When we set out to build TypePilot, our primary constraint was latency. A keyboard is the most heavily used interaction point on a smartphone. If a keyboard lags by even 50ms, the user feels it. Early on, we experimented with embedding small quantized LLMs directly into the Android app. While this provided excellent privacy, the cold-start times and battery drain were unacceptable for a utility meant to be used hundreds of times a day.

We pivoted to a cloud-first approach and benchmarked several providers. Google Gemini provided the best balance of speed, strict prompt adherence (crucial for our JSON-based tone shifting features), and robust Android integration. By relying on Android's secure system-level handling for password fields, and only transmitting explicitly selected text to Gemini, we achieved the perfect balance of premium AI capabilities and strict user privacy.