01Context
HotelKey's CRS APIs sit in the critical path of every booking. Partners integrate against them, the contact-center booking platform depends on them, and enterprise brands hold them to contractual SLAs. Historically, an integration was "done" when the partner said it worked. Defects surfaced as client-reported discrepancies in production, and a meaningful share of support volume was API-related: mismatched payloads, edge cases nobody had tested, and behavior that differed between staging and production.
02My role
I owned the analytical side of integration quality: defining what "certified" meant for a partner integration, building and running the test collections that proved it, auditing the data that came out the other side, and tracking integration health against SLAs. I was the person a brand or partner talked to when something didn't match, and the person who found out why.
03Approach
- The spec is the contract. Every test traced to a Swagger/OpenAPI definition, so a partner disagreement became "the spec says X" rather than a negotiation.
- Collections, not clicks. Postman collections per endpoint, with environment files for sandbox and staging, meant a certification cycle was re-runnable in minutes instead of re-created from memory.
- Prove it landed. A passing request isn't a passing integration. SQL audits against reservation and booking tables verified that what the API accepted matched what the database stored, which is where most "it says success but the data's wrong" tickets originate.
- Pass/fail in writing. Certification results were documented per partner, per cycle, so release-readiness was evidence, not a feeling.
- Watch it after go-live. Uptime, response time, and error rates tracked against each integration's SLA, with breaches escalated to engineering with the root-cause already narrowed down.
04Outcome
The ticket reduction came from two directions at once: fewer defects reaching production because certification caught them, and faster resolution for the ones that did because the SQL audit and SLA data pointed straight at the cause.
05What I'd do differently
Automate the run itself earlier. The collections were re-runnable, but a human still ran them; wiring them into Newman on a schedule, and generating the certification document from the run output, is the step that turns a good process into an unattended one. That's the CRS Certification Harness project on the ledger, and the SLA Sentinel is the monitoring half of the same idea.