Cormac Brick

Cormac Brick is a principal engineer at Google AI Edge developing the software and hardware techniques that bring capable AI to phones, laptops, embedded computers, and robots. His work focuses on making models small, fast, private, and reliable enough to operate locally on affordable devices.

From specialized processors to on-device AI

At Movidius, Brick worked on low-power machine intelligence and demonstrated the Movidius Fathom Neural Compute Stick at NeurIPS 2016: a USB accelerator that ran neural-network inference directly on a Raspberry Pi. After Intel acquired Movidius, he served as a director of machine intelligence in its Movidius group and later led architecture work on laptop neural processing units.

At Google, he became a technical lead for the on-device stack encompassing MediaPipe, LiteRT, and the open-source LiteRT-LM inference framework. In 2024, he coauthored Google’s introduction of the AI Edge Torch Generative API, which helps developers deploy PyTorch-built language models through Google’s mobile runtime. He also contributed to technical reports for Gemma 3, Gemma 4, and EmbeddingGemma, collaborating with Gemma developers to make their models run effectively across many devices.

How he makes small models useful

• Choose the right deployment model. Brick distinguishes operating-system models such as Gemini Nano, shared through Android’s AI Core, from models packaged inside individual applications. Shared models avoid duplicating large downloads; application-specific models offer greater customization and broader hardware reach, at the cost of additional engineering.

• Train tiny models for specific jobs. His team applies synthetic-data fine-tuning to FunctionGemma, a 270-million-parameter model, for mobile function calling. Targeted training improved an example workflow from roughly 46 percent accuracy to the mid-80s overall, with most individual functions exceeding 90 percent. The result depends on narrowly defined tasks and rigorous evaluation, not generalized intelligence.

• Keep agent context lean and actions predictable. For on-device agent skills, Brick favors loading detailed instructions only after a model selects a relevant capability. This progressive disclosure reduces context and computation; constrained decoding further improves reliability by limiting tool calls to permitted operations. The open-source Google AI Edge Gallery demonstrates this architecture across mobile applications.

• Design around real hardware constraints. Available memory, quantization, latency, and component cost determine whether a model can reach older phones, Raspberry Pi boards, or inexpensive robots. LiteRT supports portable deployment across CPUs and GPUs, while neural processing units require specialized, vendor-specific compilation.

• Combine focused models into complete products. Brick’s team built AI Edge Eloquent, an offline dictation application that combines separate Gemma-derived models for speech recognition and text polishing. The system removes verbal fillers, adapts to specialized vocabulary and unfamiliar names, and illustrates how compact, inspectable model pipelines can deliver practical on-device products.