GitLoom
Ask a codebase questions. Follow the answer back to the files.
Overview & Context
GitLoom indexes entire GitHub repositories into vector embeddings so engineering teams can query architecture, review AI-generated commit summaries, and search meeting transcripts grounded in code references.
The Problem
Onboarding to large codebases or tracing architectural decisions across commits and meetings is fragmented across documentation, git logs, and call recordings.
The Architectural Approach
Engineered an automated ingestion pipeline using Octokit and Gemini embeddings stored in PostgreSQL with pgvector. Built tRPC endpoints for source-grounded Q&A that cites specific repository files, automated diff summarization on new commits, and an audio meeting pipeline powered by AssemblyAI chapters to surface action items and issues.
Key Technical Highlights
- pgvector semantic search over repository code embeddings with citation links
- AI-generated commit diff summaries and saved technical Q&A threads
- AssemblyAI meeting audio transcription with timestamped issue extraction
Quill
Upload a PDF and ask questions grounded in its contents using embeddings, vector retrieval, and streamed AI answers.