Problem
Rental listings in Vietnam are scattered and noisy — duplicates, stale posts, vague addresses. A search that returns them all returns nothing useful. The backend had to be the thing that cleans the data, not just store it.
What I built
- RESTful APIs for core rental operations, user management, and location-based services
- Scale validated in production: 2,000+ visitors and 100+ active users within the first two weeks of launch
- Map-based search letting users find and filter available rooms by precise geographic location
- An automated crawling and ingestion pipeline pulling from high-reliability sources, with geolocation verification and a deduplication algorithm that filters spam and duplicate listings before they reach users
- LLM-powered room recommendations and conversational search, so a query in natural language can be answered without the user learning the filter UI
Architecture
Service layer
Spring Boot over PostgreSQL for listings, users, and search; Redis for the hot read paths a browse-heavy product generates.
Ingestion
Crawler → geolocation verification → deduplication → publish. Listings enter through a pipeline, not by hand.
Search
Geographic filtering on the structured path, LLM-driven recommendation and conversational query on the soft path.
Decisions and trade-offs
- Deduplication at ingestion, not query time. Cleaning on write makes reads cheap and means the noise is dealt with exactly once; the cost is that a bad rule silently drops real listings, so the pipeline needed verification before publishing.
- Crawl from few, reliable sources. Wider crawling would have grown the dataset and the spam with it. Quality of the result set was the product.
- LLM for recommendation and query understanding, not for truth. It ranks and interprets; the listing facts come from PostgreSQL. Anything else invents rooms that do not exist.