Inspiration: The Ghost in the Machine
In large organizations, there's a silent crisis brewing. Buried deep in data centers are legacy applications—some running for 20+ years—that nobody dares to touch. They work, they're stable, and most importantly, they're critical to business operations. The problem? The subject matter experts (SMEs) who built and maintained these systems retired years ago, taking their institutional knowledge with them.
I witnessed this firsthand when we needed to migrate several legacy applications together. The documentation was outdated, the code repository contained versions that didn't match production, and the actual running server had undocumented customizations. When a critical bug emerged, the team spent weeks manually piecing together how the system actually worked—reading old docs, analyzing code, and SSH-ing into servers to verify reality against documentation.
In an era where AI can write code and analyze complex patterns, why are we still manually reverse-engineering our own systems? The cost isn't just time—it's risk. Every migration, modernization, or critical fix becomes a high-stakes archaeological dig.
What I Learned
Building this multi-agent system taught me that context is everything. A single AI agent analyzing just code or just documentation will miss critical discrepancies. The real insight comes from cross-referencing multiple sources—like a detective gathering evidence from different witnesses.
I learned that orchestrating multiple AI agents requires careful coordination. Each agent needs to be specialized (documentation analysis, code analysis, server inspection) yet able to communicate findings. The challenge was designing a system where agents could identify contradictions: when documentation says the database is PostgreSQL but the code connects to MySQL, or when the repository shows version 2.0 but the server runs 1.8.
How I Built It
I architected a custom multi-agent orchestration system using AWS Bedrock's Claude 3.5 Sonnet as the reasoning engine. The system employs four specialized agents:
- Documentation Agent - Parses PDFs, Markdown, and Word docs
- Git Analysis Agent - Examines code structure, dependencies, and architecture
- Server Analysis Agent - Inspects running environments via SSH
- Summarization Agent - Synthesizes findings and identifies discrepancies
The orchestrator manages agent dependencies and ensures findings are cross-referenced. When the Git agent discovers React components but documentation claims Angular, the system flags this as a critical discrepancy.
The code is mostly written using Kiro.
The Path Forward
This is a proof of concept demonstrating that AI-powered legacy analysis is feasible. With refinement, this approach could save organizations countless hours and reduce the risk of failed migrations. The ghost in the machine doesn't have to remain a mystery.
Built With
- amazon-web-services
- bedrock
- python
Log in or sign up for Devpost to join the conversation.