Market intelligence engine

Retail investors get raw data or generic commentary, rarely analysis tied to a specific position. The tempting shortcut is to hand a model the numbers and ask it to reason, which produces confident prose built on arithmetic nobody checked.

Here the split is explicit. Indicators, backtests and statistics are computed deterministically; the model reads those results, weighs conflicting signals and writes the explanation. It never calculates anything it then reports.

Two models, chosen by task

News sentiment runs across thousands of articles, which is volume work with a narrow output. That goes to a self-hosted open-weight model. Weighing conflicting signals into a view goes to a frontier model. The split is the difference between a viable cost base and an unviable one.

Arithmetic never touches the model

Every number in a report comes from computed output passed in as context. The model's job is interpretation. A system that lets a language model calculate a return and then report it is one silent error away from being worse than useless.

Backtesting is the honesty mechanism

A signal that sounds compelling and has never been tested against history is an opinion. Running it over historical data first is what separates analysis from narrative, and it is the same instinct as building an evaluation harness before tuning a retrieval system.