MLOps engineer passionate about machine learning, energy forecasting, and building high-performance systems. I write about the things I learn.
A May 2026 survey of long-term single-object visual tracking for autonomous cinematography drones: what DJI and Skydio quietly ship, what CVPR…
A May 2026 survey of visual-inertial odometry for drones: where production stacks have consolidated, where foundation-3D research is going, what the…
A May 2026 survey of cross-view drone-to-satellite visual geo-localization: what NVIDIA Isaac ROS ships, what CVPR 2026 published, what the MAVLink…
Deploying ONNX vision models on Apple Silicon via CoreML. From 41% slower than CPU to 28 FPS real-time through compute-unit tuning, ONNX graph…
Taking ONNX models from PyTorch CUDA to 97 FPS with TensorRT mixed-precision engines. Covers CUDA execution providers, engine building, FP16…
A step-by-step guide to profiling PyTorch inference bottlenecks and exporting vision models to ONNX Runtime. Uses SAM 2 / EfficientTAM as a…
The story of building a gradient boosting library in Rust - from dissecting XGBoost models to achieving performance parity and beyond.
A practical guide to configuring gradient boosting: learning rate, tree depth, regularization, and systematic tuning strategies.
How decision trees fit into gradient boosting: from second-order Taylor expansion to the split gain formula that powers XGBoost and LightGBM.
When to choose XGBoost vs LightGBM: speed, accuracy, features, and practical recommendations based on your use case.
How LightGBM bundles sparse features to reduce memory 25×, and how both libraries handle categorical data without one-hot encoding.
How LightGBM speeds up training by keeping high-gradient samples and subsampling the rest. Plus row and column sampling strategies.
XGBoost grows trees level-by-level, LightGBM splits the best leaf. Learn when each strategy wins and how num_leaves relates to max_depth.
The optimization that makes gradient boosting fast: how histogram-based training reduces split finding from O(n log n) to O(bins) with the…
How gradient boosting optimizes in function space: from pseudo-residuals to the complete boosting algorithm, with derivations for common loss…