Minotaur

Posted on Jul 15

I Built a Tool That Scrapes Facebook Groups and Uses a Fine-Tuned LLM to Parse Flat Listings

#ai #showdev #webdev #python

Finding a flat in Hyderabad through Facebook groups is painful. Hundreds of posts daily, zero structure - just walls of text with phone numbers, random pricing formats, and "DM for details." I built a tool to fix this.

The Problem

Indian cities have dozens of Facebook groups like "Flats Without Brokers Hyderabad" where landlords post listings. But:

No consistent format (rent might be "15k" or "₹15,000" or "fifteen thousand")
Buried in comments and unrelated posts
No way to filter by BHK, location, budget, or furnishing
You have to scroll endlessly, every single day

What I Built

A pipeline that:

Scrapes Facebook groups using Playwright (headless browser, cookie-based auth)
Parses unstructured text into structured data using a cascade of regex → fine-tuned LLM → keyword enrichment
Stores listings in PostgreSQL (Supabase) with deduplication
Serves a searchable web UI (Flask) with filters for rent, BHK, location, furnishing
Runs on a schedule via GitHub Actions (scraping) + Render (web app)

Here's what the end result looks like - 304 listings, all parsed and filterable:

Architecture

Facebook Groups (16 groups)
        │
        ▼
┌─────────────────┐
│  Playwright      │  Cookie auth, scroll, expand "See more"
│  Scraper         │  Extract raw post text
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Cascade Parser  │
│                  │
│  1. Regex first  │  Fast extraction: rent, BHK, sqft, phone
│  2. LLM fallback │  Fine-tuned Gemma 3 1B (LoRA) for ambiguous fields
│  3. Enrichment   │  Keyword-based: amenities, parking, facing
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  PostgreSQL      │  Supabase (prod) / SQLite (dev)
│  + Dedup         │  Skip already-seen listings
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Flask Web UI    │  Filter, search, wishlist
│  on Render       │  Mobile responsive
└─────────────────┘

The Interesting Part: Cascade Parsing

The naive approach would be: throw every post at an LLM and get JSON back. But that's:

Slow (API call per listing)
Expensive at scale
Overkill for fields regex handles perfectly (rent, BHK, phone numbers)

Instead, I built a cascade:

Layer 1: Regex

Handles the easy stuff instantly:

# Rent: matches "15k", "₹15,000", "15000/-", "rent 15000"
match = re.search(r'(?:rent|₹|rs\.?)\s*(\d[\d,]*)', text, re.I)

# Phone: Indian mobile numbers with Unicode digit handling
normalized = text.translate(str.maketrans('𝟬𝟭𝟮𝟯𝟰𝟱𝟲𝟳𝟴𝟵', '0123456789'))
match = re.search(r'(?:\+91[\s-]?)?([6-9]\d{9})', normalized)

Layer 2: Fine-Tuned LLM

For ambiguous fields (furnished status, gated community, available dates), I fine-tuned Gemma 3 1B using LoRA:

Training data: 200+ manually labeled listings from the actual Facebook groups
Finetuned on Apple Silicon (MPS) using HuggingFace transformers + PEFT
Runs locally via MLX for inference - zero API costs
Only called when regex returns null for these specific fields

Layer 3: Keyword Enrichment

Post-processing that catches what both regex and LLM miss:

Amenities detection (lift, gym, security, power backup)
Parking type inference
Floor/facing extraction from context

The result: ~85% of fields are filled by regex (fast, free), ~10% by the fine-tuned model, and ~5% by enrichment. Average parse time per listing: under 200ms.

You can filter by fully furnished, gated community, no broker, budget - and it actually works:

And expanding a listing shows the raw Facebook post alongside the parsed structured data:

Fine-Tuning Details

# LoRA config for Gemma 3 1B
lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
)

Training data format (chat-style JSONL):

{"messages": [
  {"role": "system", "content": "Extract flat rental info from Facebook posts. Return JSON only."},
  {"role": "user", "content": "2BHK semi furnished flat in Gachibowli..."},
  {"role": "assistant", "content": "{\"bhk\": 2, \"furnished\": \"semi\", \"location\": \"Gachibowli\", ...}"}
]}

6 epochs, batch size 1, lr=2e-5. Trained in ~20 minutes on M-series MacBook.

Deployment

Scraper: Runs locally (Facebook blocks cloud IPs). Could move to a residential proxy setup.
Web app: Deployed on Render (free tier), auto-deploys from main
Database: Supabase PostgreSQL (free tier, SSL connection)
CI: GitHub Actions checks linting and deployment

Results

557 listings scraped from 16 Hyderabad Facebook groups
Searchable by rent range, BHK, location, furnishing status
Wishlist feature to save favorites
Found my current flat through it

Tech Stack

Component	Technology
Scraping	Python, Playwright (async)
Parsing	Regex + fine-tuned Gemma 3 1B (LoRA/MLX)
Database	PostgreSQL (Supabase) / SQLite fallback
Web UI	Flask, Jinja2, mobile-responsive CSS
Deployment	Render (web), GitHub Actions (CI)
Auth	Cookie-based Facebook session

What I'd Do Differently

Add notifications - alert when a listing matching your criteria appears
Better dedup - some listings are reposted with slight text changes
Map view - plot listings on a Hyderabad map by location
Multi-city - architecture supports it, just need more groups

Source Code

GitHub: github.com/MinotaurG/fb-flat-finder

If you're apartment hunting in an Indian city and drowning in Facebook group posts, feel free to fork this and point it at your city's groups. PRs welcome.

DEV Community