CheckSync @ HaQathon DOH

Taking special cargo acceptance to the next level incorporating AI-supported performance of checks and pre-checks of AWB data, while sharing check templates and results between stakeholders via 1R.

Comment

Inspiration

At the current point in time, there are many digitalization attempts for acceptance check sheets by airlines and GHAs respectively. But none of these solutions are connected to other partners, which causes additional effort or leads to process issues.

In case of GHA check sheet applications, GHAs need to keep check sheets templates of their airline customers up to date, which is time consuming and results in additional IT costs. Moreover, some check sheets (e.g DG or Live Animals) are still required to be printed by the airlines, since this is the only reference of a successful check for them, which contradicts the goal of a digital solution.

Airline check sheet applications on the other hand don't fit into GHA's IT landscapes and cause additional training effort. Moreover, these apps lack the possibility to block shipments in GHA's handling systems. This can, in the worst case, lead to an unintended and dangerous transportation of an unchecked shipment. This is why we from Swissport, Lufthansa Cargo & Lufthansa Industry Solutions partnered to build a PoC of an ultimate, interconnected check sheet solution that solves above problems.

With AI tools currently on the rise, we felt like it was also about time to investigate the usage of generative AI to assist the check process. Ultimately this will lead to a fully supported DG check, with less errors and less training efforts needed.

What it does

We built an application (OnSync) to perform special cargo checks with multiple innovations:

  • Is capable of using AI (ChatGPT) to check AWB data before physical arrival of the shipment and provides check result back to the shipper to take pre-emptive corrective action if needed
  • Fetches airline-provided check templates via 1R, making sure that the current applicable check sheets are always used. This minimizes maintenance efforts & costs on GHA side.
  • Supports users during the check with AI-assisted proposals for check answers to minimize errors and speed up the process
  • Is capable of syncing check status & answers with all relevant stakeholders, including airlines, forwarders and shippers. Everyone has transparency about the check process and outcome, receives the necessary data in their IT systems in real-time (e.g. to archive check sheets in the format required by the airlines).
  • Provides a chat assistant to the end user that can answer questions about regulations, airline-specific processes and shipment details.

With these features we address the following challenges:

  • Challenge 1: Special cargo validation – Automatic pre-checking of shipment data and provision of check outcome to forwarders/shippers
  • Challenge 3: ONE Record AI Challenge – AI-driven support of check sheet performance
  • Challenge 6: Open – Synchronization of check sheet templates and check outcome with 1R

How we built it

First step was the definition of check sheet templates for common types of DG check sheets (Non-Radioactive DG, ICE, ELI/ELM) using the current OneRecord check sheet data model. To lay the groundwork for a ChatGPT-assisted completion of the check sheets, we defined prompts that return the correct answers for the documentary questions, taking into account the shipment data (e.g. goods description, origin, destination), and fine tuned them with real world data.

Simultanously, the coders of our team worked on building an Angular frontend with Angular Material UI components that allows to select or scan shipments, perform the relevant check sheets and use AI-assistance as part of the check sheet performance workflow. Moreover, the AI Assistant with its accrued knowledge about DG regulations and available ONE Record data objects was made available to the user via a chat interface. The Generative AI that powers both features is Azure OpenAI, which allows us to access Large Language Models in a secure manner without losing the control over confidential data. In the backend we used the 1R data model to enable the exchange of check sheet data via NE:ONE 1R server with servers of other stakeholders.

Challenges we ran into

  • Late changes in storyboard for the video
  • Time went by surprisingly fast :)
  • Working with 1R API and the JSON LD data model without an SDK proved to be time-consuming
  • Fine-tuning the ChatGPT prompts to get consistent results was challenging

Accomplishments that we're proud of

  • Using ChatGPT to answer selected check sheet questions worked much better than anticipated
  • How well the multicultural team from several companies (Swissport, Lufthansa Cargo & Lufthansa Industry Solutions) with different backgrounds came together, and worked as one
  • Distribution of work throughout the team was handled very well. Team was also able to adjust the plan in real-time

What we learned

  • Some fields for check sheets are missing in the current version of the data model:
    • mandatory: boolean (is a question mandatory or optional to answer)
    • positiveAnswer: string (defined what answers lead to a positive outcome of the check; e.g. "Yes" or "Yes;N/A" for all DG check sheet questions)
  • The current way of linking the check sheet "Answer" objects via the "Question" and "CheckTemplate" object to the "Check" object in the data model is very inefficient and results in having redundant instances of the "CheckTemplate" and "Question" objects.
  • Start with the script for the video early, because late changes are time-consuming to implement

What's next for CheckSync

The PoC app built during the Hackathon is a representation of future fully integrated check sheet process, used in a GHA IT landscape. Next steps will be to work on refining the 1R data model for check sheets and a pilot with an airline and a GHA to test the exchange of check sheet data via 1R between currently used IT systems in a real world scenario. On the AI part, more advances will be made in the field, which will increase the likelihood of correct answers to be even better than now. There will be a point, when such solutions generate less errors than humans. At the same time AI generated solutions will be more adaptable to changes than pre-defined rules engines.

Built With

Share this project:

Updates