housing_rs

Taiwan publishes real-estate transaction data — but it's scattered, hard to query, and lacks analysis. housing_rs unifies nationwide transactions, neighborhood POIs, ML-based valuation, and LLM natural-language search into one Rust platform. The goal: let an investor type "Find me a property in New Taipei under NT$10M near a metro station" and get structured results.

Problem

• Government open data is messy, fragmented across APIs, and parsing-heavy
• No tool combines transactions + neighborhood data + ML pricing — investors patch it together manually
• Existing platforms are SPA + thin proxy — no real analysis

Solution

A Rust backend + ML microservice + LLM parser in three layers:

1. Rust data pipeline: tokio-cron-scheduler pulls government ZIPs on schedule; sqlx writes to PostgreSQL; PostGIS indexes handle spatial POI queries.
2. Python FastAPI ML service: XGBoost for valuation, Isolation Forest for outlier detection. Rust calls it over HTTP — keeps ML deps out of the Rust binary.
3. Claude API LLM search: User queries like "New Taipei under NT$15M, < 5 years old, near metro" parse to structured JSON, then Rust translates to SQL.
4. htmx + Alpine.js frontend: HTML over the wire — avoids SPA over-engineering. Leaflet for maps, ApexCharts for charts.

Stack

• Web: Axum 0.7 (tower ecosystem) + tokio + sqlx 0.8 compile-time queries
• DB: PostgreSQL 16 + PostGIS 3.4 (spatial indexing, nearest-neighbor)
• ML: Python FastAPI + XGBoost + Isolation Forest (separate microservice)
• LLM: Anthropic Claude API (structured output JSON mode)
• Cache / Limit: moka cache, governor rate limiting
• Deploy: Multi-stage Docker, healthchecks, request tracing

Outcomes

• M1 complete: skeleton + DB schema + transaction-data ingestion pipeline
• M2-M12 in plan: UI, map visualization, valuation accuracy benchmarking, LLM tool use
• Target M6 MVP, with 50% of investor queries done via LLM interface

Lessons

• Rust isn't only for HFT: sqlx compile-time query checks deliver huge maintenance value on mid-size projects — schema refactors no longer break SQL at runtime.
• HTTP-isolating the ML service is the right call: Rust ML crate ecosystem isn't mature; forcing it locks you into bad choices.
• LLM is not a SQL translator: Letting LLMs write SQL = open SQL injection. Structured JSON output, then your code translates to SQL — that's safe.