Projects
Case studies that show decisions, pivots, and outcomes
A selection of end-to-end machine-learning work, with the process and tradeoffs left visible.
01
FlyRank — Content Decline Triage
Detecting observed content decline and ranking editorial review opportunities. ROC AUC 0.77 with 100% precision among the top 50 held-out items.
0.77 ROC AUC144K items
See the case study
02
LLaMA-2 Fine-Tuning for Code Generation
Trained with QLoRA using constrained compute and private-data requirements. Includes failure analysis and metric-based evaluation.
+19.4% Pass@55.1 GB VRAM
See the case study
03
Landslide Image Segmentation
Self-labeling sparse satellite imagery and building a segmentation workflow for low-data, high-variance terrain detection.
5,843 reviewed164 masks
See the case study