Aster Waste Forecast

Aster Waste Forecast

$25.97
$25.97
Skip to product information
Aster Waste Forecast
  • Aster Waste Forecast
  • Aster Waste Forecast
  • Aster Waste Forecast
  • Aster Waste Forecast
Aster Waste Forecast

Aster Waste Forecast

$25.97
$25.97

Aster Waste Forecast is an AI-powered food-demand and waste-prediction solution for hospital kitchens, cafeterias, and healthcare dining operations.

It predicts how much food should be prepared for each meal period, menu item, patient unit, and service location. The goal is to reduce overproduction, control food costs, and improve sustainability without creating shortages or affecting patient care.

Aster can generate forecasts for:

  • Number of meals required by breakfast, lunch, and dinner

  • Demand for each menu item

  • Patient meal demand by clinical unit

  • Cafeteria demand from employees and visitors

  • Dietary meal demand, such as diabetic, low-sodium, renal, gluten-free, vegetarian, and texture-modified meals

  • Expected leftovers and discarded food

  • Ingredient-level consumption

  • Spoilage risk for perishable inventory

  • Recommended preparation quantities

  • Recommended purchasing and replenishment levels

For example, Aster may predict that the kitchen needs:

  • 240 standard lunch meals

  • 58 low-sodium meals

  • 32 diabetic meals

  • 18 texture-modified meals

  • 120 cafeteria portions of grilled chicken

  • 75 portions of the vegetarian option

The system can also provide a confidence range rather than a single number, such as recommending that the kitchen prepare between 115 and 125 portions.

Data sources

Aster combines operational, clinical, environmental, and historical data.

Hospital operational data

  • Current patient census

  • Expected admissions and discharges

  • Unit-level occupancy

  • Scheduled procedures

  • Emergency department volume

  • Staff schedules

  • Visitor trends

  • Hospital events and meetings

  • Holiday schedules

Food-service data

  • Historical meals prepared

  • Meals served

  • Plate waste

  • Kitchen production waste

  • Leftover quantities

  • Menu rotation

  • Recipe information

  • Portion sizes

  • Ingredient usage

  • Inventory levels

  • Supplier delivery schedules

  • Food expiration dates

  • Item substitutions

  • Cafeteria transaction history

Patient and dietary data

Only the minimum necessary information should be used.

Relevant inputs may include:

  • Diet order category

  • Unit or floor

  • Meal eligibility

  • NPO status

  • Allergy restrictions

  • Texture requirements

  • Meal cancellations

  • Patient discharge timing

It does not need to expose patient names or unnecessary clinical details to the forecasting model.

External data

  • Weather

  • Seasonal trends

  • Local events

  • School schedules

  • Public holidays

  • Supply-chain disruptions

  • Supplier availability

Weather can influence cafeteria traffic, visitor volume, and preferences for certain menu items.

The Aster Feature Store

The feature store is the organized, governed layer that holds the predictive signals Aster uses for both model training and live forecasting.

Instead of rebuilding the same calculations every time, the feature store maintains reusable features such as average meal demand, current patient census, menu popularity, spoilage risk, and expected discharge volume.

It ensures that the features used to train the model are calculated in the same way as the features used when the model makes daily predictions.

Why the feature store matters

Without a feature store, one engineering team may calculate “average demand” using the previous seven days, while the production system uses the previous thirty days. That inconsistency can make a model perform well during testing but poorly in the hospital.

The feature store provides:

  • Consistent feature definitions

  • Reusable data across models

  • Faster model development

  • Historical feature tracking

  • Data quality controls

  • Point-in-time accuracy

  • Feature versioning

  • Real-time and batch access

  • Access controls

  • Auditability

Examples of features

Census features

These describe the expected number of patients requiring meals.

  • Current hospital census

  • Census by hospital unit

  • Census by dietary category

  • Expected admissions in the next 24 hours

  • Expected discharges before each meal

  • Average census for the same weekday

  • Seven-day census trend

  • Occupancy percentage

  • Number of patients currently NPO

  • Number of patients eligible for each meal

Example:

expected_lunch_meal_patients = 286
expected_discharges_before_lunch = 14
npo_patients_at_lunch = 22
renal_diet_patients = 31

Historical demand features

These describe what was previously prepared and consumed.

  • Meals served during the previous meal period

  • Average demand for the same weekday

  • Average demand for the same menu item

  • Seven-day moving average

  • Twenty-eight-day moving average

  • Seasonal demand

  • Demand during comparable holidays

  • Meal cancellation rate

  • Tray return rate

  • Cafeteria transaction volume

  • Demand volatility

Example:

grilled_chicken_7_day_avg = 116
grilled_chicken_same_weekday_avg = 122
grilled_chicken_last_served = 119
grilled_chicken_demand_volatility = 0.08

Menu features

These describe the meal being offered.

  • Menu item

  • Recipe category

  • Protein type

  • Vegetarian indicator

  • Sodium level

  • Calorie range

  • Portion size

  • Menu rotation position

  • Historical popularity

  • Historical waste percentage

  • Substitution history

  • Preparation complexity

  • Shelf life after preparation

The model may learn that a certain meal is consistently less popular on Fridays or that a particular side dish produces high plate waste.

Inventory features

These help Aster consider ingredient availability and spoilage.

  • Current ingredient quantity

  • Days until expiration

  • Historical spoilage rate

  • Average daily consumption

  • Open purchase orders

  • Supplier delivery date

  • Storage capacity

  • Ingredient substitution availability

  • Cost per serving

  • Lead time

  • Inventory turnover

Example:

spinach_quantity_on_hand = 42_kg
spinach_days_to_expiration = 2
spinach_average_daily_usage = 11_kg
spinach_spoilage_risk = high

It could recommend prioritizing menu items that use ingredients approaching expiration, provided dietary and operational requirements are still met.

Waste features

These measure waste across the food-service operation.

  • Food prepared but not served

  • Plate waste

  • Spoilage waste

  • Trim and preparation waste

  • Waste by menu item

  • Waste by kitchen station

  • Waste by hospital unit

  • Waste by meal period

  • Waste cost

  • Waste weight

  • Waste percentage

  • Avoidable versus unavoidable waste

Example:

vegetable_soup_overproduction_rate = 12%
vegetable_soup_plate_waste_rate = 7%
vegetable_soup_average_waste_cost = $84_per_service

Cafeteria traffic features

These predict employee and visitor demand.

  • Number of employees scheduled

  • Shift-change timing

  • Visitor count

  • Cafeteria transactions by hour

  • Day of the week

  • Hospital event schedule

  • Conference attendance

  • Promotional offers

  • Average transaction volume

  • Weather conditions

Time and calendar features

  • Hour of day

  • Meal period

  • Day of week

  • Month

  • Season

  • Holiday indicator

  • Weekend indicator

  • Days before or after a holiday

  • School vacation indicator

  • Menu cycle week

Cost and sustainability features

  • Cost per serving

  • Estimated waste cost

  • Disposal cost

  • Carbon estimate

  • Water-use estimate

  • Ingredient sourcing category

  • Local versus imported product

  • Compostable waste percentage

  • Donation eligibility

These features allow the hospital to optimize not only for quantity, but also for financial and environmental impact.

Feature-store structure

It can organize its feature store around several core entities.

Entity Example features
Hospital census, occupancy, total meal demand
Unit dietary mix, tray returns, meal cancellations
Meal period breakfast, lunch, dinner demand
Menu item popularity, waste rate, cost per portion
Ingredient inventory, expiration, spoilage risk
Cafeteria transaction volume, visitor demand
Date weekday, holiday, season, weather
Supplier lead time, fill rate, delivery reliability

This structure allows the model to retrieve the right features for a particular prediction.

For example:

Hospital: Main Campus
Date: July 14
Meal period: Lunch
Menu item: Grilled Chicken

The feature store would return the relevant census, menu, inventory, historical demand, and calendar features for that exact combination.

Batch features

Batch features are updated on a schedule, such as nightly or hourly.

Examples include:

  • Thirty-day average demand

  • Historical waste rate

  • Menu popularity score

  • Supplier reliability score

  • Ingredient turnover

  • Seasonal demand

  • Average patient census

These features do not need to update every second.

Real-time features

Real-time features reflect rapidly changing hospital conditions.

Examples include:

  • Current census

  • Newly recorded admissions

  • Recent discharges

  • NPO status changes

  • Last-minute diet-order changes

  • Ingredient shortages

  • Cafeteria transaction volume

  • Cancelled procedures

  • Unexpected patient surges

A real-time feature pipeline allows it to revise its recommendation shortly before food preparation begins.

For example, if twelve patients are discharged before lunch and five new patients are admitted, the forecast can be adjusted automatically.

Point-in-time correctness

The feature store should preserve what was known at the moment a historical forecast would have been made.

Suppose a patient was discharged at 2:00 p.m. The lunch forecast created at 9:00 a.m. must not use that later discharge information when training the model.

Otherwise, the model would receive information from the future and appear more accurate than it really is.

Point-in-time retrieval prevents this form of data leakage.

Feature versioning

Features should be versioned whenever their logic changes.

For example:

menu_popularity_score_v1

might use the previous 30 days of demand.

menu_popularity_score_v2

might use the previous 90 days and adjust for holidays.

Versioning allows the team to:

  • Compare model performance

  • Reproduce previous forecasts

  • Audit historical decisions

  • Roll back problematic changes

  • Safely introduce improved calculations

Feature quality monitoring

The system should continuously detect and flag data-quality issues such as missing values, delayed data feeds, sudden or unexpected changes in incoming data, invalid dietary categories, duplicate records, incorrect units of measurement, negative inventory values, inconsistencies in patient or customer census data, stale features that have not been refreshed, and distribution drift that indicates production data is no longer behaving like the data used to train the model.

For example, if the patient census feed normally updates every fifteen minutes but has not updated for two hours, Marigold should warn users and lower the forecast confidence.

Feature ownership and governance

Each feature should include metadata such as:

  • Feature name

  • Business definition

  • Data source

  • Calculation logic

  • Update frequency

  • Owner

  • Sensitivity classification

  • Expected range

  • Version

  • Creation date

  • Last validation date

Example:

Feature: expected_patient_lunch_demand
Definition: Estimated number of admitted patients eligible for lunch
Source: ADT feed and dietary ordering system
Update frequency: Every 15 minutes
Owner: Food-service analytics team
Sensitivity: Operational, de-identified
Expected range: 0–1,000

How Marigold uses the feature store

Hospital Systems
EHR · Dietary System · POS · Inventory · Scheduling
                         │
                         ▼
              Data Integration Layer
        APIs · HL7/FHIR · Files · Streaming Events
                         │
                         ▼
                  Marigold Feature Store
 ┌───────────────────────────────────────────────────────┐
 │ Census features                                      │
 │ Dietary-demand features                              │
 │ Historical consumption features                      │
 │ Menu popularity features                             │
 │ Inventory and expiration features                    │
 │ Cafeteria traffic features                           │
 │ Waste and sustainability features                    │
 │ Weather and calendar features                        │
 └───────────────────────────────────────────────────────┘
                  │                   │
        Historical features    Current features
                  │                   │
                  ▼                   ▼
          Model Training       Daily Forecasting
                  │                   │
                  └─────────┬─────────┘
                            ▼
              Recommended Production Plan
                            │
              ┌─────────────┴─────────────┐
              ▼                           ▼
       Kitchen Dashboard          Purchasing Dashboard

Daily operational workflow

At 4:00 a.m., Marigold loads the latest census, admissions, discharge expectations, diet orders, menu, inventory, and historical demand.

At 5:00 a.m., it generates an initial breakfast and lunch forecast.

Before production begins, it checks for:

  • Census changes

  • New dietary orders

  • Procedure cancellations

  • NPO changes

  • Ingredient shortages

  • Cafeteria staffing changes

It then updates the forecast and provides recommendations such as:

Prepare 118 grilled chicken portions.
Prepare 76 vegetarian pasta portions.
Reduce vegetable soup production by 14%.
Use spinach inventory first because 42 kg expires within two days.
Hold 8% reserve capacity for unplanned admissions.

Kitchen managers can accept, reject, or adjust the recommendation. Their actions should be captured and returned to the feature store as feedback.

Feedback features

Marigold becomes more accurate when it learns from operational outcomes.

Useful feedback features include:

  • Recommended quantity

  • Quantity approved by kitchen manager

  • Quantity actually prepared

  • Quantity served

  • Quantity wasted

  • Reason for manager override

  • Unexpected patient-volume change

  • Ingredient shortage

  • Menu substitution

  • Actual meal cancellation rate

The system can learn, for example, that kitchen managers regularly increase the recommendation for a popular dish or reduce quantities before holiday weekends.

Key performance indicators

Marigold should measure:

  • Food waste by weight

  • Food waste by cost

  • Overproduction percentage

  • Meal shortage rate

  • Forecast accuracy

  • Forecast bias

  • Cost per meal

  • Inventory spoilage

  • Plate waste

  • Production waste

  • Emergency food preparation

  • Kitchen labor efficiency

  • Carbon-emissions reduction

  • Dollars saved

  • Manager override rate

The most important balance is reducing waste without increasing shortages.

A successful model should not simply recommend preparing less food. It should recommend the right amount of food while protecting patient meal availability and service quality.

Suggested positioning

Aster Waste Forecast helps hospital kitchens predict meal demand, optimize production, use inventory before it expires, and reduce unnecessary food waste—while preserving patient nutrition, dietary compliance, and meal availability.

Tagline:
Predict every plate. Prepare with purpose. Waste less.

Marigold Waste Forecast

Marigold Waste Forecast is an AI-powered food-demand and waste-prediction solution for hospital kitchens, cafeterias, and healthcare dining operations.

It predicts how much food should be prepared for each meal period, menu item, patient unit, and service location. The goal is to reduce overproduction, control food costs, and improve sustainability without creating shortages or affecting patient care.

What Marigold predicts

Marigold can generate forecasts for:

  • Number of meals required by breakfast, lunch, and dinner

  • Demand for each menu item

  • Patient meal demand by clinical unit

  • Cafeteria demand from employees and visitors

  • Dietary meal demand, such as diabetic, low-sodium, renal, gluten-free, vegetarian, and texture-modified meals

  • Expected leftovers and discarded food

  • Ingredient-level consumption

  • Spoilage risk for perishable inventory

  • Recommended preparation quantities

  • Recommended purchasing and replenishment levels

For example, Marigold may predict that the kitchen needs:

  • 240 standard lunch meals

  • 58 low-sodium meals

  • 32 diabetic meals

  • 18 texture-modified meals

  • 120 cafeteria portions of grilled chicken

  • 75 portions of the vegetarian option

The system can also provide a confidence range rather than a single number, such as recommending that the kitchen prepare between 115 and 125 portions.

Data sources

Marigold combines operational, clinical, environmental, and historical data.

Hospital operational data

  • Current patient census

  • Expected admissions and discharges

  • Unit-level occupancy

  • Scheduled procedures

  • Emergency department volume

  • Staff schedules

  • Visitor trends

  • Hospital events and meetings

  • Holiday schedules

Food-service data

  • Historical meals prepared

  • Meals served

  • Plate waste

  • Kitchen production waste

  • Leftover quantities

  • Menu rotation

  • Recipe information

  • Portion sizes

  • Ingredient usage

  • Inventory levels

  • Supplier delivery schedules

  • Food expiration dates

  • Item substitutions

  • Cafeteria transaction history

Patient and dietary data

Only the minimum necessary information should be used.

Relevant inputs may include:

  • Diet order category

  • Unit or floor

  • Meal eligibility

  • NPO status

  • Allergy restrictions

  • Texture requirements

  • Meal cancellations

  • Patient discharge timing

Marigold does not need to expose patient names or unnecessary clinical details to the forecasting model.

External data

  • Weather

  • Seasonal trends

  • Local events

  • School schedules

  • Public holidays

  • Supply-chain disruptions

  • Supplier availability

Weather can influence cafeteria traffic, visitor volume, and preferences for certain menu items.

The Marigold Feature Store

The feature store is the organized, governed layer that holds the predictive signals Marigold uses for both model training and live forecasting.

Instead of rebuilding the same calculations every time, the feature store maintains reusable features such as average meal demand, current patient census, menu popularity, spoilage risk, and expected discharge volume.

It ensures that the features used to train the model are calculated in the same way as the features used when the model makes daily predictions.

Why the feature store matters

Without a feature store, one engineering team may calculate “average demand” using the previous seven days, while the production system uses the previous thirty days. That inconsistency can make a model perform well during testing but poorly in the hospital.

The feature store provides:

  • Consistent feature definitions

  • Reusable data across models

  • Faster model development

  • Historical feature tracking

  • Data quality controls

  • Point-in-time accuracy

  • Feature versioning

  • Real-time and batch access

  • Access controls

  • Auditability

Examples of features

Census features

These describe the expected number of patients requiring meals.

  • Current hospital census

  • Census by hospital unit

  • Census by dietary category

  • Expected admissions in the next 24 hours

  • Expected discharges before each meal

  • Average census for the same weekday

  • Seven-day census trend

  • Occupancy percentage

  • Number of patients currently NPO

  • Number of patients eligible for each meal

Example:

expected_lunch_meal_patients = 286
expected_discharges_before_lunch = 14
npo_patients_at_lunch = 22
renal_diet_patients = 31

Historical demand features

These describe what was previously prepared and consumed.

  • Meals served during the previous meal period

  • Average demand for the same weekday

  • Average demand for the same menu item

  • Seven-day moving average

  • Twenty-eight-day moving average

  • Seasonal demand

  • Demand during comparable holidays

  • Meal cancellation rate

  • Tray return rate

  • Cafeteria transaction volume

  • Demand volatility

Example:

grilled_chicken_7_day_avg = 116
grilled_chicken_same_weekday_avg = 122
grilled_chicken_last_served = 119
grilled_chicken_demand_volatility = 0.08

Menu features

These describe the meal being offered.

  • Menu item

  • Recipe category

  • Protein type

  • Vegetarian indicator

  • Sodium level

  • Calorie range

  • Portion size

  • Menu rotation position

  • Historical popularity

  • Historical waste percentage

  • Substitution history

  • Preparation complexity

  • Shelf life after preparation

The model may learn that a certain meal is consistently less popular on Fridays or that a particular side dish produces high plate waste.

Inventory features

These help Marigold consider ingredient availability and spoilage.

  • Current ingredient quantity

  • Days until expiration

  • Historical spoilage rate

  • Average daily consumption

  • Open purchase orders

  • Supplier delivery date

  • Storage capacity

  • Ingredient substitution availability

  • Cost per serving

  • Lead time

  • Inventory turnover

Example:

spinach_quantity_on_hand = 42_kg
spinach_days_to_expiration = 2
spinach_average_daily_usage = 11_kg
spinach_spoilage_risk = high

Marigold could recommend prioritizing menu items that use ingredients approaching expiration, provided dietary and operational requirements are still met.

Waste features

These measure waste across the food-service operation.

  • Food prepared but not served

  • Plate waste

  • Spoilage waste

  • Trim and preparation waste

  • Waste by menu item

  • Waste by kitchen station

  • Waste by hospital unit

  • Waste by meal period

  • Waste cost

  • Waste weight

  • Waste percentage

  • Avoidable versus unavoidable waste

Example:

vegetable_soup_overproduction_rate = 12%
vegetable_soup_plate_waste_rate = 7%
vegetable_soup_average_waste_cost = $84_per_service

Cafeteria traffic features

These predict employee and visitor demand.

  • Number of employees scheduled

  • Shift-change timing

  • Visitor count

  • Cafeteria transactions by hour

  • Day of the week

  • Hospital event schedule

  • Conference attendance

  • Promotional offers

  • Average transaction volume

  • Weather conditions

Time and calendar features

  • Hour of day

  • Meal period

  • Day of week

  • Month

  • Season

  • Holiday indicator

  • Weekend indicator

  • Days before or after a holiday

  • School vacation indicator

  • Menu cycle week

Cost and sustainability features

  • Cost per serving

  • Estimated waste cost

  • Disposal cost

  • Carbon estimate

  • Water-use estimate

  • Ingredient sourcing category

  • Local versus imported product

  • Compostable waste percentage

  • Donation eligibility

These features allow the hospital to optimize not only for quantity, but also for financial and environmental impact.

Feature-store structure

Marigold can organize its feature store around several core entities.

Entity Example features
Hospital census, occupancy, total meal demand
Unit dietary mix, tray returns, meal cancellations
Meal period breakfast, lunch, dinner demand
Menu item popularity, waste rate, cost per portion
Ingredient inventory, expiration, spoilage risk
Cafeteria transaction volume, visitor demand
Date weekday, holiday, season, weather
Supplier lead time, fill rate, delivery reliability

This structure allows the model to retrieve the right features for a particular prediction.

For example:

Hospital: Main Campus
Date: July 14
Meal period: Lunch
Menu item: Grilled Chicken

The feature store would return the relevant census, menu, inventory, historical demand, and calendar features for that exact combination.

Batch features

Batch features are updated on a schedule, such as nightly or hourly.

Examples include:

  • Thirty-day average demand

  • Historical waste rate

  • Menu popularity score

  • Supplier reliability score

  • Ingredient turnover

  • Seasonal demand

  • Average patient census

These features do not need to update every second.

Real-time features

Real-time features reflect rapidly changing hospital conditions.

Examples include:

  • Current census

  • Newly recorded admissions

  • Recent discharges

  • NPO status changes

  • Last-minute diet-order changes

  • Ingredient shortages

  • Cafeteria transaction volume

  • Cancelled procedures

  • Unexpected patient surges

A real-time feature pipeline allows Marigold to revise its recommendation shortly before food preparation begins.

For example, if twelve patients are discharged before lunch and five new patients are admitted, the forecast can be adjusted automatically.

Point-in-time correctness

The feature store should preserve what was known at the moment a historical forecast would have been made.

Suppose a patient was discharged at 2:00 p.m. The lunch forecast created at 9:00 a.m. must not use that later discharge information when training the model.

Otherwise, the model would receive information from the future and appear more accurate than it really is.

Point-in-time retrieval prevents this form of data leakage.

Feature versioning

Features should be versioned whenever their logic changes.

For example:

menu_popularity_score_v1

might use the previous 30 days of demand.

menu_popularity_score_v2

might use the previous 90 days and adjust for holidays.

Versioning allows the team to:

  • Compare model performance

  • Reproduce previous forecasts

  • Audit historical decisions

  • Roll back problematic changes

  • Safely introduce improved calculations

Feature quality monitoring

Marigold should monitor the feature store for:

  • Missing values

  • Delayed data

  • Sudden data changes

  • Invalid dietary categories

  • Duplicate records

  • Incorrect units of measurement

  • Negative inventory values

  • Census inconsistencies

  • Stale features

  • Distribution drift

For example, if the patient census feed normally updates every fifteen minutes but has not updated for two hours, Marigold should warn users and lower the forecast confidence.

Feature ownership and governance

Each feature should include metadata such as:

  • Feature name

  • Business definition

  • Data source

  • Calculation logic

  • Update frequency

  • Owner

  • Sensitivity classification

  • Expected range

  • Version

  • Creation date

  • Last validation date

Example:

Feature: expected_patient_lunch_demand
Definition: Estimated number of admitted patients eligible for lunch
Source: ADT feed and dietary ordering system
Update frequency: Every 15 minutes
Owner: Food-service analytics team
Sensitivity: Operational, de-identified
Expected range: 0–1,000

How Marigold uses the feature store

Hospital Systems
EHR · Dietary System · POS · Inventory · Scheduling
                         │
                         ▼
              Data Integration Layer
        APIs · HL7/FHIR · Files · Streaming Events
                         │
                         ▼
                  Marigold Feature Store
 ┌───────────────────────────────────────────────────────┐
 │ Census features                                      │
 │ Dietary-demand features                              │
 │ Historical consumption features                      │
 │ Menu popularity features                             │
 │ Inventory and expiration features                    │
 │ Cafeteria traffic features                           │
 │ Waste and sustainability features                    │
 │ Weather and calendar features                        │
 └───────────────────────────────────────────────────────┘
                  │                   │
        Historical features    Current features
                  │                   │
                  ▼                   ▼
          Model Training       Daily Forecasting
                  │                   │
                  └─────────┬─────────┘
                            ▼
              Recommended Production Plan
                            │
              ┌─────────────┴─────────────┐
              ▼                           ▼
       Kitchen Dashboard          Purchasing Dashboard

Daily operational workflow

At 4:00 a.m., Marigold loads the latest census, admissions, discharge expectations, diet orders, menu, inventory, and historical demand.

At 5:00 a.m., it generates an initial breakfast and lunch forecast.

Before production begins, it checks for:

  • Census changes

  • New dietary orders

  • Procedure cancellations

  • NPO changes

  • Ingredient shortages

  • Cafeteria staffing changes

It then updates the forecast and provides recommendations such as:

Prepare 118 grilled chicken portions.
Prepare 76 vegetarian pasta portions.
Reduce vegetable soup production by 14%.
Use spinach inventory first because 42 kg expires within two days.
Hold 8% reserve capacity for unplanned admissions.

Kitchen managers can accept, reject, or adjust the recommendation. Their actions should be captured and returned to the feature store as feedback.

Feedback features

Marigold becomes more accurate when it learns from operational outcomes.

Useful feedback features include:

  • Recommended quantity

  • Quantity approved by kitchen manager

  • Quantity actually prepared

  • Quantity served

  • Quantity wasted

  • Reason for manager override

  • Unexpected patient-volume change

  • Ingredient shortage

  • Menu substitution

  • Actual meal cancellation rate

The system can learn, for example, that kitchen managers regularly increase the recommendation for a popular dish or reduce quantities before holiday weekends.

Key performance indicators

Marigold should measure:

  • Food waste by weight

  • Food waste by cost

  • Overproduction percentage

  • Meal shortage rate

  • Forecast accuracy

  • Forecast bias

  • Cost per meal

  • Inventory spoilage

  • Plate waste

  • Production waste

  • Emergency food preparation

  • Kitchen labor efficiency

  • Carbon-emissions reduction

  • Dollars saved

  • Manager override rate

The most important balance is reducing waste without increasing shortages.

A successful model should not simply recommend preparing less food. It should recommend the right amount of food while protecting patient meal availability and service quality.

Suggested positioning

Marigold Waste Forecast helps hospital kitchens predict meal demand, optimize production, use inventory before it expires, and reduce unnecessary food waste—while preserving patient nutrition, dietary compliance, and meal availability.

Tagline:
Predict every plate. Prepare with purpose. Waste less.

Aster Architecture: Tiers and Layers

Aster is an AI-powered food demand and waste prediction platform designed for hospital kitchens, cafeterias, patient meal services, and healthcare restaurants. Its architecture can be organized into six major tiers: data sources, data ingestion, data management, intelligence, application services, and user experience. Each tier has distinct layers that work together to predict meal demand, recommend preparation quantities, reduce waste, and protect patient dietary requirements.

Tier 1: Data Source Tier

The data source tier contains the operational systems that provide information to Aster. These systems describe how many people need meals, what they can eat, what ingredients are available, and what has historically been prepared or discarded.

The patient census layer connects to admission, discharge, and transfer systems. It captures the number of patients currently in the hospital, expected admissions, planned discharges, transfers between units, and patient location. This information is critical because patient volume is one of the strongest drivers of meal demand.

The dietary and clinical layer receives diet orders and restrictions from the electronic health record or nutrition management system. Examples include diabetic, renal, cardiac, low-sodium, gluten-free, allergy-sensitive, texture-modified, halal, kosher, vegetarian, and liquid diets. These categories help Aster predict not only total meal volume but also demand for individual meal types.

The food-service transaction layer collects information from cafeteria point-of-sale systems, online meal-ordering platforms, kiosks, employee dining systems, and room-service applications. It records what was ordered, when it was ordered, the portion size, location, customer type, and price.

The inventory and procurement layer receives data from inventory systems, ERP platforms, suppliers, purchasing systems, and warehouse applications. It includes ingredient quantities, reorder levels, purchase orders, delivery dates, lot numbers, expiration dates, storage locations, and supplier lead times.

The production and kitchen operations layer captures recipes, menus, preparation plans, batch sizes, cooking schedules, serving times, actual portions produced, portions served, and leftovers. Kitchen equipment and smart scales may also provide direct measurements of food produced and discarded.

The external context layer brings in factors that can affect demand but are not stored in hospital systems. These may include weather, holidays, local events, flu season, school schedules, staff work patterns, menu promotions, and regional emergencies.

Tier 2: Data Ingestion and Integration Tier

This tier moves information from source systems into the Aster platform. It supports both historical batch data and real-time operational events.

The API integration layer connects Aster with systems such as EHRs, ERP platforms, point-of-sale systems, nutrition applications, and inventory tools. REST APIs, healthcare integration APIs, FHIR interfaces, and vendor-specific connectors can be used depending on the source system.

The healthcare messaging layer processes hospital events through standards such as HL7 and FHIR. For example, an admission event can immediately increase the projected number of patient meals, while a discharge event can reduce expected demand.

The event-streaming layer processes rapidly changing signals in near real time. Events may include a new patient admission, a diet-order change, an ingredient delivery, an inventory adjustment, a meal purchase, or a recorded waste measurement. Technologies such as Kafka, Amazon Kinesis, or Azure Event Hubs can support this layer.

The batch ingestion layer loads historical records at scheduled intervals. It may import several years of meal orders, census data, inventory transactions, purchasing history, menus, and waste records. This historical information is used for model training and trend analysis.

The integration orchestration layer manages the movement and sequencing of data. It ensures that dependent datasets arrive in the correct order. For example, patient census and dietary-order feeds may need to be processed before the breakfast demand forecast is generated.

The data validation gateway checks incoming records before they enter the analytical environment. It looks for missing values, duplicate records, negative inventory, incorrect measurement units, invalid dietary codes, census inconsistencies, delayed data, and malformed timestamps.

Records that fail validation can be routed to a quarantine area instead of contaminating forecasts.

Tier 3: Data Management and Feature Tier

This tier stores, organizes, cleans, standardizes, and prepares data for forecasting and operational decision-making.

The raw data layer preserves data in its original form. This creates an audit trail and allows analysts to reproduce earlier forecasts or investigate data problems.

The standardized data layer converts data from different systems into common formats. For example, pounds, ounces, kilograms, cases, trays, and individual portions may all be converted into standardized measurement units.

The master data layer maintains consistent definitions for menu items, ingredients, recipes, departments, kitchen locations, dietary categories, suppliers, and units of measurement. Without this layer, one system may identify an item as “Mashed Potatoes,” another as “Mash Potato,” and another as product code MP-102.

The curated analytical layer combines related information into analysis-ready datasets. A meal-demand table might combine patient census, dietary orders, menu schedules, day of week, historical consumption, staff volumes, and weather.

The feature store layer contains reusable variables used by machine-learning models. Important Aster features may include:

  • Current patient census by hospital unit

  • Expected admissions and discharges

  • Meal participation rate

  • Historical consumption by menu item

  • Day of week and meal period

  • Dietary-category demand

  • Average waste by recipe

  • Inventory remaining

  • Ingredient expiration proximity

  • Supplier lead time

  • Cafeteria traffic pattern

  • Weather and holiday indicators

  • Menu-item popularity

  • Recent trend in meal orders

  • Kitchen production capacity

The feature store should have both an offline component for model training and an online component for live predictions. This prevents training-serving skew, where the model sees one version of a feature during training and a different version in production.

The data-quality layer continuously evaluates completeness, accuracy, validity, freshness, uniqueness, and consistency. It identifies missing values, delayed feeds, sudden data changes, duplicate records, invalid dietary categories, incorrect units, negative inventory, stale features, and distribution drift.

The governance and metadata layer records where data originated, who owns it, how it was transformed, and which models use it. This provides lineage, traceability, and accountability.

Tier 4: AI and Decision Intelligence Tier

This is the core intelligence tier of Aster. It contains forecasting, optimization, anomaly-detection, and recommendation capabilities.

Demand Forecasting Layer

The demand forecasting layer predicts how many meals, portions, menu items, and ingredients will be required. Forecasts can be generated at multiple levels:

  • Hospital-wide demand

  • Kitchen-level demand

  • Patient-unit demand

  • Cafeteria demand

  • Meal-period demand

  • Menu-item demand

  • Dietary-category demand

  • Ingredient-level demand

Aster can combine several model types. Time-series models capture seasonality and recurring demand patterns. Gradient-boosting models capture complex relationships between census, menu, weather, dietary categories, and historical consumption. Deep-learning models may be used when there is substantial high-frequency historical data.

Aster can also use a model ensemble, where several forecasts are combined to produce a more stable final prediction.

Candidate Generation Layer

Candidate generation identifies the possible actions Aster could recommend. Instead of immediately deciding how much food to prepare, the system first creates a set of feasible production options.

For example, the candidate generator may produce the following options for a soup item:

  • Prepare 80 portions

  • Prepare 90 portions

  • Prepare 100 portions

  • Prepare 80 portions initially and hold ingredients for 20 additional portions

  • Substitute another menu item if inventory is low

  • Transfer surplus from another kitchen location

Candidates are generated using demand forecasts, recipes, available inventory, kitchen capacity, dietary rules, expiration dates, and historical waste patterns.

Ranking and Decision Layer

The ranking model scores each candidate action and determines which option creates the best operational outcome. The score may combine several weighted objectives:

[
Decision\ Score =
w_1(Demand\ Match)

  • w_2(Expected\ Waste)

  • w_3(Shortage\ Risk)

  • w_4(Operational\ Cost)

  • w_5(Inventory\ Utilization)

  • w_6(Dietary\ Compliance)
    ]

The weights reflect hospital priorities. A hospital may assign a very high weight to avoiding patient meal shortages and dietary violations, while still giving meaningful weight to waste reduction and cost.

The ranking model considers trade-offs. Preparing too little can create shortages, delayed meals, and patient dissatisfaction. Preparing too much increases waste and cost. The recommended quantity should minimize expected waste without creating unacceptable service risk.

Waste Prediction Layer

The waste model estimates how much food is likely to remain unused. It can predict:

  • Preparation waste

  • Overproduction waste

  • Serving-line waste

  • Plate waste

  • Expiration waste

  • Spoilage risk

The model may identify that a menu item has consistently high plate waste, that a certain portion size is too large, or that demand drops when a particular side dish is served.

Inventory Optimization Layer

This layer translates forecasted menu demand into ingredient requirements. Recipes are decomposed into ingredients, and the system compares required quantities with current inventory.

It recommends:

  • What to order

  • How much to order

  • When to order

  • Which inventory to use first

  • Which ingredients are at risk of expiration

  • Where substitutions are possible

  • Whether surplus can be moved to another kitchen

The optimization logic can use first-expire-first-out rules to prioritize ingredients closest to expiration.

Anomaly Detection Layer

The anomaly-detection layer identifies situations that do not match expected behavior. Examples include:

  • A sudden census increase

  • A large drop in expected cafeteria traffic

  • Abnormally high waste for a menu item

  • Inventory consumption that does not match production

  • A dietary category disappearing from the feed

  • Unusual negative inventory values

  • A meal count far outside the normal range

Anomaly detection protects the forecasting process and helps operations teams intervene before incorrect recommendations reach the kitchen.

Scenario Simulation Layer

Hospital managers can simulate operational scenarios before making a decision. For example:

  • What happens if census increases by 15%?

  • What happens if a supplier delivery is delayed?

  • What happens if a menu item is replaced?

  • What happens if a kitchen reduces batch size?

  • What happens if employee cafeteria traffic increases?

The simulator shows the expected impact on food cost, waste, shortage risk, inventory, and staffing.

Explainability Layer

Each recommendation should include the main reasons behind it. Aster might explain:

Prepare 112 renal-diet lunch meals because the renal census is 8% above the four-week average, three planned discharges were postponed, and participation for this menu item historically reaches 91%.

This helps kitchen managers trust the recommendation and override it when they have information that the model does not.

Tier 5: Application and Orchestration Tier

This tier converts model outputs into operational workflows.

The forecast service exposes demand predictions to dashboards and downstream systems.

The recommendation service provides preparation quantities, purchasing recommendations, inventory actions, and waste-reduction suggestions.

The menu-planning service evaluates menu popularity, nutrition requirements, ingredient availability, cost, and historical waste. It can recommend changes to future menus.

The production planning service creates kitchen preparation plans by meal period, production station, recipe, batch, and preparation time.

The inventory service updates expected inventory consumption and generates replenishment suggestions.

The alerting service sends notifications when demand changes significantly, stock is insufficient, ingredients are approaching expiration, dietary data is invalid, or data feeds are delayed.

The human approval layer allows supervisors to accept, edit, or reject model recommendations. Human decisions should be captured and returned to the learning system.

The workflow orchestration layer coordinates the complete process. A typical morning workflow may be:

  1. Receive updated patient census.

  2. Refresh dietary-order data.

  3. Validate inventory and menu information.

  4. Generate breakfast and lunch forecasts.

  5. Create production candidates.

  6. Rank the candidates.

  7. Send recommendations to kitchen managers.

  8. Capture manager overrides.

  9. Send approved quantities to the production system.

  10. Compare actual consumption and waste against the forecast.

Tier 6: Experience and Reporting Tier

This tier provides interfaces for kitchen staff, hospital operations, nutrition teams, procurement managers, and executives.

The kitchen operations dashboard shows today’s expected meal counts, preparation recommendations, batch schedules, shortages, and last-minute census changes.

The waste-management dashboard displays waste by menu item, kitchen, meal period, food category, and waste reason. It can highlight the largest waste-reduction opportunities.

The inventory dashboard shows current stock, projected usage, low-stock risks, expiration risk, recommended transfers, and reorder suggestions.

The dietitian interface displays demand by dietary category and allows nutrition teams to review menu substitutions and dietary-rule compliance.

The procurement dashboard shows recommended orders, supplier lead times, projected shortages, purchase-order variance, and opportunities to reduce excess buying.

The executive dashboard focuses on business outcomes such as:

  • Food waste reduction

  • Cost savings

  • Forecast accuracy

  • Meal shortage rate

  • Inventory turnover

  • Expired inventory reduction

  • Dietary compliance

  • Adoption of recommendations

  • Carbon-emission reduction

  • Waste cost per patient day

The mobile and alert interface gives kitchen managers concise notifications such as:

Lunch demand is now projected to be 7% higher because 18 planned discharges were delayed. Increase chicken entrée preparation by 24 portions.

Cross-Cutting Architecture Layers

Several capabilities operate across every tier.

Security and Privacy Layer

Aster should apply role-based access control, encryption in transit and at rest, single sign-on, audit logging, and least-privilege permissions. Patient-identifying information should be minimized because most forecasting use cases only require aggregated census and dietary information.

Governance and Compliance Layer

This layer defines how data, features, models, and recommendations are governed. It includes model approvals, data-retention rules, access policies, lineage, audit trails, and documentation of model limitations.

Model Operations Layer

The MLOps layer manages model training, validation, deployment, versioning, monitoring, and rollback. It tracks which model produced each prediction and which feature values were used.

Monitoring Layer

Aster should monitor:

  • API and pipeline availability

  • Data latency

  • Missing and invalid data

  • Feature freshness

  • Prediction accuracy

  • Model drift

  • Recommendation acceptance

  • Forecast errors

  • Waste-reduction performance

  • System response time

Feedback and Continuous-Learning Layer

Actual production, consumption, inventory usage, waste, and manager overrides are returned to the system. These outcomes become labels for future model training.

For example, if Aster recommends 100 portions, the manager changes it to 115, and 113 are consumed, that information helps Aster improve its future recommendations.

End-to-End Aster Flow

Hospital Systems and External Data
                │
                ▼
APIs, HL7/FHIR, Events and Batch Pipelines
                │
                ▼
Validation, Standardization and Data Quality
                │
                ▼
Data Lake, Warehouse and Feature Store
                │
                ▼
Demand Forecasting and Waste Prediction
                │
                ▼
Candidate Production Plans
                │
                ▼
Ranking, Optimization and Dietary Guardrails
                │
                ▼
Manager Review and Approval
                │
                ▼
Kitchen, Inventory and Procurement Execution
                │
                ▼
Actual Consumption and Waste Feedback
                │
                └──────────► Model Retraining

The most important architectural principle is that Aster is not only a forecasting model. It is a complete decision system that connects hospital operations data, predicts demand, generates feasible production actions, ranks those actions using cost and clinical constraints, sends the recommendation into kitchen workflows, and continuously learns from actual food consumption and waste.

You may also like