Atlas Brief

Atlas Brief is a creator-screening product built for the RoC hackathon. It ranks creators using a hybrid model that combines semantic relevance with commercial quality, then presents the result in a guided landing page, an interactive demo, and a short methodology appendix.

What It Does

• screens the full creators.json universe
• runs the official challenge query for brand_smart_home
• exports the top 10 as JSON
• presents the results in a polished dashboard with charts, filters, and PDF export

Challenge Check

• searchCreators(query, brandProfile): RankedCreator[] is implemented in src/searchCreators.ts
• Postgres + pgvector setup is included in sql/schema.sql
• DB ingestion is included in scripts/ingest.ts
• official top-10 output is written to output/brand_smart_home_top10.json
• README setup instructions are included here
• Loom is the only submission item that must still be recorded separately

Quick Start

For the fastest local path with no OpenAI key or Postgres:

npm install
env EMBEDDING_PROVIDER=local VECTOR_BACKEND=memory npm run demo
npm run dashboard

Open:

• http://127.0.0.1:4173/ for the landing page
• http://127.0.0.1:4173/demo.html for the demo
• http://127.0.0.1:4173/methodology.html for the methodology page

Official Challenge Output

The required submission query is:

Affordable home decor for small apartments

using the brand_smart_home profile.

The generated output file is:

output/brand_smart_home_top10.json

Scoring

Official challenge ranker

projected_normalized = (projected_score - 60) / 40

final_score =
  100 * (
    0.45 * semantic_score +
    0.55 * projected_normalized
  )

This follows the recommended starting point from the challenge: semantic relevance blended with a slightly stronger projected-value weight.

Atlas demo score

commercial_quality =
  0.65 * projected_normalized +
  0.20 * engagement_norm +
  0.15 * gmv_norm

relevance_score =
  (industry_match + query_overlap + audience_fit) / 3

atlas_score =
  100 * (
    0.60 * relevance_score +
    0.40 * commercial_quality
  )

The demo score is for exploration. It uses equal treatment inside the relevance block, then blends relevance against commercial quality.

Full Setup

For the intended OpenAI + Postgres/pgvector path:

cp .env.example .env

Set:

OPENAI_API_KEY=your_real_key
EMBEDDING_PROVIDER=openai
VECTOR_BACKEND=postgres
DATABASE_URL=your_real_postgres_url

Then run:

npm run ingest
npm run demo
npm run dashboard

The intended submission path is the Postgres + pgvector flow above. The in-memory mode exists only for easy local review of the UI.

Main Commands

npm run typecheck
npm run ingest
npm run demo
npm run dashboard

Key Files

• src/searchCreators.ts - official challenge ranking logic
• scripts/ingest.ts - embedding and storage pipeline
• scripts/demo.ts - reproducible challenge run
• dashboard/index.html - landing page
• dashboard/demo.html - live dashboard
• dashboard/methodology.html - formulas and references
• output/brand_smart_home_top10.json - top 10 submission output

Notes

• local mode is the easiest review path
• OpenAI + Postgres/pgvector is the intended production-style path
• the dashboard can export both JSON and PDF from the demo page