Chintan Parikh

Chintan Parikh is a Google product manager working on LiteRT, the company’s cross-platform runtime for deploying AI models on phones, computers, browsers, and connected devices. His focus is making sophisticated models practical outside data centers, where latency, privacy, unreliable connectivity, battery life, and hardware limitations determine what applications can actually do.

Earlier in his career, Parikh moved from test engineering into business and product responsibilities. His software projects include Serene, a minimal PHP framework, and Shel, a flat-file blogging engine.

• A universal runtime for on-device AI. Parikh has helped develop and explain LiteRT’s evolution beyond TensorFlow Lite. The runtime supports models originating in PyTorch, TensorFlow, and JAX, preserves the portable TFLite format, and provides deployment paths across Android, iOS, desktop systems, browsers, and connected hardware.

• Practical neural-processor acceleration. His work on production NPU deployment addresses the gap between specialized chips and usable applications: unified hardware interfaces reduce vendor-specific integration while improving latency, power consumption, and sustained performance. Related Arm optimization work combines PyTorch conversion, selective quantization, XNNPACK, and Arm KleidiAI for on-device audio generation.

• LiteRT.js and browser-native inference. Parikh coauthored the introduction of LiteRT.js, bringing the runtime into JavaScript applications through WebAssembly and WebGPU, with emerging WebNN support. He and Matthew Soulanille also introduced the technology at the WebAI Summit.

• On-device agents beyond chat. At AI Engineer Europe 2026, Parikh demonstrated how Gemma edge models can call tools, produce structured output, interpret images, and interact with local applications. His examples included voice-driven personal tracking, privacy-sensitive workflows, hybrid device-and-cloud processing, and a Raspberry Pi robot—illustrating how smaller models become useful when paired with capable runtimes and appropriate hardware.