Classic optimization algorithms with cloud-native infra and flight-delay ML for sub-second dispatch
Tech Stack
Core Technologies
Python
Flask
Microservices
Optimization algorithms (greedy, genetic)
AWS Services
ECS/Fargate
Lambda
API Gateway
S3
SageMaker
CloudWatch
ML & Algorithms
XGBoost
ML pipelines
Event-driven architecture
Challenge
•Dispatch decisions didn't consider flight delays or real-time constraints
•Drivers spent too long idling at airports or missing optimal jobs
•Initial monolithic logic was hard to scale and tune under peak demand
•No predictive layer to anticipate schedule disruptions
Solution
•Built an allocation engine using greedy and genetic algorithms to assign drivers to services under constraints
•Implemented aggressive caching and memoization to keep response times sub-second
•Deployed microservices on AWS ECS/Fargate behind clear APIs
•Created an optimal route generation system that computes best-effort routes considering driver distribution, demand, and business rules
•Uses Amazon S3 for configuration and state storage (durable, cheap, versioned)
•Developed a flight delay predictor on AWS SageMaker with historical flight + weather data
•Engineered features (weather, airline, time, weekday, etc.) and trained a two-stage pipeline: classifier (delay yes/no) + XGBoost regressor (delay size)
•Deployed managed endpoints and wired them into dispatch via API Gateway
•Offloaded heavy traffic management functions to AWS Lambda for event-driven scaling
•Monitored via CloudWatch (latency, errors, model health)
Impact
•Achieved sub-second allocation decisions under production load
•Dispatch decisions became delay-aware, reducing idle time at airports
•More predictable service quality thanks to flight-aware scheduling
•Architecture now scales horizontally with demand and is easier to evolve
