Analytics projects rarely fail on the dashboard. They fail earlier, on data that arrives late, in three shapes, from a system nobody owns.

The engineering that matters is unglamorous: ingestion that survives a source going down, schemas that can absorb a new field without a migration weekend, pipelines that are idempotent so a replay does not double-count, and enough observability that you find out about a broken feed before the business does.

We design for that failure case first, then build the reporting layer on top of it. Where the destination is an AI system rather than a chart, the same discipline decides whether retrieval returns something trustworthy or something plausible.

We write about the architecture decisions behind this in public — event design, replay and the cost trade-offs at scale.

• Event-driven architecture with Kafka
• Building a production-ready RAG pipeline