All work
Work · Tool for investors

Deck reader

A tool for investors. You hand it a startup deck, it hands you back what is missing from it and the questions to ask the founder before the call. Never an opinion on whether to invest.

100% fictional demo deck, real mechanism.

A deck lands. You flip through it on your phone, between two other things. A team that looks solid, a market that sounds big, a nice curve. You remember the founder's energy and two figures. What you do not remember is what was not there.

And that is what matters: how many users were interviewed, who is full-time, what was tested and then dropped, where the figure on page 5 comes from. You would ask every one of those questions if you had an hour per deck. You have ten minutes.

The reader asks them for you. It hands you a report, next to the deck: what is missing, what does not hold, and the questions to ask the founder before the call.

Why the gaps are invisible

A deck is built to show what is going well. The founder spent weeks on it, and every page was reworked to convince. The gaps sit between the pages, and to the naked eye you do not see them.

Once the company has customers, at seed, the problem changes shape. There are figures, and a figure on a page is reassuring: a monthly revenue, a retention rate, four named customers. What the page does not say is where it comes from, and whether it still holds once you open the billing export.

At series A there are one or two years of figures, and the question changes again. No longer "does anyone pay", but "does the machine repeat": does the next customer cost the same as the last one, bring the same, sign without the founder in the room? That does not show on a page. It shows in a monthly P&L and in cohorts.

What the reader does

A PDF goes in, a report comes out. In between, eleven steps, each one isolated from the others so the AI cannot make up what is missing. On the right, the report of a fictional deck fills in as the steps go by.

You givedeck.pdfand, from seed on, the documents behind its figures
You getdeck.reading.pdfand the same reading as markdown, next to the deck on your disk
  1. First, for every deck
    1. The PDF is transcribed page by page, and every sentence keeps its page number. That is what makes quoting possible.

      Headerpreseed-deck.pdf · 12 pages · read on 2026-09-08Quotes verified against the PDF text layer
    2. Sector, business model, announced stage, each with its quote. The stage picks the list of questions: 24 before the first customers, at pre-seed, 26 once there are some, at seed, 32 at series A. Before customers, we look for what the founder learned in the field. After, who pays and who comes back. At series A, whether the machine repeats. The business model adjusts the list: a marketplace, a hardware company or a biotech do not get the questions of a SaaS. A later-stage deck is refused in one line.

  2. if series AIf it is raising a series A, it demands the documents before reading
    1. Always the same six: the monthly P&L over 24 months, cohorts over 12 months or more, the CRM export with the weighted pipeline, the cap table, the three-year financial model, the contracts of the top 10 customers. Every annex received is sorted into one of the six slots, with a quote the code checks. One document missing, or too short, and it stops there: no reading, but the email that asks for it, document by document. There is no coverage threshold, the list is the gate.

  3. if seed or series AIf the company has customers, it checks its figures
    1. Every verifiable statement, one per line: figures, named customers, competitors, founders, past funding. Page and quote each time.

    2. Revenue export, cohorts, financial model, cap table: every deck figure is looked up in them. Under 25% off, noted with no penalty. Up to double, a question for the call. Beyond that, blatant. With no documents, it stops and drafts the email that asks the founder for them.

    3. Only for what a document cannot settle: a named customer, a competitor, a founder's track record. It searches for and against, and only concludes with two independent sources. Otherwise it writes "unverifiable" and makes nothing up. At series A it also looks at public reviews, open job posts and the LinkedIn profile of every announced key hire.

    4. Before calling anything false, a second AI, which has seen none of the rest, looks for an honest explanation: a date, a unit, a scope. With one, it becomes a question for the call. Without one, it stops and shows you the sources on both sides. Everything read so far stays in the report, and you decide.

  4. Then, for every deck
    1. The questions are grouped by theme: team, market, evidence, money. Each theme is read by an AI that sees only its own questions. No way to excuse a gap on the team side with a nice sentence on the market side. Every answer is found, partial or absent, with the page and the exact quote. A quote that is not word for word in the deck is rejected. A figure that is not on a page does not exist.

    2. On borderline questions, two readings of the same deck do not always agree. When the result lands in the grey zone, between 65 and 80%, it reads twice more and keeps the majority answer. Disagreements stay marked in the report.

  5. if series AIf it is a series A, it puts the orders of magnitude next to the figures
    1. For the model and the stage, public orders of magnitude, each with its source and date: what a net retention, a CAC payback or a burn multiple is worth at series A. They show next to the deck figure and never enter the score. When no dated source exists, the cell stays empty.

  6. Finally, for every deck
    1. The score is the share of questions the deck answers, each theme weighing according to the stage: before customers, team and field count the most; after, traction and economics; at series A, unit economics and net retention. The report is written by an AI that sees only the answers, never the deck. It cannot be charmed by a good page 3.

The weights and the computation, stage by stage

The computation fits in three lines, and code does it. Per question: found is worth 2 points, partial 1, absent 0, multiplied by the weight of the theme. Per theme: points obtained over the maximum possible, as a percentage. Global: the average of the themes, weighted by their weight. A theme under 50% turns red and produces a question for the call, not a judgement.

What changes from one stage to the next is what weighs. The weights are public and dated, and they will move at the first reviews.

24 questions

Pre-seed

What did the founder learn in the field?

Pre-seedWeightShare of score
AProblem and field321%
BEvidence321%
CEconomics17%
DMarket and timing214%
ETeam321%
FMoney and next step214%
26 questions

Seed

Who pays, and who comes back?

SeedWeightShare of score
AProblem and customer213%
BTraction320%
CEconomics320%
DMarket and competition213%
ETeam213%
FMoney and next step213%
GReferences17%
32 questions

Series A

Does the machine repeat?

Series AWeightShare of score
AProblem and customer212%
BTraction212%
CUnit economics318%
DNet retention318%
EMarket and competition212%
FTeam212%
GMoney and next step212%
HReferences16%

The business-model block adjusts these weights: a biotech puts traction at 0, a hardware company doubles economics.

What you get

A report, as PDF and as markdown, next to the deck on your disk. Inside: one bar per theme. Question by question, the answer, the page, the quote. The three main gaps, ranked by weight. Then the questions for the call, worded as such: "Are Lena and Marc full-time on Gantrix, and since when?"

If the company has customers, add every deck figure with what confirms or contradicts it, the gaps to probe in the call, and the ready-to-send email asking for the missing documents. You send it, or you do not. At series A, every figure has its benchmark next to it, source and date included.

What it does not do

It does not say whether to invest. The score measures whether the deck answers the questions, not whether the company is good. A complete deck can describe a bad company, and a founder who presents poorly is not a founder who did nothing. Only a conversation tells those apart.

It makes nothing up. The thresholds, the score computation and the quote checks are done by code, not by an AI. It only blocks a reading on a proven contradiction, never on something vague. It never sends the email. Nothing leaves your disk except the web searches, which carry the text of the statement and never the deck.

It has limits, and they are written down. Text that only exists inside an image or a chart cannot be cited: it counts as absent. A customer with no public trace is unverifiable, not contradicted. The thresholds are a first setting. They are public, and they will move.

The questions are dated and public, and every reading is kept. At eighteen months, we look at what the companies became and publish the result, flattering or not.

See the code on GitHubRead the pre-seed, seed and series A questions

THIS KIND OF READER, ON YOUR DEAL FLOW

Your questions are not mine? The mechanism is the same, we put yours in. Want to see what it would give on your decks?

Book a chat
Repository activity · deck-reader6 commits
Repository created on 8 Sept 2026 · read daily from GitHubSee the commits