Nubank refactors 6M lines of ETL code 8x faster and at 20x lower cost using AI agent Devin

Background

Nubank's 8-year-old ETL monolith had become a bottleneck. With 6M+ lines of code and dependency chains reaching 70 levels deep, the architecture could no longer support the company's growth across three countries. A manual migration involving 1,000+ engineers and 18 months was the only available option before Devin offered an alternative.

What Was Implemented

• Deployed Cognition's Devin AI software engineer to handle autonomous ETL sub-module migrations
• Taught Devin the migration pattern through examples of previously manual migrations (fine-tuning)
• Ran multiple parallel Devin agents on individual data class migration sub-tasks (~100,000 total)
• Maintained human oversight: a project manager reviewed and approved Devin's pull requests
• Devin built its own helper scripts to automate the most repetitive mechanical components
• Benchmark evaluation set confirmed fine-tuning results (2x task completion score, 4x task speed)

Results

Individual business units (Data, Collections, Risk) completed migrations in weeks instead of months or years . Across the migration scope delegated to Devin, the project achieved an 8–12x engineering efficiency gain and over 20x cost savings . After fine-tuning, Devin's task speed improved 4x (from ~40 minutes to ~10 minutes per sub-task) and task completion scores doubled.

Lessons

• Large-scale, repetitive refactoring tasks with high variation (many edge cases) are well-suited to AI agents: human expertise is needed once to define the pattern, then the agent scales it
• Fine-tuning an AI agent on domain-specific examples significantly improves performance beyond generic capability — Nubank observed 2x completion scores and 4x speed after fine-tuning
• Human-in-the-loop review (approving PRs) provides a quality gate without negating efficiency gains
• AI-agent cost economics are most compelling for tasks combining high volume, high repetition, and scarce human resources