Ryght, a startup building generative AI tools specifically for healthcare and life sciences, is publicly launching its Preview platform today after working closely with Hugging Face’s Expert Support Program to overcome the usual early-stage technical hurdles. The company’s pitch is straightforward: pharma and biotech firms are drowning in data from labs, EMRs, genomics, and clinical trials, but the tools to extract insights from that data remain painfully manual and slow.
What makes Ryght’s approach worth paying attention to is its architecture. Rather than betting everything on a single large language model, the platform treats LLMs as pluggable components. Each model gets registered and connected to customer-specific inference endpoints, which means Ryght can swap in new or specialized medical models as they appear without rebuilding the entire system. CTO Johnny Crupi credits Hugging Face’s Text Generation Inference and Text Embeddings Inference services for making this flexibility possible, calling them “game changers in their own right.”
The partnership with Hugging Face went beyond just infrastructure. Ryght’s team needed to get up to speed quickly on a fast-moving technical landscape, and the Expert Support Program provided advisory meetings, technical workshops, and asynchronous communication channels with people who operate at the center of the open-source AI ecosystem. That kind of guidance matters more than most people realize when you’re a startup trying to separate signal from noise in a field where new models and techniques appear weekly.
For an industry as regulated and privacy-sensitive as healthcare, the pluggable approach also addresses a real governance concern. By running TGI and TEI in their own environment, Ryght avoids the rate limits and lock-in that come with proprietary embedding APIs, while maintaining control over where data flows. Whether the platform delivers on its promise will depend on whether life science professionals actually find it faster than their current workflows, but the technical foundation at least doesn’t look like it’ll be the bottleneck.
