Apppublic

manzz05/kitchenflow-v1

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes

App README

KitchenFlow-v1 — Ghost Kitchen Dispatcher

An OpenEnv environment where an AI agent acts as the dispatch brain of a ghost kitchen — deciding the perfect moment to summon a delivery driver so they arrive exactly when the food is bagged.

Everyone knows the pain of cold fries. Now you can train an AI to prevent it.

The Problem

In busy cloud kitchens, food often sits getting cold because the driver isn't there yet — or drivers idle for 20 minutes because food isn't ready. The agent must watch food prep progress and real-time traffic to time each dispatch perfectly.

Observation Space

Each step (= 1 minute) the agent sees:

Field	Type	Description
`time_min`	int	Current simulation minute
`traffic_index`	float	Road congestion (1.0 = free flow, 2.5 = gridlock)
`orders[].food_prep_progress`	0.0–1.0	How close the food is to being bagged
`orders[].driver_dist_km`	float	Distance of the assigned driver
`orders[].food_temp_c`	float	Current food temperature (°C)
`orders[].driver_eta_min`	int/null	Minutes until driver arrives (null if not summoned)
`orders[].driver_summoned`	bool	Has the driver been called?

Action Space

json

{"dispatch_decisions": {"ORD001": 1, "ORD002": 0, "ORD003": 1}}

—0 = Wait this minute
—1 = Summon driver (one-time trigger — driver heads to hub immediately)

Reward Function

Event	Reward
Driver arrives ≤ 2 min of food being ready	+10 (perfect timing)
Driver arrives ≤ 5 min of food being ready	+5 (good timing)
Each °C below 75°C at delivery	−1 (cold food)
Driver waits > 15 min (risk of cancellation)	−20
Food waits > 10 min without driver	−5
Order not resolved in time	−10 (failed)

Scores are normalised to 0.0–1.0 per order, then averaged.

Tasks

Task ID	Difficulty	Scenario	Orders	Time Limit
`T1_single_order_dispatch`	Easy	1 burger, stable traffic (1.2)	1	30 min
`T2_multi_order_coordination`	Medium	3 orders, fluctuating traffic	3	35 min
`T3_peak_hour_rush`	Hard	5 orders, traffic spike 10–20min (up to 2.2)	5	45 min

Dispatch Physics

driver_speed = 0.5 km/min ÷ traffic_index   (= 30 km/h in free flow)
driver_eta   = driver_dist_km ÷ driver_speed

optimal_dispatch_minute = food_ready_minute − driver_eta

Baseline Scores (80% Threshold Heuristic)

"Summon driver when food_prep_progress ≥ 0.80"

Task	Score
T1 easy	1.00
T2 medium	0.96
T3 hard	0.84
Average	0.93

A well-tuned LLM that reasons about ETA calculations should beat 0.93.

API

bash

# Start T1
curl -X POST https://your-space.hf.space/reset \
  -H 'Content-Type: application/json' \
  -d '{"task_id": "T1_single_order_dispatch"}'

# Step (dispatch decisions, with episode_id from reset)
curl -X POST https://your-space.hf.space/step \
  -H 'Content-Type: application/json' \
  -d '{"action": {"dispatch_decisions": {"ORD001": 1}}, "episode_id": "<from_reset>"}'

Run Baseline

bash

export HF_TOKEN=hf_...
export MODEL_NAME=meta-llama/Llama-3.3-70B-Instruct
export API_BASE_URL=https://router.huggingface.co/v1
export ENV_URL=https://your-space.hf.space
python inference.py

Local Development

bash

pip install -r requirements.txt
uvicorn app:app --host 0.0.0.0 --port 7860
# Visit http://localhost:7860 → auto-redirects to Swagger UI

OpenEnv Validation

bash

openenv validate .
# [OK] : Ready for multi-mode deployment