Systems thinking, model evaluation, and build notes from applied ML work.
Research paper
A Distill-style article adapted from the paper, with spatial leakage control, two-level validation, SHAP interpretability, seasonal UHI tables, and the original PDF.
Open research articleAttribution sketch
A lightweight SVG graph for posts about model behavior: prompt tokens at the base, feature groups in the middle, and the target decision at the top.
Benchmark surface
Pipeline trace
Benchmarking SegFormer variants (B0-B5) on NVIDIA Jetson Nano and comparing inference latency vs. mIoU. Key findings on quantization strategies and memory bandwidth bottlenecks.
How we achieved reproducible experiments and zero-downtime model updates using DVC for versioning, FastAPI for serving, and Prometheus for observability.
A deep dive into failure modes of YOLOv8 on glass and metallic surfaces. We trained a specialized augmentation pipeline and reduced false negatives by 34%.
Side-by-side evaluation of monocular depth models on real-world terrain navigation tasks. Depth Anything V2 showed superior edge preservation but higher latency.
Private