Apppublic

manzz05/kitchenflow-v2

sourceHugging Facemitupdated 5mo agoView on Hugging Face

1likes

App README

KitchenFlow-v2 — Ghost Kitchen Dispatcher

🚀 Training Notebook (Colab): https://colab.research.google.com/drive/1Sfvju98sWX4HEDqJczHpGSiK1wSux2LW?usp=sharing 🔗 Live Demo: https://huggingface.co/spaces/manzz05/kitchenflow-v2 📝 Blog Post: https://huggingface.co/spaces/manzz05/kitchenflow-v2/blob/main/blog.md

Everyone knows the pain of cold fries. KitchenFlow-v2 trains an AI to prevent it.

Training Results

Reward Curve

[image]

Before vs After

[image]

The Problem

Ghost kitchens lose money every day to the order-ready mismatch. Summon a driver too early — they wait 20 minutes and may cancel (−20 penalty). Summon too late — food drops from 72°C to 55°C (−1 per degree). The agent must learn perfect timing under real-time uncertainty.

What's New in v2

—Multi-agent: KitchenCoordinator sub-agent broadcasts prep signals
—Chaos events: Driver cancellations, traffic surges, GPS corruption
—Expanded action space: 4 actions including priority requests and GPS requery
—5 tasks: T1–T5 from easy single order to enterprise lunch rush
—Richer rewards: Sync bonus, rush completion, chaos recovery bonus

Tasks

Task ID	Difficulty	Orders	Chaos	Time Limit
`T1_single_order_dispatch`	Easy	1	None	30 min
`T2_multi_order_coordination`	Medium	3	None	35 min
`T3_peak_hour_rush`	Hard	5	None	45 min
`T4_chaos_dispatch`	Medium	1	1 event	30 min
`T5_enterprise_lunch_rush`	Hard	5	2 events	60 min

Action Space

Code	Action	Description
0	`wait`	Do nothing
1	`summon_driver`	Call nearest driver
2	`request_priority`	Ask coordinator to speed up prep
3	`requery_gps`	Use when GPS data is stale

Reward Function

Signal	Value	Trigger
`delivery_base`	+50	Order delivered
`sync_bonus`	+10	Driver arrives ≤ 2 min of food ready
`rush_bonus`	+15	All orders done in ≤ 75% of max s