Welcome to the Home Page of Ljubomir Josifovski
Current - Machine Learning (ML) / Artificial Intelligence (AI) - again.
Prior - systematic trading, research and development.
Past - automatic speech recognition (ASR) in noise, spoken documents indexing and retreival with spoken queries (https://patents.google.com/patent/GB2404040A/en), speech synthesis, machine learning (ML).
Online coordinates
Linkedin ljubomirjosifovski https://www.linkedin.com/in/ljubomirjosifovski
GitHub ljubomirj https://github.com/ljubomirj
Hugging Face ljupco https://huggingface.co/ljupco
X @ljupc0 https://x.com/ljupc0, highlights https://x.com/ljupc0/highlights, posts search, local arhive
Hacker News ljosifov https://news.ycombinator.com/user?id=ljosifov (favourites)
Substack @ljubomirjosifovski https://substack.com/@ljubomirjosifovski
Bsky @ljupco.bsky.social https://bsky.app/profile/ljupco.bsky.social, posts search (click Latest)
Mastodon @ljupco https://mstdn.io/@ljupco
Email LjubomirJosifovski at gmail dot com
Recent projects
Recent repos I want to have them easy to find, especially work that lives on a branch rather than on a tidy GH profile pin:
- LoveWork - Work that you love, so you never work a day in your life. LoveWork is a work-discovery agent, personal guide to achieving your life's work. Movie stars and successful creatives have human agents who look continuously for interesting projects, people, and opportunities. LoveWork makes that kind of representation available to everyone. (WIP)
- ChEMBLdb-query - LLM-assisted SQL over the ChEMBL database, including local-model endpoint meshes. Latest v5 allows not only multiple rounds with judges, but allows for synthetic data generation.
- ds4: REAP compact support - running compact DeepSeek V4 Flash REAP GGUFs on Apple Silicon. (this REAP25 on HF)
- hipfire: feature/gbnf-and-reasoning-budget branch - Adds llama.cpp-like GBNF, grammar-aware loop-guard handling, max thinking token budget separate from the total max token budget, a server-side reasoning-budget interruption message, and greedy temperature-zero DFlash verification under an active grammar.
- Hermes agent: /advisor plugin - A persistent second-model review plugin for Hermes Agent. A second model that reviews the main agent's work each turn, and delivers structured advice inline. On a branch off recent Hermes HEAD and with several issues raised in the PR feedback fixed. Was initially stand alone repo Hermes agent /advisor plugin, but that's now obsolete.
- llama.cpp: Ling-2.6-Flash branch - Metal support and experiments for Ling-2.6-Flash. Runs this HuggingFace gguf on Apple Silicon metal. The same gguf is to be run in ds4 -agent and -server on Apple metal with this branch of ds4 too.
- Telegram Voice & Text Chat - Automated Customer Support Prototype. WIP make it into customer advocate agent.
- ASI and ROCm local LLM tests - benchmarks from an M2 Max 96 GB MacBook Pro and a Radeon 7900 XTX 24 GB desktop.
- nemotron-3-elastic-mlx - conversion tools for Nemotron 3 Elastic inference with MLX on Apple Silicon. The models are as gguf on HuggingFace.
- dotfiles - the machine-specific configuration that keeps all of this reproducible, or at least recoverable.
Offline coordinates
In Harpenden UK, from Skopje MK.
My happy place - sharing with PK (book), VJ, KJ.
--
LJ HPD Sun 6 Oct 07:50:30 BST 2024