Here is the current project I’m working on: a customizable “Cognitive Extender” tailored to my exact needs.
It’s no secret that Google makes its own Pixel smartphone lineup. Opinions on them vary, and I won’t praise the series unconditionally—it often lacks stability and optimization—but one thing is certain: it’s an incredible playground for experimentation thanks to Google’s open platform approach. My hardware stack: Even G2 + Pixel 9 Pro + Even R1 / Midra link.
Even’s native implementation of the AI assistant, translator, teleprompter, and other features is decent, but it is clearly designed as a generic, out-of-the-box solution that relies heavily on cloud processing. The Pixel 9, on the other hand, packs a TPU core inside the Tensor G4 ("Rio") paired with LPDDR5X RAM (45 TOPS INT8/INT4, up to 8,533 Mbps on a quad-channel bus). This architecture allows you to run on-device inference with the Gemini Nano v3 model (4B+, mixed INT4/INT2 hybrid quantization) virtually instantaneously—and completely LOCALLY.
Since I'm starting this blog mid-development, some of the groundwork is already laid. First off, I quickly abandoned the idea of writing fully custom firmware for the glasses. Even's engineering team did some impressive heavy lifting under the hood: a 4-MEMS microphone array handles initial ADC (PDM/I2S) -> the uncompressed audio stream routes to an integrated ARM Cortex-M for conversion to linear PCM, noise reduction, and voice isolation -> feeds into an LC3 encoder -> and the resulting LC3 frames are transmitted over BLE 5.4 to the official Even Hub.
Right around that point, I realized that writing my own LC3-to-PCM decoder was a fool’s errand—it’s already implemented natively inside the official Even Hub. As a result, at the entry point of our custom ehpk plugin, we are already working with raw, uncompressed 16-bit Mono LE PCM bytes.
