Investment-committee memos
A full memo assembled from the deal’s own evidence base — documents, notes, references, scores — with each claim pointing at the record it came from.
Initializing workspace...
The assistant reads your fund’s own records and proposes. Fund maths — TVPI, DPI, IRR, ownership, dilution, capital accounts, fees, hurdle, catch-up, carry, concentration, runway — is calculated by the system, not by the model. That split is the whole design.
The model reads your fund’s own records and drafts. The arithmetic belongs to the system, and every write belongs to a person.
Memos, letter drafts, cited answers and proposed changes, all on evidence you can open.
Fund maths runs as code over your own records. Where an input is missing, the calculation refuses rather than returning a number that looks plausible.
A full memo assembled from the deal’s own evidence base — documents, notes, references, scores — with each claim pointing at the record it came from.
Ask questions of one deal’s documents and get answers that cite the page they came from.
The same, across the book: runway, concentration, who has not reported. Figures come from the ledger; the assistant explains them.
A quarterly letter drafted from live portfolio data, which a person then edits and sends.
Restructures a Pages document as ordinary, undoable edits — not a black-box rewrite you have to accept whole.
Reference responses summarised for the file, and records tagged automatically so search finds them later.
Column mapping on a messy spreadsheet, proposed and shown before anything is written.
Masking happens on the way out of the system, so the protection does not depend on the model choosing to behave.
Every write is a human action. The single exception is proposing a milestone change, which still requires a person to approve it.
A response that would disclose something across a boundary is withheld rather than sent and apologised for.
Text inside your own records is treated as data, never as instructions — and every AI action lands in the audit log.
AI is on every paid plan and metered in monthly credits. Allowances are set well above what a normal fund uses, exceeding one is a top-up rather than a wall, and a failed generation is refunded automatically. The free plan has no AI at all.
Larger funds can run the models under their own provider contract and data-processing terms, at which point the credit cost largely goes away and DeelSignal charges a small infrastructure fee on usage.
Enterprise workspaces can connect their own tools to the assistant through MCP servers, declared on the same connection register as everything else.
A better model writes better prose and makes better judgement calls, which makes this product better. No model makes the numbers right — those come from running code over your own records. Every model improvement helps; none of them makes the split redundant.
It does not do arithmetic, and where the inputs for a figure are missing the calculation refuses rather than returning a number that looks plausible. It also has no AI on the free plan, and no plan is unlimited — an uncapped allowance would mean an uncapped bill.
Forty minutes, a workspace shaped like your fund, and no deck. If it does not replace something you already pay for, we will say so and end early.