Code-Driven Fine-Tuning of AI Agents for Airline Disruption Management

Frank Morales Aguilera, BEng, MEng, SMIEEE

Boeing Associate Technical Fellow /Engineer /Scientist /Inventor /Cloud Solution Architect /Software Developer /@ Boeing Global Services

The airline industry is a complex and dynamic ecosystem where maintaining operational efficiency and customer satisfaction is paramount. However, disruptions caused by various factors, such as weather, technical malfunctions, and air traffic congestion, can lead to significant challenges. Effectively managing these disruptions requires intelligent systems capable of making rapid and informed decisions. The notebook titled “MLxDL/FineTuningDM_MISTRAL_DEMO.ipynb” explores the promising potential of code-driven fine-tuning to develop AI agents for airline disruption management. This essay will delve into code-driven fine-tuning, discuss its application in the notebook, and highlight its potential for creating intelligent systems, aligning with my goal of developing an AI agent.

The Essence of Code-Driven Fine-Tuning

Code-Driven Fine-Tuning in the Notebook. The notebook “MLxDL/FineTuningDM_MISTRAL_DEMO.ipynb” is a significant resource that effectively demonstrates code-driven fine-tuning in the context of airline disruption management. It utilizes the Mistral-7B-Instruct-v0.1 model and fine-tunes it on a dataset of Python code snippets representing various agent functionalities, including:

• CrewAgent: Responsible for managing crew-related operations like assigning crew members to flights and updating schedules.
• PassengerAgent: Handles passenger-related tasks such as retrieving passenger manifests and notifying passengers.
• FleetAgent: This key component orchestrates overall recovery efforts, including assessing the impact of disruptions and coordinating crew and passenger actions. It also provides reassurance about the system’s capabilities.
• Database Operations: This component creates, connects, and closes database connections.
• Analytics Functions: Capabilities for analyzing disruption impact and calculating associated costs.
• Disruption Simulation: Functionality to simulate disruption events at specific airports.

By fine-tuning the model on this code-based dataset, the notebook empowers it to generate code for handling various disruption scenarios, thus automating critical tasks and improving response times.

Conclusion

By treating code as a primary data source, we can create intelligent systems capable of automating complex tasks and solving real-world problems. This approach has significant implications for various domains and aligns with the broader goal of developing sophisticated AI agents.