[b.] creative | Auction Intel Lead Generation Platform
Auction Intel — Project Portfolio Summary
Auction Intel (auctionintel.app) — An AI-powered lead intelligence platform that discovers nonprofit auction events and delivers verified, sales-ready leads to event-services vendors.
The Problem It Solves
Companies that provide services to nonprofit auction events (auctioneers, event planners, catering, A/V rental) have no reliable way to find upcoming events before their competitors. Manually searching thousands of nonprofit websites is impossibly time-consuming. Auction Intel automates this: upload a list of nonprofit domains, and the platform uses AI web research to discover which organizations are planning auction events, extract event details, find decision-maker contacts, verify email deliverability, and deliver structured billable leads — all in real time.
Platform Overview
| Layer | Technology |
|---|---|
| Backend | Python 3 / Flask (~6,800 lines in app.py alone) |
| AI Engine | Claude Haiku 4.5 with web search via Poe Bot API |
| Database | PostgreSQL on Railway (14 tables) |
| Payments | Stripe (Payment Intents, webhook-driven wallet top-ups) |
| Resend API (20 HTML templates) | |
| Email Verification | Emailable API (real-time deliverability scoring) |
| IRS Data | PostgreSQL table of IRS 990 nonprofit filings |
| Real-Time Streaming | Server-Sent Events with reconnect+ polling fallback |
| Deployment | Railway (auto-deploy on push to main) |
| Frontend | Server-rendered HTML, Tailwind CSS, vanilla JS |
| API Client | Standalone Python CLI for headless batch operation |
Technical Architecture
| System Component | Implementation Details |
|---|---|
| LLM Service Layer | Wraps an LLM with web search into a metered SaaS with per-result billing, tiered pricing, and real-time wallet deductions. |
| Real-Time Event Streaming | SSE with numbered event replay, heartbeat keepalives, 20-attempt reconnect, automatic polling fallback, and per-domain checkpointing. |
| Semantic Cache Layer | TTL varies by result type: found events expire after the event date, not-found in 30 days, errors in 7 days, uncertain in 1 hour. |
| Multi-Stage LLM Parsing | Four extraction strategies in sequence plus data-driven status override that trusts extracted fields over model self-classification. |
| End-to-End Data Pipeline | IRS data (prospect discovery) → AI web research (event verification) → email validation (lead quality) → Stripe (monetization) → 20-email lifecycle (retention) — all in a single Flask app with zero external job queues. |
Platform Scale
| System | Metric |
|---|---|
| Lines of Python | ~9,500 across 5 core files |
| PostgreSQL tables | 14 |
| Flask routes | 50+ |
| Email templates | 20 |
| CSV output columns | 18 per lead |
| IRS filter fields | 30+ |
| Max nonprofits per job | 5,000 |
| Max IRS results per query | 10,000 |
Video Tour