I challenged myself to build a working AI product from scratch. The result: an AI that qualifies real estate leads and generates personalized email responses in under 1 second.
🔗 Try the Live Demo
➡️ Live Demo
Type any message as if you were a real estate prospect. The AI will:
- Score your lead (0-10)
- Extract property type, budget, location, urgency
- Generate a personalized email response
- All in ~700ms
This is real AI, not a simulation. Every response is generated live by Llama 3.1 8B via Groq.
What It Does
A real estate agency receives dozens of leads daily from portals and their website. Most take 2-4 hours to respond manually.
This system:
- Receives a lead via webhook (POST request)
- Qualifies it with AI (structured JSON: score, property type, budget, zone, urgency)
- Generates a personalized email response matching the lead context
- Returns everything in ~700ms
Honest Benchmarks
10 production tests, 10/10 correct:
| Test Case | Score | Correct? | Latency |
|---|---|---|---|
| Detailed buyer (budget+zone+phone) | 8/10 | ✅ | 701ms |
| One-word message | 2/10 | ✅ | 721ms |
| Message with typos | 8/10 | ✅ | 724ms |
| Seller (not buyer) | 8/10 | ✅ | 715ms |
| Rental (not purchase) | 7/10 | ✅ | 664ms |
| Spam | 0/10 | ✅ | 617ms |
| Investor (500-800K) | 9/10 | ✅ | 861ms |
| Creole language | 7/10 | ✅ | 793ms |
| Greeting only | 0/10 | ✅ | 750ms |
| Land purchase | 8/10 | ✅ | 219ms |
Average latency: 694ms.
Stack
- Backend: Python + Docker on Render (free tier)
- AI: Groq API with Llama 3.1 8B Instant (free)
- Frontend: Static HTML on GitHub Pages
- Cost: $0.00/month
Honest Comparison
| Solution | Cost/month | Latency | Quality |
|---|---|---|---|
| This project | $0 | 694ms | Llama 8B |
| Zapier + GPT-4 | $18 | ~3s | GPT-4 (better) |
| Make.com + GPT-4o-mini | $9 | ~2s | GPT-4o-mini |
| Human agent | $250+ | 2-4h | Variable |
Honest Limitations
- ❌ No email sending yet (generates but doesn't send)
- ❌ 0 customers
- ❌ Render cold start ~50s
- ❌ Llama 8B < GPT-4 on complex cases
- ❌ At scale, costs ~$9/mo (same as Make.com)
What I Learned
- Small models work for domain-specific tasks. 8B ≈ 70B on standard real estate queries.
- Building is easy, selling is hard. 24 rounds of development, still 0 customers.
- Honesty improves quality. Inflated benchmarks led to worse decisions.
Source Code
Would love feedback. Would real estate agencies pay for this? What should I improve first?
Top comments (0)