This repository encapsulates an end-to-end ML system architected for high-precision food delivery time (ETA) forecasting. The system integrates principled data preprocessing, experimental methodology, ensemble modeling, automated pipeline orchestration, and containerized deployment infrastructure.
In the on-demand logistics domain, time precision is a first-class citizen. Accurate ETA (Estimated Time of Arrival) estimation affects:
- Customer experience (trust and retention),
- Fleet management (operational efficiency),
- Resource optimization (cost reduction in high-traffic nodes).
This project simulates a Swiggy-like environment to build an ML-driven delivery time inference engine, optimizing prediction accuracy over heterogeneous delivery contexts.
Design and operationalize a machine learning pipeline that generalizes well over:
- Urban mobility variability (city_type, traffic, vehicle_condition),
- Contextual temporal patterns (order_time_of_day, is_weekend, festival),
- Service constraints (multiple_deliveries, type_of_order, ratings),
- Geospatial logistics (pickup_time_minutes, distance).
- Missing Data Strategy
- Model Benchmarking & Selection
- Hyperparameter Search & Optimization
- Stacked Ensemble Integration
| Feature Name | Description |
|---|---|
age |
Driver or account tenure; proxy for experience |
ratings |
Historical service quality, indicating behavioral patterns |
pickup_time_minutes |
Latency between order placement and pickup |
distance |
Geodesic or route-based delivery path length |
weather |
Encoded climatic category; affects delivery friction |
type_of_order |
Categorical: food class or preparation intensity |
type_of_vehicle |
Mode of delivery (bike, scooter, etc.) |
festival |
Binary indicator of peak-demand seasonality |
city_type |
Urban classification (metro, tier-2, etc.) |
is_weekend |
Binary flag for weekend status |
order_time_of_day |
Time block encoding (morning, lunch, dinner, late night) |
traffic |
Congestion proxy; ordinal or sensor-based input |
vehicle_condition |
Maintenance score or categorical proxy (new, moderate, old) |
multiple_deliveries |
Number of concurrent deliveries being handled by the same delivery agent |
- Encodes deterministic ML workflows via
dvc.yaml. - Stages include: preprocessing → training → evaluation → model registry.
- Outputs logged and version-controlled via DagsHub.
- Artifacts (datasets, models) pushed to S3 using secure, programmatic access.
- Enables scalable remote storage and pipeline portability.
- Central hub for:
- Experiment metric comparison,
- Model performance visualization,
- Commit-wise provenance tracking.
- Triggered on pull requests or commits to
main. - Validates:
- Code correctness,
- DVC pipeline integrity,
- Docker image build success,
- FastAPI application health checks.
- Ensures environment reproducibility.
- Production-grade FastAPI app encapsulated for lightweight serving.
- Final Model: StackingRegressor (LGBM + RandomForest)
- Accuracy: ~85%
- Error Distribution: Majority predictions within ±4 min deviation
- Inference Speed: ~20ms per sample on CPU
- Modular project layout using a cookiecutter template.
- Reproducible pipeline runs via DVC with data provenance.
- Metrics and model lineage tracked visually in DagsHub.
- End-to-end automation from data versioning to deployment pipeline.
- Dockerized REST API for low-latency predictions.
- Stacked Generalization: Ensemble strategy that learns a meta-model from base model predictions to minimize generalization error.
- Data Provenance: Recording the lineage and transformations of data across the ML lifecycle.
- Pipeline DAG: Directed Acyclic Graph of pipeline stages, ensuring immutable, traceable model training processes.
- CI/CD in MLOps: Continuous Integration for data science code, and Continuous Delivery for pipeline re-runs, model validation, and container redeployment.
- Entropy-based Modeling: Algorithms like LGBM utilize gain (information entropy) to optimize tree splitting for tabular data.
This project exemplifies the convergence of rigorous data science, software engineering discipline, and cloud-native infrastructure to deliver a resilient and scalable delivery ETA prediction system. From hypothesis-driven feature engineering to LGBM-stacked inference pipelines, it brings forth a deployable ML product ecosystem, suitable for enterprise-grade logistics systems.