Atlas

Inspiration

During an incident, the useful answer is rarely in one place. A release note explains what changed, a team conversation records the symptoms, and an older incident may contain the missing context.

We built ATLAS to bring that evidence together, helping teams investigate without reconstructing the same history from scratch.

What it does

ATLAS is a company-brain prototype built around OrionCart, a fictional e-commerce company experiencing checkout failures.

Users can upload company documents, investigate questions, follow citations, inspect source versions, and save incident reports. Engineering and Support demo roles receive different evidence, with filtering applied before content reaches the model.

The dashboard displays sources in recorded order. TriggerWare supplies synthetic operational signals through a real local server and its official Python SDK.

How we built it

The interface uses React, TypeScript, and Vite. A Python FastAPI service stores documents, source versions, settings, and reports in SQLite. Keyword retrieval selects relevant records, and a language model reasons over the resulting evidence.

ATLAS supports local OpenAI-compatible models and ChatGPT subscription sign-in. The subscription provider uses OpenAI’s Sign in with ChatGPT DevKit and macOS Keychain for protected credential storage.

TriggerWare executes a fixed SQL query against a synthetic OrionCart signals table. Those results are preserved in company memory and included in investigations.

We reuse App Starter’s shared organization-role enum for demo document-editing capabilities. We did not adopt its complete application scaffold.

Challenges

The main challenge was preserving context without overstating what the evidence proves. ATLAS retains source references and versions, distinguishes likely causes from confirmed causes, and displays provider failures explicitly.

Integrating model sign-in also required keeping credentials outside the browser and repository.

What we learned

A useful company brain needs traceable evidence and durable memory as much as a capable model. Provenance, permission boundaries, and visible uncertainty make an investigation easier to evaluate.

We verified that newly uploaded information can change an investigation’s answer while remaining restricted to its selected demo audience.

What’s next

Next steps include authenticated identities, resource-level permissions, semantic retrieval, and richer TriggerWare connectors.

The current application is a local prototype using synthetic data and simulated roles. Its action saves an incident report locally. It does not implement production IAM, Cognee, live Slack, Bluesky, or autonomous monitoring.