Read every deck with the slide behind each fact
Every fact pulled from a deck links to its slide. Slides that couldn't be read are flagged, not guessed, and founders can correct what was extracted.
- PDF analysis and citations Pilot
- Cited research First release
Traction
20 enterprise customers
USD 410k ARR · Sep 2026
Slide 4: “20 enterprise customers”
7 extracted facts · scroll or use arrow keys
What happens after a deck arrives
Each step is a queued job that retries on its own and shows when it fails, so a broken extraction never blocks your team from opening the deal.
- 1
Upload
The founder uploads a PDF on your portal, or an analyst adds one to a deal. The original file is kept as it was sent.
- 2
Scan
The file is scanned for malware before anything reads it. A file that fails stays in quarantine, and the founder is told it couldn't be accepted.
- 3
Extract
Text is read slide by slide, with character recognition for slides that are images. Each slide is also saved as a picture for the viewer.
- 4
Map the deck
A first pass labels each slide (problem, market, traction, team, ask and so on) and notes the stage and sector the deck describes. It doesn't score anything.
- 5
Check the numbers
Currency amounts, percentages and metrics such as ARR, MRR, burn and runway are pulled out in code and checked against each other before any model writes a summary.
- 6
Facts and claims
Company facts and claims are extracted, each with its slide number and the exact words from that slide. A quote that can't be found on its slide is dropped.
The design target is a finished analysis within five minutes for most PDF decks. It is a target for the pilot, not a measured result. If a step fails, the deal opens anyway and the analyst reviews the deck by hand.
Every fact points at the slide it came from
Click a fact in the panel above: the slide opens with the quoted region outlined. The same quote is repeated in text, so the outline is never the only signal.
| What | Examples | Cited to |
|---|---|---|
| Company facts | Name, website domain, headquarters, stage, sector, round size and currency, team size | The slide each value came from |
| Metrics | Revenue figures with their period and units, customers, pricing, growth | Slide and the exact quote; numbers cross-checked in code |
| Claims | Statements worth checking, such as "20 enterprise customers" | Slide and quote; labelled Founder reported until checked |
| Structure | What each slide covers, and which expected sections are missing | Slide map |
| Gaps | Slides with too little readable text, charts that couldn't be read, character recognition used | Flagged per slide for an analyst |
A claim, as your team sees it
“USD 410k ARR · Sep 2026”, exact words from slide 4.
- Quotes must match the slide's extracted text, or the claim is dropped.
- Numbers are read in code first, so a summary can't invent a figure the deck doesn't contain.
- Currency is kept as written (USD, EUR, INR), with the period and units.
A slide that couldn't be read is flagged, not guessed
Pitch decks are full of screenshots, charts and logos. When the text isn't there, Delsivo says so.
Image-only slides
Character recognition runs when a slide has no text layer, and the slide is marked so the analyst knows the text was recognised, not read.
Illegible charts
A chart whose labels can't be read becomes a gap such as "Slide 7 chart not legible", never a made-up number.
Missing sections
If the deck has no team, traction or ask slide, the slide map shows it, and the fit report keeps those criteria Unknown.
Founders can correct facts, and old citations keep working
A deck changes during a process. The record keeps every version and every source, so nothing your team read earlier disappears.
- Pilot
Founder corrections with provenance
Founders see the key facts read from their deck and can correct them on your portal. The correction is stored beside the original, marked as a founder correction, and your team sees both.
- Pilot
Deck versions
A replacement deck becomes version 2 on the same deal. Citations to version 1 still open version 1 at the right slide, and you can see what changed between versions.
- Pilot
Duplicate suggestions
When a company applies twice, or a referral and an application describe the same startup, Delsivo suggests a match by company domain and founder identity. An analyst confirms any merge, and the history of both is kept.
- First release
PowerPoint after measurement
The extraction service can read PowerPoint, but it is switched on only after quality has been measured on real decks. Until then, founders are asked for a PDF.
What is sent to the AI provider, and what is not
Deck analysis uses a large language model from a named provider, listed on the subprocessors page. Each call gets only what the task needs.
| What | How | |
|---|---|---|
| Sent to the AI provider | Slide text and slide images needed for the task, the slide map, and your rubric when screening | Contact details removed where possible. Deck text is marked as untrusted content, so instructions hidden in a deck are not followed |
| Not sent | Other deals, other funds' data, your team's notes and ratings, founder account details | Each call gets only the evidence for that task |
| Not used for training | Nothing a fund or founder submits | AI providers are used under terms that don't train on submitted content, confirmed per provider before customer content is processed |
| Logged | Which model ran, prompt and rubric version, timing and cost | AI logs keep no document text |
Files are scanned before any extraction or AI call.
A re-uploaded identical deck reuses your fund's earlier analysis. Another fund's cache is never used.
Every AI result is labelled as a suggestion until an analyst accepts it.
A well-made deck and a good fit are different questions
The analyzer reads how complete and clear the deck is. Whether the company fits your thesis is a separate reading from your own rubric.
| Deck analyzer | Thesis screening | |
|---|---|---|
| Answers | What does the deck say, and where? | Does this company fit our rubric? |
| Based on | The deck's slides | Extracted facts, your published rubric version |
| Output | Facts, claims, gaps, a deck quality reading | Met, Mismatch or Unknown per criterion, three separate readings |
| Who decides | An analyst accepts or edits facts | An analyst or partner, with a reason for any override |
More on rubrics and the fit report on thesis screening.
Examples use a made-up company.
Questions about deck analysis
Read your past decks with their citations
Pilots start with deals you have already reviewed, so you can check every extracted fact against slides you know.