Private equity teams have to form a view of a company from thousands of documents in a few weeks, so most of a data room gets skimmed. Winback reads all of it and cross-checks what management claims against what the underlying documents say, like a recurring revenue figure that the customer contracts do not support.
Every document is split into blocks with permanent ids, and every statement the model makes has to cite one. The server checks each citation against the real blocks before anything reaches the screen, so a made-up reference is dropped rather than shown. Benchmarks are computed in code from extracted values, never by the model, and nothing issues a verdict: an analyst accepts, edits or dismisses each finding before the memo is drafted.
My part was mostly on the Sunday: fixing bugs in what had been built on Saturday, and getting it to work on real uploads with parsers for PDF, Word and Excel, block-level document storage, background ingestion and the upload screens. The weekend itself is in Someone else had our idea.