How it works
What the tool does, in the order it does it. Written for someone who will use it or decide about it, not someone who will maintain the code.
In one paragraph
An associate gives the tool what they have on a company: the deck, the website, the founders' bios, their own notes from the first call. The tool breaks those into numbered passages and asks the AI to write an assessment where every sentence points at one passage and quotes it. A separate program then checks every quote against the source. Anything that fails gets one chance to be fixed; whatever still fails is shown as unverified rather than hidden. The associate edits the result in a document that opens the real slide behind any sentence. The tool never recommends and never scores.
Getting a company in
The aim is that nobody uploads anything. Decks arrive by email; the tool reads them as they land; by the time an associate opens the app the company exists with its deck read, its name, round and founders filled in, and, if the fund wants it, a draft already waiting.
The inbox is where that happens. It can be fed three ways:
- A connected mailbox. Gmail or Microsoft 365, read-only, watching one label or folder the fund chooses. Messages in that label with a deck attached or a deck link are picked up; nothing else is read, and nothing is ever sent, moved or deleted.
- A forwarding address. The fund adds it as a forwarding rule, or gives it to the people who receive decks. No mailbox access needed.
- A drop folder. In this prototype a folder on disk stands in for a mailbox, holding sample emails, so the whole flow can be tried end to end today.
Two switches on the Inbox page decide how far it goes on its own: import on arrival turns each message into a company with its attachments read and its fields filled; draft on arrival drafts it too.
The connectors are set up per client rather than shipped as a universal button. Google and Microsoft each require the fund's own credentials for a read-only connection, and every fund routes deal flow differently: a shared deals@ address, a label, a forwarding rule, sometimes a Slack channel. The connection flow in the app explains what gets connected and what the fund has to provide, and keeps the settings so the connection goes live when the credentials exist.
When an email gives a full link to the company's site, it is read too. A website address that only appears inside a deck is shown as a suggestion with an "Add as a source" button rather than read automatically, because the address in a deck can belong to someone else: one test deck named a domain that turned out to be an unrelated leather-goods shop.
Two fallbacks remain for anything that did not come by email: drop in files (PDF, PowerPoint, Word, or a link to one) or paste a website link. Once a deck is read, by any route, the company's name, one-line description, round, website and founder bios are filled in from it, and the associate corrects whatever is wrong.
Drafting
Each source is split at paragraph breaks into passages with short labels, such as deck:3 for the third slide or call-notes:7 for the seventh paragraph of the notes. Slide titles stay with their slide. The AI only ever sees text with these labels attached, and it is told it may point at nothing else.
It returns the assessment in a fixed shape every time: a summary a partner can read in two minutes, six to eight sections such as team, market and traction, where the sources disagree, what is missing, the strongest case against, and a list of anything in the inputs that tried to steer the reader. That shape has no slot for a score or a verdict, so neither can appear even by accident.
Every sentence carries the label of the passage it came from and a short quote from it, thirty words at most. Each sentence also says whether the source states it, whether the tool worked it out from numbers in the source, or whether it is a point about something the source does not contain.
Drafting runs in the background on the server. The page shows progress and can be closed or refreshed; coming back picks the progress up again, and the company list shows it as Drafting. If the connection to the AI drops partway through a reply, the step is retried once. If the server restarts mid-draft, the company simply shows as not drafted, ready to start again.
Checking the quotes
A small program, with no AI in it, reads every quote and asks: does this passage exist, do these exact words appear in it, is the quote short enough, and does every number in the sentence appear somewhere in the passages it points to.
If anything fails, the AI is shown the exact failures and asked once to fix the quote, point at the right passage, or drop the sentence. There is no second chance. Whatever still fails stays in the draft, marked, so the associate can see what the tool wanted to say and could not back up.
"Verified" therefore means one specific thing: these words really are in that source. Whether the sentence built on them is fair is a separate question, answered by the second reader described in the Evaluation.
The document
A company has one assessment, shown in whatever state it is in: not drafted, drafting, draft, or reviewed once the associate has changed anything.
The editor behaves like a Notion page. Click anywhere and type. Enter makes a new block, a slash offers a heading, bullet, quote or divider, and blocks drag by their handle. Sentences can be reworded or removed, but their quote and source stay attached. Clicking a sentence, or the small source label after it, opens the actual deck page in a panel beside the text with the quoted words highlighted. The panel resizes.
The AI's draft is never overwritten. The associate's edits sit on top of it, so the original and the edited version can always be compared.
Under the header sits a strip with the fit tag and the topic tags. Right after a draft is written, a second, smaller AI call reads the draft and suggests one fit tag (good fit, possible fit, not a fit) and up to four topic tags from a fixed list. Each reason has to name the statements in the draft it rests on, and a program drops any reason or tag that names none, or a "Warning signs" tag that does not point at a note about steering. "Why" opens the reasons with links to those statements. The associate can change the fit, add or remove tags, or go back to the AI's suggestion; their choice is saved next to the AI's original, never over it. The fit and tags also show on the company cards, where the list can be filtered by fit.
Send to partner opens an email ready to go: the fit, the summary and a link to the Preview of the assessment, plus a Word copy to attach. It opens in the associate's own email app, and the assessment is marked as sent, with who it went to and when.
A switch flips between Edit and Preview; Preview is what the partners receive, with the fit and tags at the top. Export gives Word, PDF, Markdown or a clean copy, all with the edits applied. A follow-up email to the founder is composed from the missing list and the open questions. Search (Cmd+K) finds anything across every company: a phrase in a slide, a sentence in a draft, a missing item, an inbox message.
Cost and time
| Step | Per company |
|---|---|
| Drafting | 2 to 3 minutes, about 30 to 50 cents |
| Fixing failed quotes, when needed | about a minute more |
| The second reader, in the evaluation | about a minute, under 10 cents |
| Filling the intake form from a deck | a few seconds, a few cents |
At 150 companies a month, the AI spend is under $100.
What a real version would add
The full plan, in order, with the questions to settle with a fund, is in the Next steps document. In short:
- The first mailbox connector, built for the first client's mail provider and routing, so the inbox is fed without anyone touching it. Then slide numbers that read as
deck p.4, better website capture, and call recordings turned into notes. - A database in place of folders, holding companies, sources, drafts, edits and the partners' decisions.
- The second reader on every draft, with its verdict shown next to each sentence alongside the quote check.
- Sign-in, and partner pages that can be sent as links.
- A record of outcomes from day one, so that after a year the tool can be judged on whether it helped the fund decide, not only on whether its quotes were real.
Limits worth knowing
- The five example companies were written by the person who built the tool. They test the mistakes that person thought of.
- Each example was run once; a second run would be worded differently.
- The second reader has not yet been checked by a person.
- Very long documents make passages too big for a quote to pin precisely.
- The number check is literal: it misses numbers written as words and can be fooled by a coincidence.