Citework

Problem

Language models will invent a leverage multiple if you let them do the arithmetic. An investment committee cannot act on a memo where the numbers might be fiction and the quotes might be paraphrases.

What I built

Citework turns a folder of deal documents into a one-page credit memo with a paper trail. The pipeline is two model calls with Python in between. Step one extracts named figures, each with a source file and a verbatim quote, and does no arithmetic. Python then verifies that every quote actually appears in the named file (whitespace folded, paraphrases fail, a 2 inside $52M does not count) and computes net debt, leverage, ARR multiple, and NRR in Decimal; a non-positive denominator is N/M, never a negative multiple. Assumptions, conflicts, gaps, and the verified numbers are written to disk before a human gate. Only after approval does step two write prose, and it is handed the numbers block, not the raw documents, so the memo cannot disagree with the table.

The world is closed: documents are the only source of truth, and file text is treated as untrusted data, never as instructions. A Streamlit app reviews the artifacts without calling the model; an offline mode replays captured responses so the ~70 tests run without a key.

Outcome

A shipped example memo for a fictional mid-market software LBO is in the repo, with every figure traceable to a quote. The pipeline has no deal-specific logic, so a different folder of documents runs the same way.