Labscout

Inspiration

The best deep-tech founders are often students two years before there is a company. By the time a startup shows up in PitchBook, Harmonic or Crunchbase, every fund can see it. Early signals already exist in public: a first-author paper, a repo picking up stars, an NSF I-Corps award, a new NY corporation filing. Nobody joins them up per person. So we built ALL SIGNAL

What it does

A VC types an investment thesis in plain English, or uploads it as a PDF. AllSignal returns current university students already publishing on it with proprietary ranking. Each student is an index card in a drawer:

A spinout score with recent momentum
A 4-stage meter: research, builder, industry pull, spinout intent
A 3-line "why meet them" brief
Thesis fit: yes / partial / no against each criterion in the fund's own memo
Evidence links behind all numbers

One click exports a partner memo. Shortlist is remembered, so re-running a thesis next week shows only who is new.

How we built it

Thesis to search: Claude (via Nicrron) turns the thesis into keywords and pulls criteria from an uploaded memo.
Find people: OpenAlex papers from 2024–26 at four NYC schools, aggregated per author. ORCID checks enrollment: an education entry with no end date means a current student.
Founder signals: 24 public sources, mapped to stages from lab to company. They include GitHub, Hugging Face, arXiv, Semantic Scholar, NSF, NIH, NY DOS, SEC Form D and YC.
Market pull: akta.pro news for the thesis.
Saved shortlists: Roundtable versions every shortlist.
Tracing: PRISM traces every run.
Stack: zero npm dependencies, Node native TypeScript, Vercel serverless, static HTML.

The displayed score is normalized so the top student in each result is 100.

Challenges we ran into

Entity resolution. "J. Chen" at Columbia is many people. A signal only counts with a name match plus a shared institution, paper, ORCID or handle. We dropped many matches rather than show a wrong one.
Pedigree bias. Citations and lab prestige reward people who are already known. We weighted velocity over totals and penalized long careers, so a second-year student with momentum can outrank a professor.
Signals arrive late. Company filings and customer logos mostly appear after incorporation. We use them to confirm a rank, not to drive it.
Flaky upstreams. PatentsView and SBIR.gov were down on build day. Fonts are self-hosted, and a failed live search falls back to a labeled snapshot, because venue Wi-Fi is unreliable.

Accomplishments that we're proud of

Explainability earns trust. In interviews, a VC trusted a ranked list only when each point traced to a source.
The software finds the people; partners judge the person. Speed of learning, resilience and the ability to recruit still need a coffee.
The product is the re-run. A weekly "who's new for my thesis" diff is what a fund would pay for, not a one-off search.

What we learned

A score needs receipts. A ranked list is only useful if every point traces back to a source, so every number in Labscout opens to its evidence.
The software finds the people; partners judge the person. Speed of learning, resilience and the ability to recruit still need a coffee.
The product is the re-run. A weekly "who's new for my thesis" diff is what a fund would pay for, not a one-off search.

What's next for Labscout

1. Talk to the buyer first. We built before interviewing, and we know that is backwards. Next, we will talk to 15–20 associates, partners and scouts at pre-seed and seed funds. We want to learn how they source student founders today, what a missed founder costs them, and what they would pay. Today's judges' feedback is our first round of that input.

2. Rework the product around what we hear. Every version gets tested against real theses from real funds. We will change the scoring weights, the card, and the memo based on what partners actually forward and act on, not on what we assumed.

3. Full build-out:

Coverage: more universities beyond Columbia, NYU and Cornell, starting with Boston and the Bay Area.
Signals: bring patents (USPTO) and SBIR/STTR back online, and add accelerator and grant-program rosters.
Accuracy: stronger entity resolution, so one name never merges two people and every match shows its evidence.
Calibration: check the spinout score against students who later founded companies, so the weights are earned, not guessed.
Workflow: a weekly "who's new for your thesis" digest, CRM export, and shared shortlists for a whole fund.

4. Prove someone will pay. We will run 3–5 funds on Labscout for a month. Then we will measure how many intros they take and how many founders they would not have found otherwise. That number decides the price.