AI & ML interests
We build and optimize open-weight models that run on hardware you own from sub-100M task specialists up to frontier MoE builds — with a focus on edge and air-gapped industrial deployment where cloud APIs aren't an option. Our research is systems-first: mixture-of-experts expert-precision laws, speculative decoding economics, quantization and structured pruning, and the reproducible benchmarking needed to tell a real gain from run noise. We publish the harness, the exact commands, the run counts, and the spread — and we publish negative results, because those are usually the useful ones. Model lineage is labeled honestly: Trained / Adapted / Optimized / Verified.
RogerAI-fm 's datasets
None public yet