Compass India: An AI-Powered Meal Demand Planning App

Compass India platform hero banner showcasing AI-driven business solutions

Share Your Requirements

To help our experts understand your business
objectives and create your customized plan.

Describe your project...

Share Your Requirements

To help our experts understand your business
objectives and create your customized plan.

Lead generation form initiation card illustration

About Compass India

Compass India operates food services at scale, across sites that function very differently. A corporate cafeteria follows a different pattern than an institutional kitchen, and even within one location, demand can change with little warning. Each day comes down to a simple question: how many meals will be served tomorrow? Planning once relied on past averages, local judgment, and quick adjustments. That worked, to a point.

Overestimates led to waste and higher costs. Underestimates disrupted service and procurement. Compass India needed more than a standard forecasting tool. It needed a system that could learn from historical POS data, adapt to each site’s behavior, and still allow human override when context mattered. The outcome was an AI-driven demand forecasting system built for real operations, where accuracy supports decisions, and trust sustains them.

Left image showing Compass India platform interface and team collaboration

Right image highlighting Compass India solution workflow and user experience

Driving Smarter Meal
Planning Through
Predictive Forecasting

At Compass India, meal demand rarely followed a set pattern. Each site behaved differently, and daily volumes shifted without warning. Events and attendance swings made averages unreliable. The aim was to forecast next-day meals accurately while keeping planning clear and trustworthy.

The platform needed to:

[ 01 ]

Predict demand per site and meal with enough reliability for operational planning.

[ 02 ]

Adapt to unique patterns across locations instead of forcing uniform rules.

[ 03 ]

Flag predictions for human review when confidence was low, letting planners step in.

[ 04 ]

Improve over time through retraining cycles without disrupting daily operations.

We used multiple machine learning models rather than a single predictor. Prophet captured recurring patterns and seasonality where historical rhythms were stable. AutoGluon ran many models in parallel and picked the best one for each site and meal period.

High-confidence outputs went straight into operational plans, and low-confidence predictions were flagged for review with planner corrections recorded. The result: meal planning moved from reacting to waste to anticipating needs with real-world predictive insights.

Turn Forecasts Into Confident
Planning Decisions

Move beyond averages and last-minute adjustments with models
that learn from real demand patterns.

Compass India case study CTA banner promoting project insights

Mobile CTA banner for Compass India case study

Our Process

Discovery & Data Understanding

Reviewed historical POS data across sites to understand variability.
Identified locations with stable cycles versus those driven by local events.
Used these inconsistencies to shape how forecasting models were trained.

Model Design & Forecasting Logic

Built forecasts at the site, meal, and day level to reflect operational reality.
Applied multiple models using Prophet and AutoGluon.
Designed the system so each site could adapt independently.

Confidence Banding & Human Review

Assigned a confidence level to every forecast.
High-confidence outputs flowed directly into daily planning.
Medium and low-confidence predictions were surfaced for planner review.

Self-Learning Feedback Loop

Logged all manual overrides made by planners.
Fed these corrections into scheduled retraining cycles.
Improved accuracy over time by learning from human judgment.

Ongoing Evolution

Designed the platform for extension rather than replacement.
Planned enhancements include event-based adjustments, blackout rules, and rolling multi-day forecasts.
Ensured the system could grow with procurement and planning needs.

Project Challenges

[01]

Highly Variable Demand Patterns

Meal demand differed sharply across sites and meal periods. Local events, workforce rotation, and attendance swings made historical averages unreliable, requiring models that could adapt site by site rather than follow uniform rules.

[02]

Balancing Automation With Human Oversight

Forecasts needed to drive daily planning without removing the planner's judgment. The system had to automate routine decisions while clearly flagging uncertain predictions for review, ensuring accuracy without slowing operations.

[03]

Data Inconsistency Across Locations

POS data quality varied between sites, with gaps and uneven history affecting reliability. The forecasting approach had to remain stable and accurate even when inputs were incomplete or irregular.

Solution Approach

We began with how meal planning works on the ground at Compass India. Demand changes often and not always for clear reasons. Planners trust data, but they also trust experience. The system had to respect both and adapt by site without hiding uncertainty.

Detailed solution workflow banner for Compass India platform

Key solution features

We used multiple models: Prophet for steady patterns, AutoGluon for site- and meal-specific fits.

Forecasts had confidence scores; high-confidence went straight to plans, low-confidence flagged for review.

Planner overrides fed retraining, letting the system improve accuracy in the background.

Core Technologies Powering
the Forecasting Platform

Every technology choice reflected Compass India’s on-the-ground meal planning. The stack enables next-day forecasts, confidence tagging, and manual review without slowing cycles, reducing waste while keeping planners in control and supporting scalable, reliable enterprise forecasting.

Forecasting and Data Intelligence

AutoGluon

Prophet

Statsmodels

Python

Pandas

NumPy

Scikit-learn

Backend and APIs Processing

Flask

JSON

JWT Authentication

FastAPI

REST APIs

User Interface and Visualization Logic

Streamlit

HTML 5

CSS 3

Javascript

Data Storage

PostgreSQL

MySQL

Time-Series Databases

Deployment, Monitoring, and Orchestration

Docker

CI/CD Pipelines

Apache Airflow

Prometheus

Grafana

OpenTelemetry

Cloud and Infrastructure Orchestration

AWS EC2

AWS S3

Load Balancers

Compass India Achieved Accurate Demand Forecasting

The forecasting platform helped align food preparation and procurement with real demand, shifting planning from reactive adjustments to data-backed foresight.

High forecast accuracy across operations

Pilot deployments achieved over 92% forecast accuracy, validating the use of machine learning for large-scale planning.

Human-in-the-loop confidence controls

Confidence scoring enabled operational teams to review and adjust low-confidence forecasts, strengthening adoption and long-term accuracy.

Reduce Waste Through
Smarter Forecasting

Build an enterprise forecasting and planning
solution driven by real consumption data.

Compass India platform CTA banner promoting enterprise solutions

Mobile CTA banner for Compass India platform solutions

Frequently Asked Questions

[ 1 ]

How does the forecasting system account for differences between sites?

Each site is treated on its own. Forecasts are built from that site’s historical POS data, not pooled averages. Local habits show up quickly, and over time, the forecasts reflect how each location actually behaves.

[ 2 ]

Can operational decision-makers still intervene if the forecast does not feel right?

Yes, by design. Every forecast carries a confidence level. When confidence is low, the numbers are flagged for review. Planning teams can adjust or override them, and those decisions feed back into the learning loop. This balance made machine learning useful for operations planning instead of something teams had to work around.

[ 3 ]

Is the solution limited to next-day forecasting only?

For now, yes. Next-day meal counts were the immediate need at Compass India. The broader enterprise forecasting and planning solution is set up to support rolling forecasts, calendar-based adjustments, and event-driven demand spikes as the system expands.

[ 4 ]

How does this help procurement teams specifically?

Forecasts that track real demand give procurement clearer signals. Stock levels stay tighter. Vendor planning improves. Emergency purchases drop. Over time, this supports predictive analytics for procurement planning without adding extra process.

[ 5 ]

Can this forecasting engine scale across new sites or regions?

It can. Models are trained per site, so new locations adapt using their own data. There is no manual tuning required, and each site benefits from shared learning across the network while keeping local behavior intact.