Superhuman runs one of the largest AI productivity platforms in the world, processing text across dozens of languages for millions of users every day. A unified data platform on Databricks powers analytics, machine learning and the pipelines behind customer-facing features. However, the infrastructure connecting that data to production applications and go-to-market workflows had grown fragile, creating a significant engineering burden on two fronts.
On the infrastructure side, the ML team had built a custom pipeline to sync data from the Databricks Lakehouse into Redis and DynamoDB. The pipeline powered eligibility rules for promotions and free trials, things like whether a user had activated a specific feature or logged enough sessions to qualify for a discount. “It was the right solution when it was built. But priorities shifted, the team moved on to other things, and the system stayed behind,” said Michael Kobelev, ML Infra Software Engineer. Recently, a routine library update broke sync jobs writing to Redis. Without centralized alerting, the team couldn’t identify all the affected jobs and some failures didn’t surface until weeks later.
On the go-to-market side, the gap between data and action was just as costly. Sales and customer success reps manually copied metrics from dashboards into PowerPoint templates, spending roughly 30 minutes per deck, multiple times a week. “Data teams build incredible pipelines and tables, but there’s still a last mile between what’s in the warehouse and what a sales rep can actually use,” said Maximilian Proano, Software Engineer, Data Applications. When the team built Deckster, an LLM-powered app that automatically generates customer-facing presentations, the first version pulled metrics live from a SQL warehouse. Latency spiked on every click, and there was no clean way to cache results or persist a user’s progress.
Both teams needed the same thing. A managed transactional layer that sat close to their Delta tables and required minimal upkeep.
