Milestones
Key improvements, architecture changes, and deployments.
Grid Detection v4: GNN Classifier + Two-Stage Reranker
Architecture ChangeGrid Detection
Added GATv2 graph neural network for page-level fiducial filtering. Combined with two-stage MLP scorer + attention reranker. Fiducial precision jumped from 78.9% to 93.5%, false positives cut by 77%.
Before
86.6%
v3-mlAfter
89.3%
v4-ml-gnnImprovement
+2.7%
Grid Detection v4-ML: 14-Feature Tracer with Anchor Attention
Architecture ChangeGrid Detection
Replaced heuristic tracer with learned two-stage model: MLP scorer for O(N) candidate filtering + self-attention reranker over top-K candidates with GRU history context. Trace F1 improved from 78.9% to 89.5% (+10.6%).
Before
86.6%
v2-mlAfter
86.6%
v4-ml-no-clfImprovement
0.0%
Grid Detection Ground Truth: 18 Projects Annotated
Data UpdateGrid Detection
Created comprehensive ground truth dataset using the Fiducial Labeler tool. 18 annotated pages across projects: villanova, yampa, stloius, wpd, siu, science-build, outback, newman, milford, merck, lodge.