Project Overview
Built an Internal Developer Platform (IDP) using knowledge graphs to accelerate developer onboarding and improve DORA metrics at openHPI. Mapped 900+ Ruby on Rails components using novel multi-run AI consensus methodology with 0.8+ confidence thresholds.
Developers can now query codebase semantics, trace dependencies, and understand architectural patterns through interactive D3.js visualizations.
Organization: Hasso Plattner Institute (openHPI)
Role: Content Manager
Technical Architecture
AI & Analysis
- LLM: Google Gemini for component classification
- Methodology: Multi-run consensus (5 runs → 0.8+ agreement)
- AST Parsing: Ruby syntax analysis
- Embeddings: Semantic similarity for clustering
Visualization
- Frontend: Next.js, TypeScript, D3.js
- Graph Engine: Force-directed layout
- Features: Semantic zoom, filtering, dependency tracing
- Documentation: Interactive markdown with cross-links
Novel Methodology: Multi-Run AI Consensus
Single LLM runs showed 40% variance in component classification. The solution:
- Run 5 independent LLM analyses on each component
- Statistically aggregate results
- Only keep classifications with 0.8+ agreement across runs
- Flag low-confidence items for human review
Result: Reliable component classification with quantified confidence levels, enabling trust in the knowledge graph for real developer workflows.
Challenges & Solutions
LLM Analysis Inconsistency
Problem: Single-run LLM analysis showed 40% variance across runs, making results unreliable for production knowledge base.
Solution: Multi-run consensus methodology with statistical aggregation, achieving reproducible classifications with quantified confidence.
Large Codebase Overwhelm
Problem: 900+ components to manually analyze would take months.
Solution: Automated AST parsing + semantic embedding + LLM-powered classification with human validation for high-impact components. Achieved 68% source code coverage in 2 months (against a 6+ month manual estimate).
Developer Onboarding Bottleneck
Problem: New developers took 2-3 weeks to become productive.
Solution: Interactive portal with component relationships, workflow documentation, and visual dependency graphs. New developers now productive in 1 week.