Work, Studio product
Sage Prism
An LLM research pipeline that reads SEC filings and finds what changed.
The problem
Public companies file tens of thousands of pages a quarter. The interesting part is what changed since the last filing, and finding it by hand takes an analyst a day per company. Sage Prism exists to find it for a video channel aimed at retail investors, with every claim traceable to a filing line.
The approach
A daemon watches the EDGAR daily index for a rolling set of the largest companies by market value, downloads each filing, and runs two detectors: a numeric one over XBRL series that survives tag drift and restatements, and a narrative one that diffs sections year over year. Only filings with a signal go to research, where a model builds a fact base that must pass a mechanical grounding gate before it is stored. A separate command line step turns fact bases into scripts. The daemon never writes scripts and the script step never does research, which is what keeps the output grounded.
Where it is now
The pipeline runs on a single machine with a state machine per filing, retries and a dead letter after three failures. A person records, edits and uploads everything. Nothing publishes itself.
Built with