How Nubank split a 6-million-line ETL monolith with Devin

Nubank had Devin carry out the repetitive sub-tasks of breaking its ETL monolith into sub-modules, with engineers reviewing every change. Cognition reports large gains in speed and cost.

The problem

Nubank's eight-year-old ETL system had grown to over 6 million lines, with dependency chains up to 70 deep. Splitting it meant moving about 100,000 data class implementations. The plan called for roughly 1,000 engineers over 18 months.

How they did it

1. Collected examples of migrations that engineers had already done by hand.
2. Fine-tuned Devin on part of those examples and kept the rest as a benchmark.
3. Had Devin complete migration sub-tasks on its own while engineers reviewed and approved the changes.
4. Let Devin write its own scripts for recurring mechanical steps, such as detecting country extensions from file paths.

Results

• As reported by Cognition: an '8x engineering time efficiency gain' and '20x cost savings'.
• The story describes 'a 12x efficiency improvement in terms of engineering hours saved' and an '8-12x faster migration'.
• After fine-tuning, Cognition reports a 'doubling of Devin's task completion scores' and a '4x improvement in task speed'.

As reported by the source (Cognition (Devin) customer story); AgentGid did not measure these figures.