Carlos Mendez

Three Tesla M40 GPUs, released in 2015, considered e-waste by most. I used one to train a production extraction model in under two hours. The AI hardware gatekeeping narrative is wrong.

I benchmarked seven small language models on entity extraction for my knowledge graph. Four of them scored literally zero. The one that worked best? A 1.5B model I fine-tuned myself.

When picking a base model for fine-tuning, the best raw performer isn't always the right choice. I chose a smaller model over a better-scoring one, and it paid off.

I went from benchmark results to a deployed extraction model in 48 hours. LoRA training, GGUF conversion, and Ollama registration on consumer hardware, no cloud required.

Four aging 4TB drives. One already dead. 1.67 terabytes of production data serving five Docker containers, a GitLab instance, and an AI memory system. I handed the migration to my AI assistant at 10pm, went to bed, and woke up to a faster, healthier storage array. Here's what actually happened.