Case study · shipped internal product

AGI House — the attendee research engine

A static ops console that scored and briefed real AGI House hackathon attendees — people explorer, Firecrawl-sponsored event report, and founder-impact timelines — lifted from the repository and running here on anonymised fixtures.

The problem

AGI House runs dense founder/engineer dinners and hackathons. Selecting who to admit, brief sponsors, and later measure “did this community actually elevate companies?” required more than spreadsheets: multi-role scoring over an enriched applicant working set, a dated event briefing artifact, and longitudinal elevation research with explicit estimation labels.

How it worked

Intake

Applicants land in a working-set JSON the People UI reads.

Deep research

Step flags track profile research + detailed scoring completion.

Role scores

scores_v2 vectors for founder / eng / research / PM / investor / media.

Event report

Hackathon briefing with sponsors, people embed, projects, company chart.

Impact

Founder timelines + company ROI over post-event elevations.

Live interface (lifted)

Framed viewport shows the real HTML/CSS/JS. Controls work against anonymised local JSON — no network calls required for data.

app/people.html open ↗

What it produced — counted volumes

Each figure is a separate denominator. Do not merge them into one “people researched” headline.

applicants_working_set312People UI working setlen(applicants.json)

applicants_with_scores_v2200Scored in working setcount scores_v2 present

applicants_step_8_1_and_8_2_complete202Deep research + detailed score flagstruthy step flags

report_attendees_with_company_data250One-event company-data chartagi_house_report.json → companies_chart

founders_impact_records165Impact explorer rowslen(founders-impact.v1.json)

founders_impact_distinct_company_ids122Distinct company ids in impactunique company.id-new

events_catalog20Events in product cataloglen(events.json)

companies_ui_json151Companies UI filelen(companies.json)

projects25Hackathon projects filelen(projects.json)

full_enriched_db_objects13,733Private full corpus — count only, not shippedstream count Full-enriched-DB.json

final_2023_hack_db_objects2,391Private 2023 hack DB — count only, not shippedstream count Final_2023-hack-DB.json

What was hard / what it got wrong

Score free-text often named employers and GitHub identities — useful for operators, unsafe to publish raw.
New_founders-impact.v2.json is referenced on GitHub Pages but absent locally; the UI falls back to v1 (165).
Some companion files (dinner-guests.json, companies-new-format.json) are malformed on disk; the shipped UI path uses companies.json (151).
A prior portfolio pass merged 13,733 + 2,391 + 312 into one figure — retracted. Labels here keep denominators separate.

Real vs reconstructed

Real (lifted): People explorer, Firecrawl event report HTML, Impact people/companies UI + scoring/ROI logic, event shell, companies/events pages, score vector shapes, event/project/company counts.
Anonymised fixtures: names, emails, employers→sector·size, social URLs dropped, score comments scrubbed to numeric summaries, impact person/company labels synthetic, dates month-precision.
Not shipped: Full-enriched-DB.json, Final_2023-hack-DB.json, raw dinner-guest keys, any secret/env files.
Not invented: no fake Firecrawl SaaS dashboard; the report is the hackathon analysis product where Firecrawl appears as a sponsor brand.

Provenance: PROVENANCE.md