AudienceLab

Local-first, persona-driven multi-agent simulator for influencer marketing. It simulates influencer follower behavior on a local SNS and uses those actions (like/comment/follow) to predict marketing performance via engagement proxies.

Hackathon context: OpenAI Coxwave Hackathon, Seoul (January 2026). This repository is a prototype built for a live demo, not production software.

What It Demonstrates

• Start from a pre-selected influencer shortlist (search is out of scope)
• Explain why each candidate fits the brand goals and tone
• Simulate follower behavior using persona agents on a local SNS
• Aggregate engagement proxies (likes, comments, follows) to forecast campaign performance
• Surface metrics and qualitative signals in a reporting dashboard

Closed loop: data → personas → simulation → metrics → ranking + rationale

System Architecture

graph TD
    Shortlist[Pre-selected Shortlist] --> Dashboard["Dashboard (React + Vite)"]
    Dashboard -->|Configure run| Orchestrator["Agent Orchestrator / Runner"]

    subgraph Swarm["Follower Persona Swarm (N agents)"]
        AgentA["Agent A"] --> SNS["SNS-Vibe (SvelteKit + SQLite)"]
        AgentB["Agent B"] --> SNS
        AgentN["Agent N"] --> SNS
    end

    Orchestrator -->|Spawns| AgentA
    Orchestrator -->|Spawns| AgentB
    Orchestrator -->|Spawns| AgentN
    Orchestrator -->|LLM decisions| OpenAI["OpenAI API"]
    Orchestrator --> Shared["shared/simulation/*.json"]
    Orchestrator --> Logs["dashboard/public/simulation/*.jsonl"]
    Shared --> Dashboard
    Logs --> Dashboard
    Shared --> Eval["eval-agent"]
    Logs --> Eval
    Eval --> EvalOut["dashboard/public/evaluation/latest.json"]

Loading

Repository Layout

Path	Role
agent/	Persona-based multi-agent simulator (Playwright + OpenAI)
dashboard/	Simulation + reporting UI (React/Vite)
sns-vibe/	Lightweight SNS for fast local runs (SvelteKit + SQLite)
eval-agent/	Evaluation scripts (metrics + LLM comment quality)
shared/	File-based contracts and outputs (JSON Schema)
docs/	Architecture, evaluation, troubleshooting, quickstart
images/	Demo screenshots

Quick Start (Fastest Demo Using SNS-Vibe)

Prerequisites:

• Node.js 18+
• Python 3.11+
• uv (Python package manager)
• OpenAI API key

1) Start SNS-Vibe

cd sns-vibe
npm install
bash scripts/reset-db.sh
npm run dev -- --port 8383

2) Start Dashboard

cd ../dashboard
npm install
npm run dev

Open the URL printed by Vite (typically http://localhost:5173).

3) Configure Agent

cd ../agent
cp .env.sample .env
# Set OPENAI_API_KEY and SNS_URL=http://localhost:8383
uv sync
uv run playwright install chromium

4) Run Simulation

uv run python local_agent.py
# or: uv run python runner.py --num-agents 3

5) (Optional) Run Evaluation

cd ../eval-agent
cp .env.sample .env
uv sync
uv run python evaluate.py

Outputs and Contracts

• Simulation status: shared/simulation/{simulationId}.json
• Live agent logs: dashboard/public/simulation/*.jsonl
• Evaluation output schema: shared/evaluation/result-schema.json
• Evaluation snapshot: dashboard/public/evaluation/latest.json

See shared/README.md for the file contracts and schema locations.

Documentation

• docs/QUICKSTART.md
• docs/ARCHITECTURE.md
• docs/EVALUATION.md
• docs/TROUBLESHOOTING.md

Component READMEs:

• agent/README.md
• sns-vibe/README.md
• eval-agent/README.md
• shared/README.md

Demo Screenshots

Known Limitations

• Influencer search/ranking is mocked; the shortlist is assumed
• Single-simulation workflow only (no parallel runs)
• Local-only infrastructure; no persistent storage
• Engagement is a proxy, not real conversion data

Team (Hackathon)

Name	Focus
Sungho Park	Simulation + SNS
Seunghyun Lee	Crawler