Project Name
Home Hunt
Home Search Product
An AI-powered home search app with addictive Tinder-style UX. Built to have efficient inference economics: the app sifts through 60k London listings for just $0.01 per user
Efficient inference makes this product viable.
Aggressive optimisation was achived with daily evals run over OpenRouter's available models. Combined with detailed analytics of inference costs, through extensive database logs and dashboards. This information allows daily scheduled agents to propose daily recommendations on optimising costs further.
Per user commute calculation
Unlocks good matches in overlooked areas for the user and better filtering. Commutes are calculated from actual TFL queries to a commute postcode specified by the user.
Support Agent
Answers questions about how the app works, and can answer user-specific questions to the logged in customer.
Inference Optimisations
* Model selection: 8x decrease in costs both across multimodal models and text-to-text only models. An automated daily eval runs tests of new models on OpenRouter for potential savings.
* More efficient prompts: Firstly, overall prompts got 2x shorter in tokens. Secondly, a higher accuracy in ranking is achieved by providing the entire listing as context for the ranking model.
* Prioritising what to rank: a simple linear regression model selects the most likely homes to be generally desirable. Expected to improve over time as usage of the platform increases.
* Not ranking everything: Some homes are by default not appropriate to the user: they are too expensive, have a commute that's too long, or are of a niche listing type (auctions, cash only, houseboats) which will be caught by a different classification pass.