COSMED runs Raincurve in production for continuous log monitoring and incident intelligence

The challenge

COSMED operates software and infrastructure where failures have real consequences and where every investigation has to be explained afterwards. As delivery velocity increased, so did the volume of logs and pipeline events engineers had to read to understand a failure.

Investigations followed a familiar pattern: an alert in one system, a manual search across logs in several others, and a root cause assembled by hand from whatever evidence could be found. The information was there. Connecting it was the slow part.

The approach

Raincurve was deployed for continuous monitoring of PSCV logs alongside the surrounding infrastructure telemetry. Rather than replacing existing tooling, Curve-1 ran on top of it:

• Pre-filter scoring flagged high-risk sequences as they happened, instead of after an incident was already visible.
• Trajectory encoding compared current pipeline and infrastructure behavior against known failure trajectories.
• Hypothesis generation proposed ranked root causes from incomplete evidence, each with its supporting signals.
• Incident clustering grouped related symptoms into a single incident narrative rather than a queue of separate alerts.

Because every stage is inspectable, engineers could see why Curve-1 reached a conclusion and check it against their own understanding — important in an environment where investigations need to be defensible.

The results

Across more than 16,000 CI/CD trajectories evaluated in production, Curve-1 predicted failures with an AUC-ROC of 0.897. Operationally, the deployment contributed to a 50% reduction in mean time to resolution: engineers started investigations with a ranked hypothesis and its evidence instead of an empty search box.

MTTR reflects the combined effect of Raincurve and the team's workflow during the evaluation period.