Has anyone successfully implemented Tejas Chopra's MLOps scaling strategies for real-time inference? Hitting a wall!

I'm trying to apply some of Tejas Chopra's principles for scaling machine learning models in production, particularly around efficient real-time inference. We're struggling with latency spikes under load, even after optimizing our model serving architecture. Curious if anyone has successfully navigated this using his recommended approaches and could share practical tips or common pitfalls to avoid. Really hoping for some actionable advice here!