optimization
End-to-end tuning: fine-tuning, distillation, joint training.
The anchor essay. Why assembled GenAI stacks stall below the bar, why no vendor in the stack is positioned to fix them, and the end-to-end loop that is: evaluate, fine-tune, and own every layer against your number.
The library behind our agent-optimization engagements, out of stealth: a PyTorch-shaped training loop for agent runtimes, with textual gradients, batch-level updates, and constraint-preserving steps. Why we built it instead of adopting what existed.
A field note on legal tool-call optimization: thin tool routing, per-tool argument builders, and failure-driven prompt/program optimization for higher argument accuracy with less context pressure.
A ten-point component lead all but disappeared in the stack that shipped. The product boundary, not the leaderboard, is where the decision belongs.