Background

Sharing a project I did during my position at Etraveli Group.

To give some high level background for you who are not in the deep flight industry. Most online travel agencies selling flight tickets source them using third party systems. A large portion of tickets come from a type of system called GDS-systems. Think of them like a mega large system filled with millions of flight tickets. To use them, they require their own very complex command language, which requires a lot of training.

The Problem

This problem came to my mind during one meeting. A large organization, thousands of people, it's obvious that only a hand few of people will be really good at using the GDS-systems…

But what if, everyone could use these systems by just using natural language? More or less, very similar to vibe coding. Natural language —> a specific command language.

The Solution

This comes to the solution, let's use a couple of AI agents, APIs and custom coded tools. Turned out to be a bit more complex compared to my previous builds, but I am always up for a challenge.

To explain the workflow:

1. User inputs in Slack what they want to search for in the GDS system (e.g. "Find the lowest priced flight ticket from LON-NYC on December 11th")

2. It give the message to the main "Orchestrator" agent, responsible for calling the right agents and make sure the workflow is working

3. The first "Query" agent gets the message, and uses a LMM model to reform the message to a correct API request format according to the GDS-systems documentation

4. The query gets sent to the "GDS Connector" agent, an agent with a tool which can send API requests to a GDS-system. It has knowledge of all the different APIs and chooses the most relevant one to send the request to.

5. The agent retrieves an API response, which is usually in JSON-format an really hard to read. So for everyone in the company to understand it, its forwarded to the "Explainer" agent.

6. The "Explainer" agent has deep knowledge about GDS-systems, and uses an LLM model to create an explanation of the API response, and puts it in a clear and structured format.

7. The final answer is sent back to the "Orchestrator" which returns the output back to the user in Slack.

Some Conclusions

Whilst the code for this AI system will remain private and in the ownership of the company, I can share with you some learnings from this project & want i enables. Well yes, as I intended, now everyone within the company, meaning from 30 to now —> thousands of people being able to use the complex GDS-systems. Your welcome :)

First one being, nearly everything is possible today to do with AI agents. As long as you can build a custom function for what you want the agent to do — it can do it.

Secondly, AI agents are a lot more reliable when just given a single task or specialisation — think of it being an employee, specialised in one area. Thus this workflow have a total of 4 different agents, all working together as a team.

A single thought that came to my mind after building this. Image in the near future, it would not be totally impossible to build a whole company consisting of just a large network of AI agents. Every agent assigned a different role, and "manager" agents responsible for every agent working together as a team. Scary.... but also so interesting for what we can achieve.