Shipped: Building Graphos: Memorise Massive Codebases

After 6 months of building, I’m finally starting to get some very interesting results with Graphos, my open-source project aimed at creating a codebase memory based on graph structures.

The core idea: instead of dumping massive codebases and docs directly into giant context windows, Graphos transforms Abstract Syntax Trees (ASTs) and documentation into queryable graphs. This is particularly useful for large legacy or complex codebases where business rules and documentation are outdated, fragmented, or entirely missing.

What works today:
Because the graph structure pre-digests and isolates the relevant architectural context, I can already generate C4 architecture diagrams and arc42 documentation on large codebases 100% locally using a small 9B parameter model.

Next steps & Help needed:
Right now, the core pipeline works, but my main focus is shifting toward cleaning up the repository, refactoring, and making the graph-theoretic approach much clearer and more robust.

I’m currently leaning into leveraging graph algorithms (like community detection, centrality metrics, and traversal patterns) to automate architectural discovery and trace data flows cleanly without relying on heavy compute.

If you know of any solid research papers, literature, or architectural patterns around code-to-graph representation, software knowledge graphs, or graph-based static analysis, I would love to hear your recommendations! What directions or algorithms would you explore next?