I build tools for understanding how machine learning systems behave in practice — how they scale, where they regress, what they fail to evaluate, and how their behavior can be measured more reliably.
My current work spans robotics data systems, performance regression infrastructure, retrieval evaluation, distributed training, and transformer inference.
Tools for exploring state-space and workspace coverage in LeRobot datasets.
The project uses robot state trajectories and URDF-based forward kinematics to analyze workspace coverage, episode behavior, under-explored regions, and dataset structure through interactive 3D visualizations.
Performance regression infrastructure for pull requests.
It compares benchmark evidence, identifies material regressions, produces bounded repair guidance, and verifies fixes through a protected benchmark gate while leaving the final merge decision to a human.
Evaluation infrastructure for search and retrieval systems.
It provides reproducible benchmark datasets, TF-IDF, BM25 and hybrid retrieval baselines, ranking metrics, weak-query analysis, benchmark artifacts, and regression detection across runs.
An ML systems project built from first principles to study distributed training, transformer inference, KV-cache behavior, memory scaling, serving, observability, and performance.
| Project | Focus |
|---|---|
| Distributed Training Profiler | Communication overhead, scaling efficiency, memory use, and distributed-training bottlenecks |
| Benchmark Guardian | Benchmark regression detection for GitHub pull requests |
| ML Reproducibility Auditor | Reproducibility and engineering-quality analysis for ML projects |
| Distributed Training Simulator | Data-parallel scaling and all-reduce communication analysis |
| Autograd Engine | Reverse-mode automatic differentiation with dynamic computation graphs |
- ML systems and AI infrastructure
- Distributed training and communication
- Transformer inference and memory behavior
- Performance benchmarking and regression detection
- Search and retrieval evaluation
- Robotics data systems
- Reproducibility and observability
I prefer to measure systems before optimizing them, make assumptions explicit, and keep performance and evaluation claims backed by reproducible evidence.