SENA
SAYGIN ŞENYÜZ

An end-to-end ML pipeline over 180,519 orders. Three columns that are only filled in after delivery (Delivery Status, Days for shipping (real), shipping date) were detected and removed, taking the feature set from 53 to 25. With the leakage gone the baseline F1 is 0.39; hyperparameter and threshold tuning takes it to 0.71. Train 0.713 / test 0.714 — no overfitting. The threshold was measured at four points (0.35–0.55). Data ceiling: added features left F1 unchanged — the true drivers (weather, traffic, carrier reliability) are absent from the dataset.

A 33-node agentic workflow: classifies incoming mail by urgency and document type, archives attachments, and answers questions by querying Notion as a tool. A memory-backed LLM agent with a command interface.

AI4I 2020 dataset; 10,000 records at a 3.4% failure rate — an imbalanced problem. Target leakage was detected and removed, then logistic regression, a decision tree, random forest and XGBoost were compared under stratified 5-fold cross-validation. Two physics-based features derived from the dataset documentation — temperature delta and mechanical power — lifted F1 from 0.66 to 0.85 and cut missed failures from 32 to 15. Model behaviour was validated with SHAP, and the demo runs in the browser.

Connects WO-2601's model to a system that decides — the OP-25 section above is it, running live. The 200 trees were ported to plain JavaScript and run at the edge; across 5,000 test rows they differ from the Python original by 1.9·10⁻⁷. It picks the threshold from the planner's cost structure and says when it is not earning its keep: a four-row table on Shipping Mode alone makes the same call on 99.03% of orders.

Cost function, gradient and the update rule written by hand — no library shortcut.