Inside Meltwater Engineering.

We moved 1.23 TB of sentence embeddings out of our database cluster and into Amazon S3 Vectors, cutting the workload's run cost by roughly 90% and extending semantic search from 30 days to 24 months.

We rebuilt our journalist research pipeline from the ground up, splitting one generalist agent into focused specialist components. The result was 40% lower costs, around 10% higher prediction accuracy, and a system we can finally improve one piece at a time.

Keeping nearly 1,000,000 journalist profiles accurate used to be a fully manual job. We built an AI agent, Admiral, that cut research cost per profile by 85% and time per profile by 90%. Two lessons did most of the work: the value of a self-review step, and keeping the agent's toolset as...

The Meltwater Paris AI team attended the inaugural PyTorch Conference Europe at Station F. Here are our key takeaways on inference maturity, GPU kernel tooling, hardware co-evolution, and what it means for our AI platform.

How we reduced inference costs by 45.5% while improving accuracy by 3.02% by rethinking how our entity-level sentiment system processes documents.

How we tracked down and fixed a memory leak in a Kubernetes-hosted Node.js service after upgrading from Node 18 to 22, uncovering issues with closure retention, caching logic, and V8 engine changes along the way.

How we built a simple yet powerful event-driven system using MongoDB change streams and AWS SNS/SQS to replace polling-based architectures and deliver real-time user data changes to downstream consumers.

How we used pgcopydb to migrate a 1TB PostgreSQL database from AWS to Azure, covering migration strategies, Kubernetes execution, and the pitfalls we encountered along the way.

From merging pull requests to growing teams — how a 50/50 Team Lead experiment at Meltwater became a full identity shift, one uncomfortable conversation at a time.