A maritime anomaly on its own tells you very little. A vessel goes dark, or turns somewhere it shouldn't, and an analyst spends hours pulling together weather, news and vessel history to work out whether it means anything. Detection is the easy part; the context around it is the work. Windward detects these anomalies from AIS and remote sensing data, and we built the layer that explains them, part of MAI Expert™.

Given an anomaly, with its timestamp, coordinates and vessel class, a Step Functions pipeline fans out to a separate Lambda for each source: a news feed filtered to the right time and place, a web search where a model writes the queries, and a weather API. A model on Amazon Bedrock then reads what came back and decides whether it explains the event or whether something is still missing, and if something is missing it writes another query. That self-reflection step is what makes the output worth reading, because a fixed set of queries either over-fetches or quietly returns nothing, with no way of telling which.

Filtering runs in two passes. Amazon Rerank cuts the obviously irrelevant while keeping recall high, then a model scores what survives from 0 to 100 on time, location and vessel type, and only what clears a threshold is kept. A final model writes the report, citing every source so an analyst can check it. We evaluated on past anomalies using LLM-as-a-judge across six criteria, after aligning the judges against human ratings.