HostelHound

I built HostelHound to help travelers compare hostel prices across dates and decide when to book. It tracks 15,000+ hostels across nearly 800 cities, with a calendar for nightly prices, property-level price history, and forecasts for price drops and rises.

A Rust scraper records availability and prices in TimescaleDB, and a SvelteKit frontend serves the search and planning tools. Travelers can save and share multi-stop itineraries with live hostel quotes. The ChatGPT connector exposes 10 MCP tools for discovery, comparisons, forecasts, recommendations, and trip planning.

The forecasting pipeline runs daily. LightGBM models estimate drop probability, price direction, and the size of a potential change. Nightly outcome tracking compares forecasts with later prices so I can monitor precision rather than judge the models only by offline scores. Email alerts let travelers watch prices without repeatedly checking the calendar.

I also built a browser monitor on self-hosted GPU infrastructure. LangGraph agents explore the site across account types, Playwright checks exercise known workflows, and a separate model checks candidate findings. In one recorded evaluation, 331 reports were consolidated into 46 unique production defects, leading to 39 fixes or mitigations. Provenance checks and duplicate suppression helped keep repeat reports from inflating the results.

The scraper, database, and web app run on an Oracle Ampere A1 VM. Terraform manages infrastructure, Cloudflare Tunnel provides ingress, and PgBouncer pools database connections. The map view uses MapLibre to browse cities on a globe.