Research
Our research is about running frontier-level AI on the hardware people already own. We study how models can reason in fewer tokens, keep their quality at lower precision, and fit on anything from a single GPU to a phone.
Publications
- How we taught Swift to stop overthinkingHow we found the tokens behind reasoning loops in Qwen 3.8 27B, penalized them at inference time, trained the behavior into a LoRA and recovered the accuracy it cost.Reasoning efficiencyOct 30, 2026Read
- Fine-tuning SAM 3 to read mapsSAM 3 selected the correct region on 40% of our land-use maps. Fine-tuned on a small set of annotated maps, it selected 80%.Computer visionJan 2026Read
- A multi-stage text-to-SQL pipeline for faster, more reliable answersA two-stage text-to-SQL agent that answers in seconds on locally hosted Qwen3-235B models, where LlamaIndex's agent workflow took minutes.Natural language interfacesDec 2025Read