I built this system alone at a travel startup, and I still run it. I can’t show the company’s code, so I only describe the structure and the decisions.
How it used to work
To start selling a new hotel, three sources had to be matched. The rate descriptions from the booking system (GDS) mixed room details, perks, and conditions in one paragraph. Hotel websites looked different for every hotel. Luxury travel agency perks were on yet another site. When we emailed hotels to confirm details, only half replied. No one knew how far each hotel had gotten.
What I changed
- AI-written sentences are not saved as is. The AI only extracts facts like “breakfast credit, 30 dollars, per person, daily.” Code writes the sentences guests see. So the same rate always shows the same sentence.
- AI matches rooms by looking at evidence instead of fixed rules. It looks at three things together: size, beds, and view from the website, the booking system’s original text, and each hotel’s product codes.
- The AI that did the work does not check it. A different AI, which doesn’t know the conclusion, checks again from the source alone.
- One full pass should leave no open work. When unclear cases piled up, it was hard to see what needed a person. So every item is closed as done, on hold, or excluded.
- Once there were more than ten kinds of AI, I built a single control board. Each AI only did its own job, so no one noticed when work got stuck between them. The board looks for problems often, and a supervisor AI tries to fix them first.
- Claude and Codex read the same task manual. If one is blocked, the work moves to the other.
Things went wrong too
- While re-extracting perks, it deleted perks that were fine. Since then it never deletes existing perks. It only adds.
- A room missing from a website scraped in July was there in September. Now “not on the website” is decided by checking the current page directly.
- Breakfast-only credits were merged into general credits. Now the types are kept separate, and per-person and maximum-guest conditions are saved on their own.
Now
AI looks after thousands of hotels every day. I only handle what needs to be asked of the hotel and risky decisions.