Building a Personalized Cognitive Extender: Edge AI with Smart Glasses & On-Device LLMs

Nikita Arsenov — Mon, 24 Aug 2026 09:56:46 +0000

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.

Hello World: From Sysadmin to Reinforcement Learning

Nikita Arsenov — Fri, 21 Aug 2026 08:01:21 +0000

Hey everyone, I'm Nikita.

I'm starting this blog as a home for my technical notes, side projects, and random deep dives. To be completely honest, I don't have a rigid roadmap for where this blog is heading—part of it is simply wanting a space for creative and technical self-expression.

A Quick Background

I’m a Ukrainian-Russian engineer by training, with around 4 years of experience in system administration. Over the last 1.5 years, I’ve been actively diving into Machine Learning, with a particular interest in cognitive science and reinforcement learning.

The Spark

Like many hobbies that turn into obsessions, it started with a single project: building and training a small-scale, AlphaZero-inspired model combining CNN + DQN + LSTM architectures.

It wasn't an overnight success. It took roughly three months of relentless tweaking—redesigning the architecture, fixing gradient issues, and fine-tuning reward functions—before the model finally started showing real, intelligent results. That breakthrough moment was when everything clicked: I realized just how much I love this field, and experimenting with ML officially became my main passion.

What to Expect Here

Going forward, this blog will be a mix of:

Ongoing ML and reinforcement learning experiments
Notes and thoughts on cognitive architectures
Practical takeaways, debugging stories, and engineering logs

Thanks for stopping by! If you're interested in ML, cognitive systems, or just like following along with hands-on technical experiments, feel free to connect or drop a comment.

Tags: #MachineLearning #DeepLearning #ReinforcementLearning #DevLog #Beginner

DEV Community: Nikita Arsenov

Building a Personalized Cognitive Extender: Edge AI with Smart Glasses & On-Device LLMs

Hello World: From Sysadmin to Reinforcement Learning

A Quick Background

The Spark

What to Expect Here