found.

Every Yale startup and every opportunity to fund it, verified, in one place.

Madison Andoniadis · MGT 409 final project · gpt-6-luna

The problem

  • Yale's startup scene is spread over separate sites: 122 startups on the Pitchbook directory, 130 opportunities on the Tsai CITY board.
  • One filter at a time, 5 pages of listings, founders hidden.
  • Listings are self-reported: 16 of 97 are misleading.
  • Founders want local support before seed.
  • As co-lead of the Startup & Entrepreneurship Club: a known gap that needs a fix.
Directory

What's new

1place for both Yale sources, multi-filter searchvs Pitchbook + Tsai CITY separately
✓verified people: Yale email, proven founder claims, and a way in for people who want to join a startupvs one-way directories
AIthat can't invent, and an Auditor that fixes the datavs ChatGPT guessing

Compared: Yale Pitchbook, Tsai CITY board, Harmonic, Crunchbase, LinkedIn, ChatGPT

Two small agents, lots of plain code

Member / visitorpublic site (React) Site owner (admin)admin console (React) FastAPI backend auth · Yale email · claims token budgets · audit trail admin approval queue Agent 1 · Scout gpt-6-luna, 1 call, no tools question → search filters Agent 2 · Listing Auditor gpt-6-luna, tool: read_page (≤2) → verdict, fix, quoted evidence quotes checked in code SQL searchcode writes thereply + cards SQLite122 startups (Pitchbook)130 opportunities (Tsai)corrections.json Startup's websiteown domain only,visible text, fenced filtersauditadmin approves → fix saved

Scout: ask in plain English

Scout
  • One luna call turns the question into filters.
  • Code searches and writes the sentence, so nothing is invented.
  • 587 tokens, 1.75 s, $0.00007 a question.
  • 19/20 test questions fully correct.

The Listing Auditor fixes the directory

Auditor
  • Reads the startup's own website (visible text only).
  • Suggests a fix with quotes verified in code.
  • A person clicks Apply; the fix survives data refreshes.
  • 16 misleading listings found in 5 min for $0.03.

Cost

Per runTokens (in / cached / out)Cost
Scout question557 / 0 / 30$0.000071
Listing audit2,252 / 11 / 146$0.000297
$0.18a month at 100 members
$15a month at 10,000 members
−59%tokens from the 1-call Scout, with no accuracy loss19/20 vs 17/20 correct

Safety: 16/16 attacks stopped

  • Hidden and visible prompt injection on websites: stripped, flagged, ignored.
  • Invented evidence removed in code; fixes need a person.
  • Budgets per run, per visitor, per site.
  • Yale-only accounts, proven founders, separate admin console, no SSRF.
Budget

Impact

252Yale startups and opportunities in one search
16misleading listings caught~5 h by hand → 5 min
120s → <20stimed pilot: the same task on the two Yale sites vs found. (6× faster)

From project to real website

  • Host on Render or Railway with a persistent disk and backups (data lives in one SQLite file today).
  • Real email from a dedicated sender (Yale shared mailbox or Resend), plus a weekly "new events" digest.
  • Real people: pilot with the SOM Startup & Entrepreneurship Club, then Tsai CITY and Yale Pitchbook to get founders claiming listings.
  • Cost: about $8/month for a club pilot, about $67/month Yale-wide (estimates).

https://github.com/mandoniadis1/found