The Challenge
A PropTech startup wanted to give estate agents the ability to generate instant property valuations during on-site visits — even in areas with poor signal. Existing solutions required connectivity and had multi-second latency that broke the conversation flow.
Approach and Execution
The model was trained on 800K property images with associated sale prices, then distilled and quantised to run inside Core ML on iOS. I built the full React Native camera pipeline, on-device inference layer, and the server-side telemetry pipeline that fed back user interactions to retrain the model monthly.
Results
Key Insight
The biggest challenge was not the model itself but the data quality pipeline. Listing images varied wildly in angle, lighting, and crop. Augmentation alone was insufficient — we needed to build a systematic feedback loop where agent corrections refined the training distribution over time.
Outcome
The feature shipped in v2.1 and became the app’s highest-rated feature within 60 days of launch. The offline capability was the standout differentiator in rural markets.
React Native · Core ML · TensorFlow Lite · BigQuery