We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Why chase after massive, resource-hungry AI when smaller models can deliver smarter, safer results?
I’m starting a series of projects built with Elixir and Go, targeting Retrieval-Augmented Generation (RAG) and local AI implementations. If you're interested in following along, join my newsletter to keep up with everything I’m working on.
PrivateAI Hub: Your Local AI Assistant
PrivateAI Hub (still working on the name) is a local AI assistant designed to securely organize, manage, and process your documents, emails, and screenshots, with a focus on privacy and security.
It all began as a Retrieval-Augmented Generation (RAG) solution built for a client using OpenAI's API and popular RAG strategies like pre-retrieval optimizations, query rewriting and expansion, re-ranking, and context compression. Everything was running smoothly, but a major concern surfaced: security. When working with confidential documents, secure handling is essential—just imagine if your chat history were leaked to unintended recipients (Sounds familiar?). Alarming, right? The only effective solution was to run everything locally. However, running local models demands resources for efficient performance or fine-tuning. This led me to explore different approaches and techniques to make the system better.
Elixir shines in this scenario, not just through libraries like Bumblebee or Axon for neural network handling, but by leveraging the Actor Model for managing complex workflows. With Elixir, you can create a network of processes that operate in parallel, each handling tasks like listening for events, updating context, and refining prompt generation and context retrieval. This approach allows you to break down tasks into manageable layers or groups, where smaller models categorize prompts and determine the best retrieval strategies in parallel—yielding highly efficient and organized results.
I've been using the LattePanda Sigma board in my mini PC setup to build an AI-powered NAS-like system for processing my documents and managing all my Git projects. The performance has been quite impressive, and everything runs smoothly using Llama 3.2 models. However, as a prototype, it’s still a way off from being something I could develop into a product to sell.
The new LattePanda MU board is even more intriguing, as it offers the possibility of using two boards in parallel, creating a more compact solution with similar performance. I'm continuing to explore different configurations and ideas to optimize the setup. My vision is to create small device you can bring everywhere, with a powerful AI assistant that can handle everything from organizing your files to generating content, completely secure and private.
RedoMap
This project began as a simple solution to help people easily find and explore new places. I developed a web app that integrates seamlessly into any website and, with just a few clicks, provides in-browser navigation. No registration or installation is needed—simply scan the QR code, and you're ready to go!
Over time, we expanded to offer our users—hotels, tour operators, and others—tools to create and enhance content for a better guest experience. These tools include audio guides, detailed descriptions, chat support, and more. The latest version now incorporates AI-driven tools that empower our clients to create high-quality content effortlessly.
Yuriy Zhar
Passionate web developer. Love Elixir/Erlang, Go, Deno, Svelte. Interested in ML, LLM, astronomy, philosophy.
Enjoy traveling and napping.
This website began as my personal blog, but after building it, I realized I enjoyed the process of development more than writing posts. I kept adding new features instead of content, so I decided to shift focus to sharing ideas and projects on this page instead.
With the rise of AI, I started building more RAG (Retrieval-Augmented Generation) applications and AI tools, so most of my projects now center on this area. I primarily work with Elixir and Go, and occasionally use Python when fine-tuning models is needed.