AI travel agents have a data problem before they have an AI problem
Language models understand the traveler's request. The harder problem is providing trustworthy product identity, pricing, availability, provenance and freshness.
NOTES Things we learned
Things we learned
the expensive way.
Field notes on keeping data pipelines alive, getting AI agents past the demo, and building software that survives an audit. No thought leadership — just what we ran into and what it cost.
Modern aviation data pipelines: APIs, NDC and the gaps between them
The useful question in aviation data is not API versus scraping. It is which source should be authoritative for each requirement — and how to normalize the result.
Why travel activities data is an entity-resolution problem, not a scraping problem
The same experience can appear as several listings, options, languages and price structures across different systems. Collecting the pages is only the beginning.
Travel data collection at scale: the scraper is the easy part
A collector proves that data can be extracted. A production travel-data pipeline has to keep the dataset complete and trustworthy while the systems underneath it keep changing.
HTTP 200 is not uptime: the silent failures that break travel data pipelines
The most dangerous external-data failure is the one that looks successful: the job ran, requests returned 200, rows landed — and the dataset is wrong.
Why scrapers die in month three
Most scraping projects work perfectly for about eight weeks. What happens after that is predictable, and it has almost nothing to do with how well the scraper was written.
Your AI pilot didn't fail on the model
AI pilots stall for reasons that have very little to do with which model was chosen. The failure is almost always in the layer underneath — and it was there before the pilot started.
Nobody asks about the audit log during a demo
Access control and audit trails are the least demonstrable parts of regulated software — and close to impossible to retrofit convincingly. A note on why they go in first.
Something in here sound familiar?
If a pipeline keeps breaking or an AI pilot won't ship, we've probably seen that specific version of it. Thirty minutes, no deck.
Talk to an engineer → RSS