Atlas
A market research app that reads a live stream of financial news, turns it into warnings and predictions for individual stocks, and shows them next to fundamentals and model-written analysis.
Overview
Atlas brings together the three things you check before looking at a stock: what the company's numbers say, what the news says, and what a careful reader would conclude from both. A browser lists stocks with live data; each one opens a page with fundamentals, recent warnings and an AI-written analysis.
How it works
- A news streamer fetches new articles and publishes them to Kafka.
- A warning streamer has a language model read each article and turn it into warnings for the specific stocks it affects.
- A prediction streamer combines new warnings with fundamentals and a live quote, has the model update that stock's outlook and target price, and stores one prediction per stock in MongoDB.
- The Flask API reads the warnings and predictions and serves them with fundamentals and the written analysis.
- The React frontend shows it all; a single command, npx github:CosmicVeil/Atlas, clones, installs and starts the whole thing.
Decisions worth explaining
Three stages joined by a message broker
Each stage only reads from one topic and writes to the next, so a slow or crashed stage doesn't take the others down, and messages wait in Kafka until it recovers. A single script doing all three would be simpler, but one failure would stop everything.
The model judges; the code does the arithmetic
The language model sets the outlook and target price, but the potential return is calculated in code. Language models are unreliable at arithmetic, and a number computed in code can be tested. If a model call fails, nothing is written, so a made-up fallback can never pass for a real prediction.
Works without any API keys
Without keys, Atlas falls back to local fundamentals and simulated analysis instead of failing. Anyone can try it in one command, and real providers switch on when keys are added.
Tests that never touch the network
The backend tests mock the news source, the model provider, Kafka and MongoDB, so they run offline, quickly and the same way every time.