Abdalla would rather run the model on a box he can kick than rent a personality from a cloud bill. Fine-tuning, quantization, RAG — the whole “please fit in RAM” conversation. The bots live here. The spreadsheet can wait.
At Eleva they take meetings so humans do not have to. At Planet Land they answer WhatsApp. If it sounds like science fiction and also like a ticket in Jira, that is the job.
Local models
Qwen-family models, GGUF, Q8-class quantization. Trade context size against latency like a person who has paid for RAM. Local, server, and the occasional phone that should not melt.
Agents that do work
WhatsApp bots, Telegram bots, virtual AI brokers, meeting reports. Abdalla builds agents that capture leads and write PDFs, not agents that write poems about synergy.
Pictures on a box
ComfyUI, local VLMs, scene generation wired into apps. Not a slideshow of “AI art.” Actual pipelines with a progress bar that is allowed to tell the truth.