Full-Stack Optimizations for Agentic Inference with NVIDIA Dynamo

Create an optimized inference engine layer specifically for 'agentic' workloads (multi-step, scratchpad-heavy). This should handle speculative decoding for agent reasoning tasks.

Suggested repo: AgentVroom

"Make your autonomous agents run 3x faster with agent-aware inference."

Estimated effort: 150h