Catering Assistant

Process Description:

Airlines submit catering requests for their ad-hoc flights via free-form emails. These emails typically include essential details such as flight information, specific services required (e.g., trash disposal, cleaning), and a list of requested food items. Upon receiving the email, a catering company employee manually reviews the request, cross-references the items with the available inventory, and suggests substitutions when necessary. The finalized order is then confirmed with the airline and entered into SAP for processing.

Goal:

The Catering Assistant aims to streamline and automate the extraction of key information from emails and match it to the company’s inventory list. The system pre-processes incoming emails, identifies and highlights relevant details (e.g., flight information, food items, and services), and pre-fills an Excel template with the extracted order details. A company employee reviews the suggested list, makes any necessary adjustments, and confirms the order with the airline, ensuring accuracy and efficiency.

Implementation Approach:

The solution leverages advanced Natural Language Processing (NLP) and Large Language Models (LLMs) to extract structured data from unstructured email content and attachments. The extracted information is cross-referenced with the original email to ensure accuracy, and matching inventory items are highlighted for easy review. The system also generates a comprehensive Excel file containing all extracted details, simplifying the review and confirmation process. Additionally, the system intelligently suggests alternative items when requested products are unavailable, ensuring minimal disruption to the workflow.

Challenges:

  • Email Variability: Emails arrive in diverse formats, composed using different email clients and written by individuals with varying styles. This necessitates robust email processing capabilities to handle inconsistencies and ensure accurate data extraction.

  • LLM Output Refinement: LLMs may modify extracted text (e.g., correcting spelling errors or paraphrasing), requiring sophisticated algorithms (e.g., string similarity, positional analysis) to accurately map the extracted information back to the original email content.

  • Inventory Substitutions: To handle unavailable items, the system must intelligently recommend suitable alternatives. This required enriching the inventory database with detailed item descriptions, which were previously unnecessary for human operators.

  • Multilingual and Ambiguous Requests: Emails may be written in multiple languages or contain ambiguous requests, demanding advanced contextual understanding and multilingual support.