Our client, an American company operating globally, specializes in organizing and coordinating private flights worldwide. The company sought a way to automate its internal processes related to knowledge management. Addepto was selected to design and implement an intelligent system for creating aviation documentation integrated with client’s infrastructure.
Case Study Shortcut
Challenge
Managing the Complexity and Interdependencies in ASR Documentation
The Aviation Service Request (ASR) documents required the inclusion of numerous interdependent variables such as aircraft type, passenger requirements, number of stops, and legal or technical constraints. Although the format was standardized, the underlying logic varied significantly for each request, making manual processing prone to errors and inefficiencies.
Adapting General-Purpose LLMs to Domain-Specific Needs
While Large Language Models offered powerful capabilities, they lacked readiness for aviation-specific use cases. The challenge was to convert a general-purpose model into a domain-specific tool that could reliably understand and act on complex flight-planning logic without introducing hallucinations or compromising on accuracy.
Integrating Seamlessly with Existing Infrastructure and Ensuring Cost-Effective Deployment
The client required the solution to work within their existing infrastructure (TripGrade on AWS) without major overhauls. Balancing performance and cost-efficiency was critical, initial attempts to use Amazon Bedrock models failed performance benchmarks, requiring a pivot to OpenAI’s models and further refinement through prompt engineering and microservices integration.
Goal
- Automate and Streamline ASR Document Creation
- Ensure Accurate Handling of Complex and Interdependent Flight Data
- Maintain Standardization While Adapting to Case-Specific Logic
Outcome
The tool developed by Addepto did not replace the operators who manually filled out ASR documentation but significantly improved their work by speeding up information retrieval and reducing the number of potential errors.
The chatbot acted as an assistant, accelerating the planning process by automatically creating and submitting ASR documents.
Users benefited from a more efficient planning process with the automated creation and submission of ASR documents. This ensured that all essential details for trip preparation were accurately collected, leading to well-organized and successful trip outcomes.
Before
- Manual ASR documentation with potential errors and slow information retrieval.
After
- Automated ASR document generation with reduced errors and accelerated planning processes using an AI-powered chatbot.
Integrate those solutions in your company
Case Study Details
Approach
Understanding Domain-Specific Requirements
- The project began with in-depth consultations to understand the nuances of private flight planning.
- This helped define the scope and ensure the chatbot could meet specific business needs around ASR (Aviation Sourcing Report) document generation.
Model Selection and Testing
- Initial testing was conducted using Amazon Bedrock (LLaMA 3 and Mistral) for cost-effective deployment on AWS.
- Despite strong accuracy, these models underperformed in real-time use, leading to the adoption of OpenAI models (GPT-3.5 Turbo / GPT-4) for better performance.
Domain Adaptation via Business-Specific Knowledge
- Rather than relying solely on general-purpose models, Addepto connected the LLM to internal SQL databases.
- Built custom logic to guide the LLM on how to use structured business data effectively for generating precise ASR documents.
Microservices Architecture and API Integration
- Designed a scalable microservices architecture using FastAPI for integration with existing client systems like TripGrade.
- Enabled modular, maintainable, and cloud-native deployment on AWS infrastructure.
Agent Logic and Prompt Engineering
- Developed precise agent logic, linking specific prompts to specific actions.
- Implemented dual response validation and prompt testing to ensure reliability and accuracy of chatbot responses.
Accelerated Development with ContentCheck
- Used Addepto’s proprietary ContentCheck tool to automate testing and evaluation of prompts.
- This significantly reduced development time and improved the consistency of the chatbot’s outputs.
Our Team Expert Opinion
We are aware that general-purpose LLMs are seldom ready for immediate business application, so we have developed various methods to overcome their limitations and selected the ones that achieve the Client's business objectives in the most optimized, time- and cost-effective manner. For instance, instead of resorting to costly LLM fine-tuning, we implemented automated prompt engineering for database integration. Our proprietary tools streamlined the validation process, making it both efficient and swift. Ultimately, we delivered a fully customized solution based on the market-proven, state-of-the-art technologies.
Michał Tarkowski Engineering Manager at Addepto
