An integrated, end-to-end AI platform that predicts menu demand, optimises ingredient procurement, forecasts food prices, and empowers suppliers purpose-built for Sri Lanka's small-to-medium restaurant ecosystem.
Domain Overview
Small-to medium-sized restaurants in Sri Lanka face chronic operational inefficiencies driven by unpredictable demand, seasonal price volatility, and limited access to data-driven tools. Manual estimation, ad-hoc supplier relationships, and no structured waste tracking result in stock-outs, over-purchasing, and significant food waste — challenges that established solutions from large supply-chain contexts cannot directly address.
ARIMA and exponential smoothing are well-established for short-term commodity prediction. However, they fail to capture nonlinear demand patterns and Sri Lanka–specific cultural surges.
Random Forest, Gradient Boosting (LightGBM), and LSTM networks outperform classical methods on volatile, multivariate datasets. Most existing implementations assume large, clean historical data.
Multi-Criteria Decision Making (MCDM) with AHP and ML-based ranking evaluate suppliers by cost, reliability, and freshness — primarily in large supply chains.
Federated Learning enables collaborative model training without sharing raw sales data. Differential Privacy and secure multi-party computation supplement this for SME settings.
EOQ and JIT models optimise static inventory but cannot adapt to perishable, demand-volatile restaurant environments. AI-powered adaptive buffers offer dynamic correction.
Weather data, cultural calendars, and temporal features significantly improve demand forecast accuracy for hospitality businesses — yet remain underutilised in small-scale settings.
Research Gap
Existing forecasting frameworks require large, structured datasets unavailable to typical Sri Lankan restaurants that maintain only manual logbooks.
Most tools predict demand in isolation without converting forecasts into recipe-mapped ingredient procurement plans, making outputs non-actionable.
Inventory buffers are set statically and never adjusted based on actual kitchen waste — missing a critical closed-loop feedback mechanism.
Restaurant owners resist sharing raw transactional data with suppliers, yet collaborative forecasting requires some form of aggregated demand intelligence.
An integrated end-to-end system combining lightweight hybrid forecasting, adaptive waste-driven inventory control, price prediction, and privacy-preserving supplier dashboards.
Incorporates the Sri Lankan cultural calendar, local weather, and SME operational realities — features no existing solution currently addresses holistically.
How can an integrated system be created that forecasts restaurant menu demand, anticipates ingredient price fluctuations, recommends reliable suppliers, and reduces food waste — specifically for Sri Lankan restaurants operating with limited structured data, low digital infrastructure, and privacy-sensitive supply relationships?
Research Objectives
Technologies Used
System Components
Each component is independently developed and seamlessly integrated into a unified decision-support pipeline for Sri Lankan restaurants and their suppliers.
The foundational forecasting engine of the system. Designed specifically for small Sri Lankan restaurants with limited historical data, this component combines a lightweight ML core with a weighted fuzzy-logic contextual engine to produce context-aware daily and weekly demand predictions.
The ML core uses LightGBM and Facebook Prophet models trained on historical sales data augmented with engineered temporal features — lag-1, lag-7, and rolling averages. The contextual engine overlays real-time OpenWeatherMap meteorological data and a standardised Sri Lankan cultural calendar (Vesak, Sinhala New Year, Poya Days) to refine predictions.
Unlike binary rule triggers, the system applies Fuzzy Logic Scaling — a continuous adjustment factor (δ) that gradually reduces forecasts as rainfall intensifies rather than switching on/off. Overlapping modifiers (e.g., festival during a storm) are resolved via a Weighted Priority Hierarchy.
FINAL FORECAST FORMULA
Ŷfinal = Ŷbase × (Σ wiδi) / Σ wi
Where Ŷbase = calibrated ML baseline, δi = fuzzy adjustment factor, wi = historical reliability weight
Research Gap Addressed
Existing models assume large, clean data. This component uses proxy data (supplier invoices, logbooks), operates on <1 year of records, and encodes domain expert knowledge through transparent rule-based overrides — making it the first forecasting system tailored to Sri Lankan SME constraints.
SYSTEM ARCHITECTURE — COMPONENT 01
| Model | MAE | RMSE | R² | MAPE |
|---|---|---|---|---|
| Proposed Hybrid Model | 3.42 | 4.89 | 0.88 | 12.5% |
| LightGBM (standalone) | 3.42 | — | — | — |
| Facebook Prophet | 5.12 | — | — | — |
This component transforms demand forecasts into physical inventory actions. It computes total ingredient requirements through recipe mapping, dynamically calculates reorder thresholds, and features a pioneering Adaptive Buffer Algorithm that continuously updates safety stock levels using actual kitchen waste data.
The dynamic inventory planner calculates the target stock level as:
TARGET STOCK CALCULATION
Starget = Lreorder + Bsize
Btarget = Wavg × Dbuffer
Bnew = Bold(1 − α) + Btarget · α
α = smoothing factor · Wavg = mean daily waste over lookback window
The adaptive buffer treats kitchen waste as an active data signal rather than a passive metric. By tracking daily unsold food and applying exponential smoothing, the system autonomously corrects safety buffers without human intervention — linking predictive analytics with real kitchen operations.
The Next.js / React frontend presents complex risk states (OK / Low / Stockout) in a simple, visually intuitive dashboard, purpose-designed for restaurant staff with limited digital literacy.
ADAPTIVE BUFFER FEEDBACK LOOP
Most Sri Lankan restaurants rely on manual cost estimation with no predictive capability for ingredient price fluctuations. This component delivers a machine learning–based food price prediction engine and an intelligent supplier ranking and recommendation system — integrated into a unified decision-support interface.
Using historical food selling prices, ingredient cost data, and current market prices, the component forecasts weekly and monthly dish selling prices. Models benchmarked include ARIMA, Prophet, Regression, and LSTM — evaluated on RMSE, MAPE, and R².
Feature engineering enriches the dataset with rolling price averages, seasonal indicators, and a computed cost index per dish. Dish-ingredient ratio mapping and supplier-ingredient mapping enable precise cost-to-price linkage.
A multi-criteria ranking algorithm scores each supplier on cost efficiency, delivery reliability, and freshness ratings. AHP-based MCDM ensures transparent, explainable recommendations. The system alerts restaurant owners when significant price fluctuations are detected and supports PDF/Excel export of supplier rankings.
SYSTEM ARCHITECTURE — COMPONENT 03
KEY FUNCTIONAL REQUIREMENTS
Small suppliers serving Sri Lankan restaurants lack demand visibility, leading to inefficient delivery scheduling and stock imbalances. This component builds a privacy-preserving supplier demand forecasting dashboard that aggregates ingredient-level demand signals from multiple restaurants without exposing raw sales data.
The system collects dish-level sales data from restaurants, converts it to ingredient-level demand via predefined recipes, and applies an ensemble of ARIMA (stationary time-series), Holt-Winters (trend + seasonality), and Random Forest Regression (nonlinear demand) models — weighted by rolling MAPE/RMSE performance.
Using Federated Learning (FL), each restaurant trains its local model and shares only model parameters (gradients) — never raw transactional data. An optional Differential Privacy layer adds controlled noise to further protect individual restaurant identity. This approach enables collaborative forecasting while maintaining competitive confidentiality.
Supplier-facing reorder points are computed using Economic Order Quantity (EOQ) and safety stock from forecast error. The interactive dashboard (React/Dash) provides time-series demand graphs, inventory heatmaps, stockout/overstock alerts, and drill-down analysis with CSV/PDF export.
SYSTEM ARCHITECTURE — COMPONENT 04
Project Timeline
All four components follow the academic year timeline (September 2025 – August 2026), spanning data collection through final presentation.
Individual proposal reports submitted by all four team members. Research gaps identified, methodologies defined, and systems architectures drafted for all four components.
Data collection completed, preprocessing pipelines built, and initial ML model training underway for all components. Baseline forecasting models benchmarked.
Full system integration milestone. Hybrid forecaster integrated with inventory management. Supplier dashboard connected with federated learning layer. End-to-end pipeline functional and under evaluation.
Final research paper submitted. All four system components validated against holdout datasets. Viva voce conducted.
Schedule Overview
Research Documents
All project documentation including proposal reports, checklists, and the final research paper. Download links will be updated as documents are finalised.
Initial project charter defining scope, team roles, and research objectives for Project 25_26J_393.
IT22073846 · Lightweight Hybrid Food Demand Forecaster component proposal.
IT22642950 · AI Enabled Inventory Management System
IT22261946 · Food Price Prediction & Supplier Recommendation component proposal.
IT22249470 · Supplier Demand Forecasting Dashboard component proposal.
Assessment checklists and milestone compliance documents for all project phases.
Combined research paper covering Components 01 & 02 (IEEE format). Currently under revision.
Complete final research document integrating all four components — due August 2026.
Presentations
Presentation decks from all assessment milestones. Future presentations will be added here upon completion.
The Team
Final year undergraduate students at the Sri Lanka Institute of Information Technology (SLIIT), pursuing B.Sc. (Hons) in Information Technology. Project ID: 25_26J_393.
Research Paper Primary Author · Adaptive Buffer Algorithm
Hybrid ML + Rule-Based Forecasting · LightGBM, Prophet
ARIMA · LSTM · AHP Supplier Ranking · Decision Support
Supplier Demand Forecasting · Supplier Dashboard · EOQ
Supervisors
Project Supervisor · Department of Information Technology, SLIIT
supunya.s@sliit.lk
Co-Supervisor · Department of Information Technology, SLIIT
chathurya.k@sliit.lk
Get In Touch
Have questions about our research? We'd love to hear from you.
it22073846@my.sliit.lk · 0719355359
it22642950@my.sliit.lk · 0759089188
it22249470@my.sliit.lk · 0764233231
it22261946@my.sliit.lk · 0776474109