Behavioral Machine Learning — Sternson Lab
Mark observation intervals on the timeline and compare against lab labels. 8,200 frames · 0.85 mean ROC-AUC.
Work
Try demos on this page. Each case study still has the full problem, contribution, approach, limitations, and what I learned.
Try the demos here — then open the full case study for depth.
Mark observation intervals on the timeline and compare against lab labels. 8,200 frames · 0.85 mean ROC-AUC.
Tap a scenario — see Pass / Review / Blocked before opening the walkthrough.
Operator view
Missing supplier ID
Deterministic rules decide the verdict. The explainer only translates evidence — traced with Langfuse.
June 2026 – Present
Open walkthroughDoes the model actually use the tissue concept it claims? Click a claim.
Evidence card
Tumor epithelium
TUM vs NORM
Necessity
0.46
Sufficiency
0.51
Specificity
0.80
High specificity + necessity → prediction depends on the named concept, not noise.
K-Scope · Phikon-v2
Open full demoEnd-to-end analyses with results you can inspect.

Clean Macro F1
0.86
TF-IDF vs LSTM vs DistilBERT under synonym, typo, and word-drop noise.

ROC-AUC
0.884
Decode left vs right choice from Neuropixels population activity.
Live UI previews — open a demo or the full case study.
Context-aware personal AI — RAG + structured user state for proactive nudges.
Interactive exploration of pathology-style features and adaptive extraction paths.
Analytics, campaigns, and other team collaborations.

AAF District 15 campaign — strategy, creative, and delivery.

CAISY beta conversation analytics → prompt & targeting recommendations.