Recommendation Systems at 30,000 Feet: Engineering Serendipity in Aviation
Building recommendation systems for airlines is nothing like Netflix or Spotify. 95% of users are anonymous, purchases are infrequent and high-stakes, and a 'wrong' recommendation can cost real money. Here's what I learned engineering serendipity in aviation.
This Is Not Netflix
When people think "recommendation systems," they think Netflix suggesting movies or Spotify building playlists. These scenarios share a key advantage: abundant user data.
Now imagine building a recommendation system where:
- 95% of your users are completely anonymous
- The average user interacts 2-3 times per year
- A single purchase can cost hundreds or thousands of euros
- You're serving 5+ different airline brands
Welcome to airline recommendation systems. This has been my world at zeroG (Lufthansa Group) for the past three years.
The Cold Start Problem at Scale
For most platforms, cold-start users are 10-20% of traffic. For airlines, it's 95%+.
What we CAN use: contextual signals (route, travel dates, cabin class, market, device type), aggregate patterns, and our generative AI city model that scores destinations across interest categories.
The Art of Measuring Impact
Building the recommendation system was hard. Proving it worked was harder. I built an AB testing framework from the ground up handling two different segmentation paradigms — logged-in users (by ID) and non-logged-in users (by market-behaviour clusters).
The key insight: statistical significance doesn't equal business significance. We built automated dual-reporting showing both statistical and business impact per segment. This framework eventually led to the creation of a new team dedicated to performance measurement.
The 20x Optimisation
The biggest improvement wasn't a better model — it was a better API architecture. Vectorised scoring, connection pooling, response caching, and lazy loading achieved a 20x performance improvement. Same model, same accuracy, but now fast enough for real-time use.
What Airlines Can Teach Silicon Valley
- Anonymous users aren't empty — they're just different. Every interaction carries contextual information.
- Measurement infrastructure > model sophistication. A mediocre model with great AB testing will outperform a brilliant model you can't evaluate.
- Performance IS product quality. A 5-second recommendation will never be deployed at the point of maximum impact.
- Cross-partner generalisation is hard. Culturally-aware, market-specific adaptations matter enormously.
Based on 3+ years building recommendation systems at zeroG (Lufthansa Group).
