Atlas

• 

  Landing

• 

  Sample Graph of a Repo

• 

Inspiration

The "painful Tuesday" framing: new engineer, 12 microservices, outdated Confluence doc

What it does

The three outputs (3D graph, edge panels with code evidence, agent-ready markdown), with emphasis on edges as first-class objects

How we built it

the 5-stage scan pipeline (Clone → Discover → Embed → Probe → Synthesize), the role of each service, the tech choices

Challenges we ran into

code embeddings vs. prose embeddings, extracting meaningful edges (not just imports), grounding the assistant, 3D navigation usability

Accomplishments that we're proud of

the unbroken evidence chain, humans+agents sharing one map, runs on a laptop with no GPU, honest confidence model

What we learned

LLMs at the end not the start, edges > nodes, ONNX quantization is production-ready, grounding is an architecture problem not a prompt problem

What's next for Atlas

private repos, incremental scanning, polyglot probes, MCP server, team annotation layer

Built With

• backboard.io
• claude
• cursor
• factory.ai
• python
• typescript
• vercel