About Me

I'm Hasib Ziai, a software developer with a B.S. in Computer Science from Cal State Fullerton. I build across a wide range of projects, from web apps on Cloudflare Workers to C++ tooling and Java mods for Minecraft. Alongside development, I have 8+ years of experience in systems administration, supporting a network of small businesses in Arizona. I managed their entire IT operations, including their recovery from a ransomware attack and the preventative measures to stop the next one. I also audited their IT practices and built an automated pipeline to move them to the cloud.

Outside of work I'm drawn to the security side of things, specifically network security, wireless signal analysis, open source intelligence (OSINT) gathering, application security, and operating systems. I was a member of CSUF's Offensive Security Society, and I've since presented common RFID attacks and other security topics to school administrators, campus safety, and law enforcement. I like learning how systems break so I can rebuild them harder, better, faster, stronger.

I've always been interested in music systems in games, especially where online social interaction and music intersect. My interest in networking and the rollback systems used in fighting games got me making my own mods, and eventually adding dynamic NPC dialogue to games. That's what led me deep into open source LLMs and the projects around them. Back in 2023, to get those NPCs to actually hold a conversation, I started running models locally and spent a lot of time in the KoboldAI and SillyTavern communities. I put those models to work as NPCs on the Minecraft server I was running at the time, where they could hold simple conversations in character. Getting that right meant experimenting with different prompting styles and different ways of storing conversation memory to keep token usage down. I used the same techniques in a few similar side projects.

That work led to about a year with PygmalionAI, a community effort to train a model for casual dialogue, which was exactly what I needed. We tried a few base models and ended up using Llama for our 7B fine-tune. I eventually stepped away when the project's direction drifted from what I was building, but I came out of it knowing how training and fine-tuning works. One of the models we trained is up on my Hugging Face as a quantized build.

Around the same time, I worked with db0 and others on their distributed computing system, the AI Horde, which KoboldAI plugs into. Volunteers donate spare compute to run models for others, and every finished job earns you "Kudos", which are purely for leaderboard bragging rights. I put my homelab to work on it, and it climbed into the top 10 pretty quickly.

These days I'm spending a lot of my time in Java, finishing an instrument mod for Minecraft. I'm also trying out newer open models like Xiaomi's MiMo and learning to train my own LoRAs for them. My daily driver is CachyOS with Hyprland.