AI-powered "synthetic focus group" that predicts consumer purchase likelihood by simulating hundreds of virtual customers with realistic demographics.
Build a system that delivers both quantitative ratings and qualitative feedback in minutes at 12,000x lower cost than traditional human focus groups while achieving 7-12% higher accuracy.
5-Agent Modular System:
USER NICHE → Agent 0 (Topic Research)
→ Agent 1 (Product Research)
→ Agent 2 (Demographics Analysis)
→ Agent 3 (Persona Generation)
→ Agent 4 (Intent Simulation)
→ REPORT
| Agent | Purpose | LED Range | Output |
|---|---|---|---|
| Agent 0 | Topic Research | 500-599 | 5-10 ranked ebook topics by demand |
| Agent 1 | Product Research | 1500-1599 | Comparable products + data sources |
| Agent 2 | Demographics | 2500-2599 | Customer profiles (validated via triangulation) |
| Agent 3 | Persona Generator | 3500-3599 | 100-500 synthetic personas (reusable) |
| Agent 4 | Intent Simulator | 4500-4599 | Purchase intent predictions + recommendations |
- SSR (Semantic Similarity Rating): 90% correlation with human responses
- ParaThinker: 8 parallel reasoning paths, eliminates tunnel vision
- Triangulation Validation: Cross-validate demographics from 3+ sources (78-85% accuracy)
Instead of asking AI "Rate this 1-5" (unrealistic distributions), we:
- Generate 8 independent reasoning paths per persona
- Use semantic embeddings to map text to intent scores
- Achieve realistic distributions matching human surveys
Phase: Architecture & Research Complete
Ready to Build:
- ✅ Complete 5-agent design (see
Docs/4-agents-design.md) - ✅ Data gathering research (Reddit PRAW, YouTube API, Playwright)
- ✅ LED breadcrumb instrumentation defined
- ✅ Validation methodology established
Next Steps:
- Build Agent 0 (Topic Research) - Week 1 MVP
- Build Agents 1-3 (Product → Demographics → Personas) - Weeks 2-3
- Build Agent 4 (ParaThinker Intent Simulator) - Weeks 4-5
Beta Phase (Current):
- $0 per product test (using Claude Code subscription)
- Unlimited testing during development
Future (if scaling):
- Optional API-based SaaS with metered pricing
- Estimated: $1.05-$1.15 per first run, $0.50 per persona reuse
vs Traditional:
- Human focus group: $5,000-20,000 per product
- Savings: 12,000x during beta
Test 70+ book titles to find optimal one (real case study: author increased sales significantly)
- Product concepts validation
- Ad copy testing
- Pricing strategy
- Audience segmentation
- Market research at scale
Purchase-Intent/
├── .claude/
│ ├── agents/ # prd-simplifier, session-summarizer
│ └── commands/ # /end-session
├── Context/
│ └── 2025-10-22/
│ └── COMBINED-SESSION-HANDOFF.md # Session decisions
├── Docs/
│ ├── 4-agents-design.md # Complete architecture (v2.0)
│ ├── Research-customer-data01.md # Data gathering research
│ ├── PurchaseIntent-overview.md # Project vision
│ └── ANTI-OVER-ENGINEERING-GUIDE.md # Development philosophy
└── CLAUDE.md # Development guidelines
Data Gathering:
- Reddit API (PRAW) - 60 req/min free tier
- YouTube Data API v3 - 10k quota/day
- Playwright - Amazon/Goodreads scraping
Processing:
- Claude API - Demographic extraction
- Sentence Transformers - Clustering
- ParaThinker - 8-path parallel reasoning
Storage:
- JSON files - Persona inventory
- Git - Version control
- 5-Agent Architecture - Complete technical specification
- Research Report - Data gathering tools & methods
- Session Handoff - Development decisions
- Anti-Over-Engineering Guide - Development philosophy
This is currently a solo project in active development. See CLAUDE.md for development guidelines.
[Add your license here]
- Research Papers:
- SSR:
Docs/LLM-Predict-Purchase-Intent.pdf - ParaThinker:
Docs/ParaThinker.pdf
- SSR:
- GitHub: https://github.com/Bladed3d/PurchaseIntent
Built with Claude Code