Local project preview

Incident Atlas

A self-improving agent team for building a living knowledge base from scattered team documents.

Open GitHub repository Visit project website Watch demo video

Project description

Incident Atlas coordinates focused agents, scores the quality of each completed run, and uses reviewer feedback to improve the next attempt. The team built the full workflow during the hackathon, using Band for the core agent loop and Replay.io for evaluation, communication, or runtime support. The demo includes a visible activity trace so judges can inspect what each agent did and where human approval was required.

Project links

Technical architecture

LlamaIndex ingestion agents normalize new documents, an evaluator checks retrieval quality, and a curator agent rewrites weak metadata before re-indexing.

Setup instructions

Clone the public repository, install dependencies, copy .env.example to .env.local, add the documented API keys, and run the development command in the README.

Lessons learned

The team found that clear handoffs and fast evaluation loops mattered more than adding more agents. Band handled orchestration while Replay.io made the result easier to inspect and improve.

Additional screenshots or links

https://docs.llamaindex.ai/
https://github.com/run-llama/llama_index

Tools used

Band

Replay.io

Back to project gallery