Five specialized agents prepare, review, challenge, and synthesize a fund's document set. Every figure is tied to code-verified arithmetic, not the model's mental math.
Five specialized agents move a fund's documents from raw upload to a reconciled, severity-ranked finding set. Each stage hands verified figures to the next.
Upload the fund's document set. Each PDF is profiled once and cached, and every agent reads the native file rather than a lossy text dump.
The Preparer extracts one cross-document figure set: NAV bridge, balance sheet, capital accounts, and stated performance.
Two agents work in parallel. The Reviewer validates disclosure and ILPA expectations; the Challenger adversarially stress-tests valuations and assumptions.
The Synthesizer reconciles every output at an EQR gate, then returns a severity-ranked finding set with a PBC list.
The pipeline is built to be checked, not taken on faith. Arithmetic is deterministic, review is adversarial, and every finding points back to a figure on a page.
Numbers are checked by deterministic code, not the model's mental math. Ten cross-statement checks run on every engagement.
An independent Challenger exists to break the Preparer's work, not to agree with it. A finding survives only if it holds up to a second, skeptical pass.
Each finding traces to a specific document and a passed or failed check, so a reviewer can vouch it in seconds.
Prompts and checks adapt to the structure in front of them: private equity, venture, hedge, credit, or real estate, with ILPA 3.0 expectations built in.