Engineering notes, deep dives, and build logs from the people making LynxKite.
Evaluating toxicity is a multi-faceted, highly context-dependent problem. In collaboration with Hummingbird Bioscience, and with support from NVIDIA Inception, we built a context-first toxicity knowledge graph — extracting schema-constrained claims from regulatory documents, literature and trial data, then using structural modeling and graph-based ranking to turn...
Read postNVIDIA NIMs are a good option when the container already exists, but new biology models show up on Hugging Face all the time and we got tired of waiting for a vendor image. A LynxKite Inference Microservice (LIM) is a Python row function with @lim on...
Read postFine-tuning an LLM used to mean writing 400 lines of Python just to get started. We got tired of watching domain experts give up before the first training run. So we built a LynxKite LLM Training plugin: a visual, drag-and-drop workflow for the entire fine-tuning process....
Read postLynxKite has a flexible graph neural network designer where you can choose from a variety of GNN layers. Graph attention networks are a great choice in many applications. In addition to providing good predictions, GATs are unique in that the attention mechanism gives us insights into...
Read post