ApartmentFinder

An automated apartment finder that scrapes listings and uses a vLLM to evaluate them based on customizable criteria.

Features

• Scrapes apartment listings from Craigslist
• Uses Claude 3.5 and Sonnet Models to evaluate listing photos and descriptions
• Scores listings based on customizable criteria like:
  • Price
  • Square footage
  • Number of bedrooms/bathrooms
  • Visual aesthetics and layout
  • Location and neighborhood
• Stores results in SQLite database
• Deduplicates listings to avoid re-evaluating

Usage

Install the requirements

uv pip install -r requirements.txt
cd frontend
pnpm install

Change the config.py file to desired craiglist query url, and specify model settings. For those who wish to manage API keys in .env, write in following format:

ANTHROPIC_API_KEY=<anthropic-key>
OPENAI_API_KEY=<openai-key>
DATABASE_URL=<database-url>
SECRET_KEY=<secret-key>

And specifically for (optional) email server functionality:

SMTP_HOST=
SMTP_USER=
SMTP_PORT=
SMTP_PASSWORD=
SMTP_FROM=

Note that DB secrets point differently in dev and prod. (This is also probably going to happen with running the scraper as an independent service).

Run backend through docker compose up --build Run the frontend from the sub-repository using pnpm dev

If you want to observe the background task logs run celery -A app.celery_app worker --loglevel=info

Migrations

Use the makefile to generate and apply. Use migrate-init if first time managing migration in dev environment

TODO: Put intial alembic stamp in a fly.io intialization script. Something like:

# Check if alembic_version table exists
if ! psql $DATABASE_URL -c "SELECT 1 FROM alembic_version LIMIT 1" >/dev/null 2>&1; then
    # First time setup - stamp and upgrade
    alembic -c app/alembic.ini stamp head
fi

# Always run upgrade after
alembic -c app/alembic.ini upgrade head

Roadmap

In no particular order:

• Neighborhood evaluation via Google Maps streetview + walkability scores + distance to park + crime statistics, etc
• Fan out chain-of-thought evaluation for important criteria:
  • break down responses to complex questions like is this sound proof
  • considering floor plans, proximity to the street, insulation etc
  • a great place for it to shine is surfacing details you might have missed ahead of going, especially disqualifiers against a tour
• Deploy + regularly update db via cron job
• Explore better vLLM benchmarks + quantify performance/cost losses when using structured outputs
• IMPORTANT, batch API support to reduce costs
• Use context manager to streamline SQLAlchemy sessions
• Email notification system for top listings