Many people already ask ChatGPT or Claude about their stocks, and for a lot of questions that works well. What is a P/E ratio? Summarise this earnings call. Explain this filing. We use the same kind of model inside Iris.
But watching a portfolio is a different job from answering a question. It means noticing, every morning, the few stories out of thousands that reach the stocks you own, including the ones you would never think to ask about, and being able to show why.
Three things a chatbot on its own struggles with
- It starts from your question. A model can explain how a supplier relates to your stock once you ask. It will not reliably raise the question itself, for every holding, every day.
- It infers relationships fresh each time. Without a stored map of who supplies whom, every answer re-derives the connections from training data, which may be out of date or simply wrong.
- Its paper trail can be invented. Language models can produce a confident explanation, and occasionally a confident source, for something that is not true. When the subject is your money, you need to see where a claim came from.
How Iris divides the work
Iris uses language models where they are strong and structured data where they are not. Models read news and filings into structured facts, judge whether a story describes a real event, and write the explanation you read. Hivemind’s knowledge graph, a stored map of supplier, customer and competitor relationships, decides what connects to what. Your portfolio decides what matters.
When you ask Iris a question, it starts from that same context: your holdings, the graph, and the day’s stories, rather than a blank prompt. And every story it shows you links to its source.
What this is not
It is not a claim that Iris is smarter than a frontier model; it uses them. It is not a forecast engine: Iris explains what moved and through which link, and does not predict prices. It is a different division of labour, one we think suits a job where being traceable matters more than being fluent.