2026 · Recommendation, simulation, evaluation & UI
FoodFlow: Recommendation & Fulfillment Simulation
We connect Top-K recommendation to merchant exposure, courier matching, and peak-hour simulation, measuring accuracy, fairness, ETA, timeout rate, and platform utility together.
- 0.4509
- Seq-Tuned Recall@20
- 24.98 min
- KG + Batch mean ETA
- 0.5697
- best mean platform utility
Problem
Offline recommendation gains may create concentrated exposure, overloaded merchants, courier imbalance, and delivery delays downstream.
Approach
We implemented eight recommenders/rerankers, serviceability constraints, KG explanations, batch and route-aware dispatch, multi-step simulation, and a shared Streamlit workbench.
Outcome
All 60 tests and the full mock pipeline pass. Seq-Tuned leads accuracy, Logistic-LTR improves coverage and exposure, and KG-Tripartite + Batch has the best current mean platform utility.