A stronger model cannot recover information it never receives. Kshetrajna Raghavan (Shopify) built the redesigned system that achieved ~75x lower inference cost and close to 2× improvement on selected quality metrics. MLOps World Austin, November 17 mlopsworld.com/

The catch with conversion has always been the retraining bill: with standard approaches it took 33× the token budget ScaleOp reaches the same quality at 3–18× instead Ravi Satya Durga Prasad Yenugula (2K Games) MLOps World Austin, November 18, 2:10 p.m. mlopsworld.com/

80% accuracy is a production target, not a ceiling. The remaining 20% is where your guardrails earn their keep. Suneet Malhotra ( @MotoSolutions ) redesigned the QA pipeline around Claude Code. MLOps World Austin, November 16, virtual session day mlopsworld.com/

Union.ai supports customers running 12 million or more containers on a control plane that holds no customer data. That was not the goal. Niels Bantilan will cover this architecture. They’re a Silver sponsor for MLOps World Austin mlopsworld.com/

Data engineering’s bottleneck has moved. Infrastructure is no longer the main bottleneck. Meaning and clarity in the data are. When machines take actions, semantic ambiguity becomes catastrophic. Joe Reis · MLOps World Austin, November 18, 2:10 p.m. mlopsworld.com/

Instacart cut feature store latency in half by running fewer nodes. Using Dragonfly’s multi-threaded engine, hundreds of nodes became roughly 100. This is part of our 7th annual MLOps World | GenAI Summit. Hope you can join us!

Nalin Dadhich, Senior Machine Learning Engineer, NVIDIA, built the synthetic data that took Cosmos 3 to #1. → Big Ideas, Innovations and Creative Explorations track. MLOps World Austin, Tuesday November 17, 2:10 p.m. mlopsworld.com

“Alternating between inference sampling and gradient updates creates severe low GPU duty cycles (~40%) during LLM reinforcement learning pipelines.” → Hardware and Chips track. MLOps World Austin, Wednesday November 18, 11:30 a.m

“A leaderboard win is not a product.” Kalpesh Sutaria (Director, NVIDIA) worked on Nemotron Retriever, which ranks #1 on RTEB but still wasn’t runnable for customer production. → Hardware and Chips track. MLOps World Austin mlopsworld.com

In their own words, cost came up more than anything else. @CoreWeave sits right in the middle of these discussions. They are our Platinum sponsor for MLOps World Austin. 📅 November 16 to 19 → Details and registration mlopsworld.com

A multi-agent pipeline that worked, but was slow and expensive. Parallelizing the independent roles cut runtime by roughly 40%. A per-role comparison of model tiers cut per-run cost by roughly 40% more. 📅 November 17 and 18, mlopsworld.com

A multi-agent pipeline that worked, but was slow and expensive. Parallelizing the independent roles cut runtime by roughly 40%. A per-role comparison of model tiers cut per-run cost by roughly 40% more. 📅 November 17 and 18, Austin. mlopsworld.com

Sandeep Mannapur (Staff Performance Engineer at ServiceNow) had three weeks of degraded responses already shipped by then, and every monitoring signal was green the entire time. → MLOps World Austin, Monday November 16, 12:05 p.m., virtual session day. mlopsworld.com

Jazmia Henry (Lead AI Research Engineer, collide.) is in our Evals and Benchmarks track, led by Suhas Pai. You leave with the case for decomposing reward into dimensions you can verify one at a time. 📅 Monday November 16, virtual session day mlopsworld.com

Attendees will learn from real case studies, and get a better idea of which metrics survive contact with production, and a better sense of where your evals may be giving false positives. 📅 November 17 and 18, 📍 Etter-Harbin Alumni Center, Austin. mlopsworld.com

Autonomous research agents that propose, implement, and refine ML solutions are priced out by the loop, not the model. → Recursive Self-Improvement track. MLOps World Austin, Wednesday November 18, 11:30 a.m., Etter-Harbin Alumni Center mlopsworld.com

67% of a multi-agent pipeline's latency came from a small minority of agent roles. The bottleneck wasn't the model. It was sequential chaining. Nadia Rauch (Chubb) on how parallelizing independent roles cut runtime by ~40%. Austin, Nov 17 mlopsworld.com/

Autonomous research agents are priced out by the loop, not the model. Vashishtha Patil (Amazon, Alexa+) tests a small open-weight model running the loop end to end, calling a frontier advisor only at set points on a metered budget. Austin, Nov 18 mlopsworld.com/

Where are your evals giving you false positives? KOHO, Wisedocs and Drew Crawford bring real case studies: governance on AWS, sensitive failure slices, the limits of LLM-as-a-judge, and agents evaluated against other agents. Austin, Nov 17-18 mlopsworld.com/