The problem
The immune system is the most powerful therapeutic tool in biology. We can sequence it, profile it, and perturb it. We cannot yet design with it. Patient-by-patient response remains a black box, navigated through trial and intuition.
CAR-T therapy is the sharpest example. In blood cancers, response rates reach 60 to 90 percent. In solid tumors, the same therapeutic concept lands at 9 percent. The gap is not biological will. It is a design problem at combinatorial scale.
The wedge
We are not building a predictive model. We are building a design optimization system that uses reinforcement learning, with the wet lab as the environment.
A foundation model proposes constructs. The lab measures them. The model updates. Each cycle compounds proprietary data that no public dataset can substitute for.
Why now
Two curves crossed. Foundation models can now internalize the structure of immune biology from the public corpus: sequences, structures, single-cell states. Wet-lab automation can now return signal at the cadence a learning system needs. The two have not been wired together at the level of a closed loop. We are wiring them together.
First proof
We are proving the platform on CAR-T design, the highest-stakes and most data-rich corner of immunotherapy. The first model trains where the data is densest. What works here generalizes to TCR-T, bispecifics, and the next generation of cell therapies.
Vision
A world where therapeutic design is an optimization problem, not a craft. Where every patient’s biology is the starting state of a model, and the answer is generated, validated, and shipped. We are building the company that makes that ordinary.