Building an AI-Powered Apartment Hunter for Tel Aviv

ByDevin

# Building an AI-Powered Apartment Hunter for Tel Aviv

## The Problem

Finding apartments in Tel Aviv is chaos. Facebook groups, Yad2, scattered sources, thousands of irrelevant listings. **Goal:** One command. Qualified apartments only.

—

## The Solution: Apartment Hunter V2

### 3-Layer Filtering System

**Raw Input:** 469 Facebook posts
**Layer 1 – Smart Filter:** Remove noise (69% elimination)
**Layer 2 – GPT-4o-mini:** Semantic scoring
**Result:** 10 qualified apartments (₪8,000-12,000/month)

### The Pipeline

1. **Apify scrapes** Facebook groups (6 seconds, $0.12)
2. **Smart pre-filter** scores posts 0-100
3. **GPT-4o-mini** evaluates against criteria
4. **Telegram alert** for 8+/10 scoring apartments

### Key Metrics

– Posts → Qualified: 469 → 10 (97.9% noise eliminated)
– Signal/Noise: 7x improvement
– Cost per run: $0.23 (70% reduction)
– Time: 3 minutes

—

## The Wins

✅ **System works end-to-end**
– Apify reliability confirmed (6-second execution)
– GPT scoring accurate (9-10/10 for legitimate apartments)
– Telegram integration working immediately

✅ **Cost optimization**
– Original: 469 posts through GPT = $0.47
– Smart filtered: 145 posts = $0.14
– **70% cost reduction achieved**

✅ **Architecture validated**
– 3-layer filtering beats single-layer
– Source-level keyword filters reduce noise early
– Semantic filtering catches human judgment

—

## The Failures & Learnings

❌ **Apify cost limit not enforced**
– Runs complete in 6-30 seconds anyway (naturally cheap)
– Monitoring works, manual caps sufficient

❌ **mrkatplace endpoint unstable**
– Fallbacks working (Apify + native scraper)
– Learned: Always build redundancy

✅ **Spotify OAuth complexity**
– URI registration vs API format mismatch identified
– Framework ready for tomorrow

—

## What’s Next

✅ Apartment Hunter V2 ready (manual on-demand)
⏳ Wolt automation (Playwright coming tomorrow)
⏳ Spotify playlists (OAuth framework complete)
🔍 Apify monitoring (alerts if runs exceed $0.30)

—

## The Bigger Picture

**Layered AI filtering beats brute-force:**

– Skip expensive operations on noise
– Use smart scoring to identify high-value inputs
– Apply semantic reasoning only where it matters
– Cost drops 70%, quality improves 7x

Same pattern: apartment hunting, lead qualification, content scoring.

**One system. Multiple applications. Massive leverage.**

—

Building AI infrastructure for real problems.
Devin Pillemer — Tel Aviv, March 8 2026

Similar Posts