How it works
What happens between upload and memo
- 01
Drop in the deck and data room
Paste the text or upload the PDF. A pitch deck, a board update, a CIM, a data-room folder. Ten pages or fifty. Meridian works out what it is reading: the sector, the stage, the document type. Nothing to configure.
Text-based PDFs up to the document limit. If a file is scanned images rather than extractable text, Meridian says so.
- 02
Meridian reads and structures the memo
The whole document is read in one pass where it fits. For longer data rooms the pieces are kept in sequence and merged. The model is told what a diligence memo includes: the business, the market, the numbers, the traction, the risks and the questions.
Documents longer than the processing limit are trimmed and the word count of what was dropped is reported. Nothing is dropped silently.
- 03
Review risks and questions
Risks come back with a severity (high, medium, low) and the exact line from the deck the risk is based on. Questions are tied to what the memo found, not to a generic checklist. Every evidence quote is located in the source before it appears.
A risk quote that cannot be found in your document is dropped as evidence but the risk itself is kept. A bad reference is worse than no reference.
In detail
The five stages
Written out because the trustworthy part of this product is not the model call. It is what happens either side of it.
- 01Intake
Getting the text out
Pasted text is already usable. A PDF is a set of rendering instructions that usually contain extractable text and sometimes contain only images of text. Meridian pulls the text out server-side, normalises whitespace and spacing, and if what comes back is near-empty it tells you the file may be a scanned export rather than running a memo on nothing.
- 02Reading
One pass, whole document
The full deck goes to the model in one request where it fits, because a concentration-risk line means something different next to a slide that shows 22% of ARR from one customer. Documents longer than one pass are split on paragraph or section boundaries, never mid-sentence, and the parts are reassembled before the memo is structured.
- 03Verification
Every evidence quote checked against your text
The model returns an evidence quote for each risk it surfaces. Before you see it, that quote is located in your own source document, first as an exact match, then insensitive to curly quotes and extra spaces, then by the longest substring that does appear. A quote that cannot be found is removed as evidence. A highlight on the wrong line is worse than no highlight.
- 04Structuring
Memo sections in order
Business summary, market and comparables, key metrics, traction, risks ranked by severity, founder questions. The sections are written in the order an investor reads a first-pass memo, so the most urgent information is at the top. A clean deck with no red flags comes back with few high risks, because rating everything high would make the rating useless.
- 05Questions
Specific to this deck, not the template
The founder questions are generated from what the memo found, not from a fixed list. If the deck shows high churn and a concentrated customer base, the first two questions are about those, with the rationale for why each one matters. The goal is five questions you can ask in the first call rather than thirty you have to read first.
Questions