Case study · Venture capital · Screening · AI

This week's best pitch deck, read first, with every score tied to its slide.

You're a junior investor at a fund that backs young startups. Founders email you about 50 decks a week (assumed), and you read them in the order they arrive. Deck Rank ranks them on upload so you call the best fit first. Every score quotes the slide it came from.

Try the demo Code Three fictional companies. No real founder decks yet.
11 to 13 sfrom upload to extracted values, 3 runs on the same deck
49values from a 12-slide deck, all from fixed rules
9 of 12target numbers covered; the AI returned none for the other 3 rather than guess
66unit tests, one fails if any quote isn't on its slide
For
A fund that backs young startups
Users
You, the junior investor, and the head of the fund, who decides where your time goes
Data
Three fictional companies' decks
Built with
Python, Streamlit, Postgres, an AI language model
TechniquesData engPDF chunked per slideData engRule-first extraction: patterns, tables, layoutsData engAppend-only values with provenanceAILLM fallback for missing numbers onlyAIExact-quote groundingData scienceWeighted score with a must-have capEval66 unit tests, quote-on-slide check
01 · VISION

Your fund is first to close the best deal that comes in.

Faster funds reach the founder first. The vision: every deck ranked the moment it arrives, by one shared rule, with every number one click from its slide, so your fund is first to be sure, and first to close.

02 · THE PROBLEM

Four things go wrong today.

1 · The best deck waits

Decks are read in the order they arrive, each in full. The best one can sit behind 49 others.

2 · Two investors pick a different "best"

Each junior investor ranks in their head.

3 · "Why this one?" means reopening the deck

When the investment committee asks, the evidence isn't at hand.

4 · The first call starts with basic questions

Decks leave out must-have numbers, and nobody notices until the call.

03 · THE SOLUTION

Rules read first. The AI fills gaps. You call the founder.

  1. Fixed rules pull out the numbers first, slide by slide, using text patterns, tables and layouts.
  2. An AI language model fills only the numbers still missing. Its quote must match the slide text exactly, or the value is rejected.
  3. One weak must-have pulls the score down. Weighted average times the lowest must-have score divided by 5.
  4. You get a ranked list and a scorecard per company. Missing must-haves are flagged, with a drafted founder email for you to review.
04 · HYPOTHESES AND IMPACT

One job, one bet, one number each.

Your jobHypothesisMetric and target, then resultExpected impact
When decks arrive, I want them ranked on upload, with each number tied to its slide, so I call the best fit first.If decks are read and ranked on upload, the best-fit deck gets a first-call decision sooner, because you start at the top of the list.Time from deck received to first-call decision, top 5 decks: down. Baseline unobserved. Upload to values: 11 to 13 sFirst calls booked sooner (unobserved)
When my colleague and I rank the same batch, I want one shared rule, so we agree on the best.If every junior investor starts from the same fund weights, two investors' top 10 lists overlap more, because the rule is written down.Top-10 overlap between two investors on the same batch: up. Baseline unobservedFewer disagreements on who to call
When the committee asks "why this one?", I want the formula and source quote, so I answer in one click.If every score shows its formula and source quote, you answer faster, because the evidence is one click away.Time to answer a "why" question: down. Baseline unobservedEach yes or no explainable to the fund's investors
When a deck misses a must-have number, I want it flagged with a drafted request, so the first call skips basic questions.If missing must-haves are flagged at upload with a drafted request, fewer first calls start with data missing, because the request goes out the day the deck arrives.Share of first calls with a must-have still missing: down. Baseline unobservedShorter path to a decision (unobserved)

Must not get worse: the number of top-10 companies later turned down stays flat or goes down. No junior investor has tested it.

05 · THE VALUE

What it's worth, and what it asks of you.

You open the laptop and this week's best deck is #1. The first call is booked the same day.

You, the junior investor
You get
This week's best deck is #1. The first call is booked the same day. You explain the rank to the investment committee in 1 click.
Will you believe it?
Medium. Every number has a quote to check. Not yet tested with junior investors.
How fast
Low delay. Numbers are pulled from a deck in 11 to 13 s (observed). Time to decide on a first call is unmeasured.
What it takes
Medium. No copy-paste, but decks are uploaded one by one.
What you give up
Low. You keep your own judgment: you can lock in your own value and add personal weights on top. You give up one more tool outside the inbox.
The head of your fund
They get
The fund is first to be sure about the best decks that come in. It gets into contested funding rounds early. Each junior investor handles more decks, with no new hire. Every yes or no is explainable to the fund's own investors.
Will they believe it?
Low to medium. Depends on junior investors using it. No data on deals won.
How fast
High delay. The effect on deals won shows over quarters, not days.
What it takes
Medium. Setup, plus the team agrees on fund weights once.
What they give up
Medium. Founder decks can go to a cloud database and an outside AI model. Junior investors give up some freedom to rank their own way.

The weak link is belief, for both. For you: a 20-deck test (hypotheses 1 and 2). For the head of the fund: a lost-deal audit. For decks that other funds later backed, how long did they wait for a first call?

06 · LIMITATIONS

What this doesn't prove yet.

07 · SCOPE AND NEXT STEPS

What it won't do, and what comes next.

Non-goals
  • Finding companies that haven't sent a deck. That's Alpha Scout
  • Memory (sub-problem 5): why a company was turned down. The status log keeps date, who and score, not the reason
  • Sending email automatically
  • Deciding to invest or pass. You decide
  • Valuing the company, syncing with the fund's contact database (CRM)
Next
  1. Accuracy first: compare extracted numbers with decks checked by hand.
  2. A 20-deck test with two junior investors: arrival order vs ranked list, and top-10 overlap.
  3. A practice committee with 5 "why" questions, deck only vs scorecard.
  4. A lost-deal audit for the head of the fund.
  5. Own formulas for Founders, Market and Tech Moat.
RESOURCES

Go deeper.

Try the demo, step by step three steps, in order
  1. Upload your deck (PDF), or use the 3 made-up samples.
  2. See where it ranks in your list, read from the top.
  3. Open it and check each number against the slide it came from, and see what's missing.

Files stay in your browser session only and are never saved. AI is off unless you turn it on.

Main learnings sample deck, speed, worked example, tests
  • Sample deck: 12 slides produced 49 values, all from the fixed rules. They covered 9 of 12 target numbers. The AI model was asked for the other 3 and returned none rather than guess.
  • Speed: upload to extracted values took 11 to 13 seconds, over 3 runs on the same deck. This supports "ranked on upload". It says nothing yet about time to a first-call decision.
  • Worked example: a Financials score of 3.16 out of 5, built from four inputs: 71% gross margin (revenue left after direct costs), 24 months of runway, revenue compared with burn, and funding stage.
  • Tests: 66 unit tests (small automated code checks). One fails if any stored quote doesn't appear in its source slide.
Hypotheses in full before and after, and the test for each

1 · SPEED

Before -> after
read in arrival order, in full -> ranked on upload, best first
Hypothesis
If decks are read and ranked on upload, then the best-fit deck gets a first-call decision sooner, because the junior investor starts at the top of the list instead of reading in arrival order.
Metric
Time from deck received to first-call decision, top 5 decks. Down. Baseline unobserved. Must not get worse: the number of top-10 companies later turned down stays flat or goes down.
Test
Two junior investors, same 20 decks, arrival order vs ranked list. "Top 5" = the 5 decks they'd call first after reading all 20.

2 · CONSISTENCY

Before -> after
ranking in each junior investor's head -> one shared formula and fund weights, personal view kept separate
Hypothesis
If every junior investor starts from the same fund weights, then two investors' top 10 lists overlap more, because the ranking rule is written down instead of remembered.
Metric
Top-10 overlap between two junior investors on the same batch. Up. Baseline unobserved.
Test
Same 20 decks. Compare top-10 overlap when ranked by hand vs with the fund weights.

3 · EVIDENCE

Before -> after
reopen the deck -> formula, inputs and source quote per score
Hypothesis
If every score shows its formula and source quote, then the junior investor answers the committee's "why this one?" faster, because the evidence is one click away.
Metric
Time to answer a "why" question on a ranked company. Down. Baseline unobserved.
Test
Practice investment committee with 5 "why" questions, deck only vs scorecard.

4 · GAPS

Before -> after
missing must-have found on the first call -> flagged at upload, moved to the bottom, founder email drafted for review. Data older than 180 days flagged
Hypothesis
If missing must-haves are flagged at upload with a drafted request, then fewer first calls start with data still missing, because the request goes out the day the deck arrives.
Metric
Share of first calls with a must-have still missing. Down. Baseline unobserved.
Test
Count it across the same 20-deck batch.
How it's built slides, fixed rules, AI fill-in, score
  1. The deck PDF is split into one chunk per slide.
  2. Fixed rules pull out the numbers first, using text patterns, tables and slide layouts.
  3. An AI language model fills only the numbers still missing. Its quote must match the slide text exactly, or the value is rejected.
  4. Each value is stored with its source slide, the version of the code that read it, and the date.
  5. The score is a weighted average multiplied by the lowest must-have score divided by 5. So one weak must-have pulls the whole score down.
  6. The junior investor gets a ranked list, a scorecard (score breakdown with sources) per company, and next steps for missing data.
Scope what this build covers
  • PDF decks that founders send in
  • Numbers pulled out with the quote they came from
  • Weighted score, must-have cap, fund weights plus personal weights
  • Ranked list, scorecard, junior investor can override a value with a reason
  • Missing-data flag plus a drafted founder email for the junior investor to review
  • Data older than 180 days flagged, quote re-checked on refresh
Trade-offs what was chosen, over what, and the cost
ChoseOverGainCost
fixed rules first, AI model secondAI model for everythingsame result every run, cheap, easy to checkmisses numbers the rules don't cover
reject value if quote doesn't matchaccept whatever the AI model returnsno numbers without a sourcefewer values filled
must-have cap on scoreplain weighted averagea weak must-have can't hidepunishes short decks from very early startups. may need tuning
keep every old value, add new onesoverwrite on re-runfull history, safe re-runsmore stored rows, needs code to pick the current value
email drafted, not sentauto-sendjunior investor stays in controlone manual step
Terms on this page plain definitions
  • Pitch deck: slides a founder sends to ask for money.
  • First call: the first meeting between the fund and a founder.
  • Investment committee: the group at the fund that approves each investment.
  • The fund's own investors: the people and institutions whose money the fund invests.
  • Burn: cash a company loses each month.
  • Runway: months left before the company runs out of cash.
  • Must-have: a number the fund needs to see before a first call.
  • Fund weights: how much each score category counts, agreed once by the team.
Related project and links Alpha Scout, code 
SHARED GOAL: see the best startup in the fund's focus area first.
             be the first fund ready to invest.

Alpha Scout = SEE FIRST.  company hasn't sent a deck -> find it, score it
Deck Rank   = ACT FIRST.  deck arrives -> fund is sure -> first call

Alpha Scout case study

How it was built with PM OS the product process, step by step
StageArtifactRepeatable with
OpportunityBusiness proposal: problem, product, personas (README section 1)/opportunity-analysis
PRDUser stories one-pager: advantage and business outcome per persona/prd-draft
MetricsNorth star: invest in more great deals, faster, backed by data that holds up to questions/define-north-star
Build7-phase roadmap, extraction pipeline, 66 product tests + 7 demo tests/code-first-draft
Launch5-minute demo script and walkthrough deck/prototype
ShowcaseThis page, name audit, goals and state files, public demo/portfolio-case-study, /name-audit, /goal-gate

Skills are mapped from the artifact each stage produced. They show how each stage repeats with PM OS; the original working sessions were not logged.

PM OS skills used

/code-first-draft /define-north-star /goal-gate /name-audit /opportunity-analysis /portfolio-case-study /prd-draft /prototype