engblogs

A concise look at modeling device capabilities, feature flags, and scalable analytics to optimize feature deployment across Netflix's diverse device landscape.

A data-driven framework for modeling device capabilities and feature flags at scale to optimize analytics, enable granular feature delivery, and measure reach across diverse devices and capabilities.

A comprehensive device capability data model coupled with feature flags enables scalable analytics and granular feature enablement across Netflix's diverse devices.

Cross-domain prompt-injection study uncovers transferable attack shapes that span 21 scenarios, and offers practical benchmarking and hardening guidance to secure AI agents beyond siloed domain testing.

Cross-domain prompt injection study reveals that most winning attacks share transferable rhetorical shapes across domains, implying domain-specific testing misses the real exposure and urging cross-domain benchmarks and design principles to separate trusted instructions from untrusted input.

How GitHub's code search achieves memory-speed case-folding via a branchless ASCII-first loop and byte-space Unicode folding, implemented in Rust as the casefold crate.

Rust-based, branch-free case-folding at memory speed for code search, combining an ASCII-first, decode-free byte-space fold with SIMD-like vectorization and a compact Unicode table, implemented in the casefold crate.

A focused examination of BeeKEM, a decentralized CGKA protocol for local-first apps, exploring causal ordering, cross-fork security, and the tension between forward secrecy and concurrency in decentralized E2EE.

Avatarin uses GPT-Realtime to power a 24/7 multilingual retail agent for Yamada Denki, turning sales expertise into proactive, context-aware conversations that blend voice, text, and visuals.

GPT-5.6 advances price-performance for enterprise AI by delivering Luna, Terra, and Sol with lower costs, faster modes, and optimized inference to enable high-volume, efficient workflows.

A case study of Avatarin’s 24/7 multilingual retail agent powered by GPT-Realtime that translates sales expertise into proactive, voice-enabled, cross-channel shopping conversations using retrieval-augmented generation.

Analyzes how GPT-5.6 and its Luna/Terra/Sol lineup advances the price-performance frontier through efficiency improvements, lower pricing, and API Fast mode to accelerate enterprise workloads.