Back to ProjectsAI-Powered Developer Tooling — IBM Bob Hackathon · May 2026
Seenior — AI-Powered Developer Onboarding
Ramp-up time: hours → under 60 secondsTech Stack
ReactNode.jsGoogle Gemini APIGitHub REST API (Octokit)Mermaid.jsD3.jsXYFlow
Objective
Cut the time it takes a developer to understand an unfamiliar codebase from hours of manual file-reading to under a minute.
The Problem
Onboarding onto a new repo means piecing together architecture and business logic by hand, one file at a time — slow, and easy to come away with the wrong mental model.
Approach
Built an end-to-end LLM app that pairs Gemini 2.5 with the GitHub REST API (Octokit) to auto-generate documentation, architecture/UML diagrams, and comprehension quizzes from any public repository. Diagrams render as interactive, clickable Mermaid.js visualizations through a custom validation pipeline built with D3.js and XYFlow.
Challenges
- LLM-generated Mermaid syntax frequently came back malformed, so raw output couldn't be trusted directly — needed a validation and retry layer before rendering.
- Balancing diagram detail against readability so large repositories with many files still produced a diagram a human could actually parse.
What I Learned
- Validating and repairing LLM output programmatically is more reliable than trying to prompt-engineer perfect output every time.
- Turning static LLM text output into an interactive, navigable visualization changes how useful it is — the diagram becomes something you explore, not just read.