AI deal flow scoring for VC: how a pre-seed fund finds founders early

Part 01

The brief: a short weekly list of strong leads

The fund invests at pre-seed and wanted to find exceptional founders early: people coming out of strong companies and universities who are starting something new, before the round is public and before other funds are in the conversation.

Two constraints shaped the whole build. The first was precision. The partners wanted a short list of strong leads each week, not a feed of hundreds of names. A tool that flags 300 people a week moves the sorting from one spreadsheet to another and saves nobody any time. Every lead on the list had to be worth a partner's attention.

The second was timing. At pre-seed, a lead found after the round is announced has already been found by someone else. That ruled out the obvious design, which is to watch funding news and startup databases, and pushed the system toward signals that appear months earlier.

Part 02

Why the press is the last signal on the ladder

Most deal-flow tools watch funding announcements and startup news. Those are good records of what has already happened. For a pre-seed fund they arrive too late: by the time a round is in the press, it has closed.

Earlier signals exist, and they tend to arrive in order. We call it the signal ladder. First a person's role changes to founder, or to "stealth". Then they start building something new. Then a company appears in the official company registry, and it posts its first founding roles. The round and the press come last.

Each rung is public. Each one usually comes before the next. A system that reads the first four gets to a founder before most funds have a reason to look.

The gap can be long. While building the system for the fund, we found a company that was in the official company registry about 21 months before any press covered it. That is one case, not a benchmark. It still shows how much time sits between the early rungs and the last one, and that time is where a pre-seed fund wants to be.

The early rungs are also noisy. A registry lists every new company, not only the ones a VC cares about. In one European country alone, the national statistics office counted 113,289 VAT registrations in 2024, and 37,473 of them were private limited companies (BV/SRL). Nobody reads that by hand every week. That is why the scoring step exists.

Part 03

AI deal flow scoring for VC: the 100-point rubric

Every person who shows up in the sources goes to an AI judge. The judge does not decide whether the fund should invest. It answers three narrow questions, gives points for each and writes down why.

A person needs 65 of 100 to be flagged. The second axis works as a gate. Take someone with a perfect background and a perfect thesis fit but no sign of starting anything: they score at most 60, below the line, so they are never flagged. An impressive person who is happy in their job is not a pre-seed lead, and the arithmetic makes sure the system agrees.

Below the line, nothing disappears. Every person under 65 stays in the dashboard as "reviewed", with the reason the judge gave. A partner can open that list at any time and see what the system filtered out and why. If the filter is wrong, the mistake is visible there, with its reasoning, instead of hiding in a black box.

Before the system ran for real, we calibrated the rubric on cases the fund already knew. In that calibration it flagged, on its own, a startup the fund had already backed. Nobody had told it about the investment. That is not proof of accuracy, and we do not present it as one. It does show the rubric pointing where the partners had already pointed.

The partners see what the system flagged, and also what it filtered out and why.

Alpgency

Part 04

The pipeline, from public sources to a partner's desk

The scoring is one step in a short chain. Each step has one job:

1. CollectSignals are gathered from public professional data and official company registries.
2. ScoreThe AI judge scores every person on the three axes and writes its reasons next to the points.
3. DeduplicateThe same founder can appear in several sources. Duplicates are removed, so a partner never reads the same person twice.
4. DeliverFlagged leads land in a dashboard and in a weekly digest for the partners.
5. ReviewA partner reads each lead and decides what to do. The system stops at the flag.

The system was live in production 9 days after kickoff, in 2026. The scope was narrow on purpose: one fund, one rubric, one weekly list. The partners did not get a new platform to learn. They got a dashboard and a digest that answer one question: who should we talk to this week?

Part 05

Guardrails: the system flags, a person decides

A system that reads public data about people falls under the GDPR. We treated that as a design input from the first day. Four rules hold for the fund's system:

• A person reviews every lead. The system ranks and explains. It does not decide.
• Nothing goes out automatically. No emails, no connection requests, no messages. Any outreach is a partner's call.
• It uses public professional data. It does not scrape LinkedIn.
• It rests on a legitimate-interest assessment, and anyone can opt out.

The reason scraping is out has a date and an amount. On 5 December 2024 the CNIL, France's data protection authority, fined Kaspr €240,000. Kaspr sold a browser extension that pulled the contact details of LinkedIn users, including people who had restricted the visibility of that information to their connections. Its database held about 160 million contacts. A fund should not build its deal flow on data collected that way.

Part 06

What other funds can take from this

The rubric belongs to the fund. The method travels: we build these systems for venture capital funds at any stage, in any European market, and set the rubric, the signals and the threshold per fund, for its own thesis and stage. If you run a fund and want to build something similar, these are the decisions that mattered most:

1. Write down your own ladderList the signals that show up before the round in your segment, in the order they appear. A seed fund in Munich and a pre-seed fund in Lisbon will not watch the same things.
2. Make one axis a gatePick the signal you cannot do without and give it enough weight that no other score can carry a lead over the line without it.
3. Keep the rejects visibleStore the reason for every score, including the low ones. Without them, nobody can check what the filter removed.
4. Calibrate on deals you knowRun the rubric over companies you already know before it runs for real. If it misses them, change the rubric before anyone relies on it.
5. Stop at the flagLet the system rank and explain. Leave every contact with a person.

The same shape works on the other side of the fund. For Decelera, a venture capital firm, we built an agent that finds potential LPs for the fund and scores each one against the fund's criteria. The signals differ and so does the rubric. The chain is the same: public sources in, a scored short list out, and a person deciding what to do with it.

FAQ

Frequently asked questions