NLP · Search · 2025
Property Search NLP Parser — natural-language search that pays for itself
A natural-language property search that replicates realestate.com.au's beta NL search, using hybrid routing to cut LLM cost ~90% while improving location accuracy. Built to target an REA machine-learning role.
By the numbers
What it does
It turns a plain-English query like “3 bed house in Brunswick under 900k with a yard” into a structured search — the same class of feature realestate.com.au shipped in beta. I built it end to end to prove I could, and to signal genuine domain initiative toward an REA machine-learning engineer role.
The engineering decision that mattered
Sending every query to an LLM is accurate but expensive. Sending everything to rules is cheap but brittle. So I built hybrid routing: rules handle the common, unambiguous cases; the LLM is reserved for the genuinely hard ones.
- ~90% cost reduction — roughly $13.8K/yr vs $138K/yr at scale, because most queries never touch the model.
- +20% location-extraction accuracy over the naive baseline, by handling suburbs, landmarks, and fuzzy phrasing deliberately.
- 32 unit tests, 95% coverage — because “it worked on my example” isn’t evidence.
Accuracy the business can’t afford isn’t accuracy. The routing is the product.
Why it’s here
It shows how I actually think about production AI: not “can a model do this,” but “what’s the cheapest, most reliable way to do this at scale, and how do I prove it holds.”