Arclo

• 

  Ai Codebase Review & Learning Oracle

Inspiration

Every developer has been there — a code review tool flags a bug, you apply the fix, and three other things break. Existing tools review files in isolation. They hallucinate fixes. They output jargon that junior developers cannot act on. We wanted to build something that actually understands a codebase the way a senior engineer does — seeing every connection, every dependency, every ripple effect before touching a single line.

What it does

Arclo is an AI-powered codebase intelligence agent. You paste a repo URL and it:

Builds a live dependency graph of every file, function, and import Finds bugs ranked by blast radius — how many other files they affect Explains every issue in plain English, no jargon Generates a fix, writes a unit test, runs it in a sandbox, and only commits if the test passes Flags uncertain fixes instead of guessing — honest uncertainty beats confident hallucination

How we built it

Gemini 1.5 Pro via Vertex AI for bug analysis, fix generation, and self-verification Google Cloud Agent Builder orchestrating a 10-step pipeline from repo ingestion to commit MongoDB Atlas storing the dependency graph as nodes and edges, with vector search for semantic pattern matching GitLab MCP for live commits back to the repository FastAPI backend on Google Cloud Run React + Vite + React Flow for the graph UI and fix dashboard Docker sandbox on Cloud Run for safe test execution before any fix is committed

Challenges we ran into

Building a dependency graph that actually captures cross-file relationships accurately across different languages and patterns Getting Gemini to reliably self-check its own fixes without just agreeing with itself — the hallucination guard required careful prompt engineering Keeping the sandbox isolated enough to be safe but fast enough to be useful in a live demo Ranking bugs by blast radius in real time without making the graph query too slow on large repos

Accomplishments that we're proud of

The hallucination guard actually works — Arclo says "I am not certain" instead of shipping a bad fix Full loop from repo URL → dependency graph → bug analysis → fix → test → commit, end to end The confidence scoring system (0–100) with three tiers of action feels genuinely useful, not just a feature checkbox Plain English bug cards that a junior developer can actually read and act on

What we learned

What's next for Arclo

GitHub support alongside GitLab Multi-language depth — current version handles Python and JavaScript best; adding Go and Java Codebase health score over time — track how a repo improves (or degrades) across weeks Team mode — assign bugs to developers, track fix velocity, spot which modules keep breaking IDE plugin — bring Arclo's graph intelligence directly into VS Code

Built With

• atlas
• auth
• docker-apis-&-integrations-gitlab-mcp
• google-cloud-agent-builder-database-mongodb-atlas
• google-cloud-build
• google-cloud-vertex-ai
• javascript
• languages-python
• mcp
• mongodb
• mongodb-vector-search-devops-&-cloud-google-cloud-run
• react
• react-flow-ai-&-models-gemini-1.5-pro
• tailwind-css
• typescript-frameworks-fastapi
• vite