engblogs

Lyft's Feature Store architecture and optimization: a platform-of-platform data system unifying batch, online, and streaming features, with governance, low-latency serving, and lifecycle tooling to accelerate ML at scale.

Lyft’s Feature Store architecture combines batch, streaming, and online serving to unify feature management, detailing optimization, governance, and developer-friendly improvements that scale ML for a rideshare platform.

Scaling from a single Engagement Agent MVP to a dispatcher-driven, enterprise-scale multi-agent platform that delivers high-throughput, fair, and observable AI-powered outreach at scale.

Comprehensive analysis of a Venezuela BGP route-leak involving AS8048 (CANTV), examining probable routing-policy misconfigurations and highlighting ASPA and RFC9234 as paths toward a safer, more verifiable BGP ecosystem.

NarrativeTrack introduces a benchmark for evaluating temporally grounded narrative understanding in multimodal language models, using entity-centric reasoning and a Compositional Reasoning Progression framework that scales from entity existence to changes and ambiguity.

A designer-aligned pipeline for improving UI generation models that leverages feedback via commenting, sketching, and direct manipulation to finetune LLMs with multimodal data and human evaluations, outperforming traditional RLHF baselines.

Introducing the Instructed Retriever, a system-aware retrieval architecture that propagates system specifications (user instructions, labeled examples, and index schema) through query generation and retrieval to enable robust instruction-following, low-latency enterprise agent search.

One-year retrospective on Google Public Sector's AI-enabled cloud, security-first innovations, and agentic technologies like Gemini that accelerate government missions across defense, civilian agencies, and education.

Practical guide to designing language-learning support systems that empower learners through choice, forgiving feedback, and low-pressure practice.

A practical guide to replacing monads with algebraic effects in OCaml 5, using the Handled_effect library to implement clean, composable Hardcaml FPGA testbenches and simulations.

BCBS 239 compliance is reframed as an AI-powered, governance-driven, lakehouse-enabled transformation on Databricks that automates risk data aggregation, accelerates regulatory reporting, reduces costs, and scales for evolving rules like DORA and Basel IV.

MIT researchers reveal how AI trained on de-identified electronic health records (EHRs) can memorize patient data and outline practical privacy evaluation methods to curb leakage in clinical foundation models.