21CQ2 began as research

After building applications on cloud AI APIs, some basic questions recurred: how do these models work, and how do they produce the output we see as users? This was the impetus for 21CQ2's interpretability research program. Along the way we found that open weight models could be trained to specialize in unique documents at ~$0 token cost—from a wall outlet. This practice is that discovery in applied form.

We run every engagement end to end: scoping, deployment, validation, and training.