RepoLens

• RepoLens

Codebases are black boxes. Understanding architecture, dependencies, and security risks can take weeks. RepoLens turns that process into a conversation.

-> What it does? RepoLens is an AI-powered codebase intelligence and security platform.

• Ask questions about any repository: and get answers with exact file citations.
• Semantic code search: powered by vector embeddings and pgvector. -Functional UI Prototype of an Attack Surface Analyzer built with React Flow, designed to eventually map out APIs and gateways

Instead of reading thousands of files, developers can interrogate their codebase in plain English.

-> How I built it?

Built from scratch without LangChain or heavyweight abstractions.

Repositories are parsed, chunked, embedded, and indexed for ultra-fast retrieval. User queries are converted into vectors and matched against the most relevant code chunks using cosine similarity before AI synthesis.

-> Challenges

• Serverless limitations: Vercel environments lack git and tar, so we engineered a fully in-memory GitHub repository extraction pipeline.
• AI rate limits: Large repositories triggered Gemini 429s, leading us to build aggressive retry and exponential backoff systems for resilient imports.

-> Accomplishments

• Built an interactive visual threat-mapping engine with React Flow.
• Achieved near-instant semantic search on large repositories using raw pgvector queries.
• Created a dependency-free repository import pipeline capable of handling private branches securely.

-> What I learned?

• Raw SQL + pgvector delivers dramatically better performance than ORM abstractions for vector search.
• Reliability matters as much as AI quality - graceful retries and failure handling are essential for production systems.

-> What's next?

-Automatic GitHub webhook synchronization, enabling continuous re-indexing so RepoLens stays perfectly aligned with the latest code changes.

Built With

• nextjs
• pgvector
• postgresql
• prisma
• react