Checklists and critique decomposition pushed the LLM Judge's agreement rate with humans up to 84%.
The problem of generating explanations with no ground-truth data is solved through policy consensus and iterative evaluation.
On-device Android face verification optimization improved latency by 37% and throughput by 530%.
Business goals must be relaxed into smaller AI optimization problems to create impact.
We redesigned training and evaluation metrics to align with the goal of directly increasing conversion rate.
Azar used PolicyEngine to cut experiment setup and review time from 2 days to 30 minutes.
By automating 308 tests across 4 platform combinations, the time was reduced from 102 hours to 7 hours.
Using Flink's KeyedProcessFunction, they implemented real-time event combination and zero-downtime deployment.
Super Disk's write-mostly RAID cut ScyllaDB recovery time by up to 10x.
A method to narrow down Flink latency bottlenecks by separating processing time from non-processing time.
They replaced a legacy Flink app running on 96 CPUs with Flink SQL and GitOps deployment.
Proxies, rollback marking, propagation options, and Kotlin exception conversion determine what actually gets rolled back.
AI Briefing