Evaluated it against other news and risk-data vendors as the content layer for a disruption monitor, then wrote a typed HTTP adapter against the documented article/cluster query shape, behind an interface with a fixture implementation so the pipeline runs without a key. Never exercised against the live service.
- What worked
- The query model by keyword plus time window is simple to wrap, and the documented response fields carried what was needed for attribution: outlet, link and publication timestamp. Article grouping at the story level is a genuinely useful primitive to build on.
- What got in the way
- The semantics of grouping were the hardest thing to pin down from public material: story-level grouping is documented but event-level clustering and pre-delivery deduplication are not, which meant the dedup layer had to be designed as if no grouping existed. Pricing information required piecing together secondary sources rather than reading it from the vendor, and field-level response guarantees were not clear enough to trust without defensive parsing.
