AWS Accelerates Oracle-to-PostgreSQL Migration with Amazon Bedrock-Based Multi-Agent GAMMA

The AWS GenAI Innovation Center and Professional Services developed GAMMA (Generative AI for Migration and Modernization Accelerator), a multi-agent framework based on Amazon Bedrock. By using AI to handle complex Oracle stored procedures and business logic transformations that were difficult to automate with existing tools, it reduced manual effort by an average of 51% and achieved a code conversion accuracy of 72% in enterprise-scale migrations.

GAMMA's 3-Step Workflow

GAMMA consists of three stages: Modernization assessment, planning, and execution. First, multi-agents construct a dependency graph between schemas and procedures to identify complexity hotspots and risk areas. Next, they identify business domain boundaries across the entire portfolio and classify optimal strategies such as direct conversion or refactoring. In the execution stage, LLM-based code conversion and semantic retry mechanisms analyze and resolve errors, generating a confidence score from 0 to 100 to prioritize review by subject matter experts (SMEs).

Accumulation of Knowledge Items (KI) and Performance Improvement

Repetitive transformation patterns are accumulated as Knowledge Items (KI) for reuse. Nine key mapping rules, such as mapping NVL to COALESCE and SYSDATE to CURRENT_TIMESTAMP, are stored as KI. As KI accumulates, the average confidence score rises from 71% to 89%, and the number of retries decreases from 4.2 to 0.9. Consequently, transformation time is reduced from approximately 45 seconds to 18 seconds, and the proportion requiring SME review drops significantly from 65% to 18%.

Benchmarks and Customer Case Studies

As of August 2026, benchmarks showed that converting over 300 mappers across three applications took 3.5 hours and cost approximately $95, with 100% of mappers converted 1:1. In the case of Mitsubishi Electric Building Systems Group, GAMMA was adopted for a project migrating 5,000 databases from Oracle to PostgreSQL, increasing the accuracy of previously unconvertible code to 67% and achieving an object-based success rate of 90%. This reduced the manual workload by 48%, equivalent to approximately 200 person-months.