evovest
Rewrote the GPU backend of a gradient-boosting library with KernelAbstractions.jl, so the same kernels run on NVIDIA, AMD and Apple GPUs.
- NeuroTabModels.jl Evovest
Rebuilt training for a tabular deep learning library on Lux.jl and Reactant.jl, and added TabM, ModernNCA and numerical embeddings.
- ParallelParticleSwarms.jl SciML
Restructured the GPU kernels of a particle swarm optimizer so they scale with problem size, and rewrote its L-BFGS local search.
- RAMTools CERN-HSF
Built tools for RAM, a ROOT-based format for genomic alignment data, during Google Summer of Code 2025: multithreaded conversion from SAM and BAM, and index-free region queries.
Things I can help with
- Porting CUDA-only code to kernels that run on any GPU vendor
- Profiling and restructuring GPU kernels
- Compiling and differentiating models through XLA and Enzyme
- Multithreaded I/O and columnar data formats
- Benchmark suites that catch regressions
The background is a particle swarm searching a randomly generated landscape for its lowest point, the method ParallelParticleSwarms.jl runs on GPUs.