Experiment 01
AMD BC-250
50%+ higher peak tokens/s
Ornith · Vulkan tuning + MTP
Read the experimentKnowix / Independent research
AI Lab
Local models. Hardware I own. Experiments I can measure.
I’m building toward an independent lab for local AI. The work starts here: getting more from accessible hardware, understanding inference bottlenecks, and building useful systems around local models.
Experiments & results
01 published studyWhat I’m investigating
Inference on accessible hardware
Vulkan kernels, quantization, and speculative decoding. I measure complete generation runs to see which changes are worth keeping.
Ornith resultsLocal agents with working memory
Long-context models are one part of the problem. I’m exploring compaction and retrieval as the next steps toward longer-running local assistants.
The 70K context experimentThe hardware journey
From Scrap to Spark, toward a lab.
Scrap to Spark is the hardware side of this work: building practical AI compute from second-hand equipment, starting with a $130 board. AI Lab brings the experiments, results, and next research questions together.