engblogs

A practical, production-grade guide to engineering reliable AI agents by moving from vibe checks to continuous evaluation (CE) with regression testing, tool orchestration via the A2A Protocol, and end-to-end observability using OpenTelemetry, shadow deployments, and data-driven evaluation across Cloud Run, ADK, and Vertex AI.

ASPA extends the existing RPKI framework to cryptographically validate the entire AS path in BGP, helping prevent route leaks by authorizing upstream providers, with Cloudflare Radar's ASPA deployment monitoring and practical steps for creating ASPA objects and applying RFC9234 BGP roles.

Radar expands post-quantum readiness from client to origin connections, adds a post-quantum compatibility tester, and unveils a public Key Transparency audit for end-to-end messaging, plus ASPA-based routing visibility with API-accessible data.

Redesigning Turnstile and Challenge Pages at global scale to deliver a unified, accessible, multilingual security verification experience.

A critique of the WHATWG Web Streams API's usability and performance issues, arguing for a first-principles, pull-based alternative grounded in async iterables and explicit backpressure with batched, byte-oriented chunks.

Detecting and mitigating toxic combinations - how converging bot signals, misconfigurations, and exposed admin endpoints can drive security incidents and how edge protections surface and neutralize them.

An in-depth look at instantiating UI components with distinguishing variations using design-space sampling, symbolic inference, and an LLM-driven mimetic sampler, demonstrated via the Celestial tool to map and visualize component design spaces.

Scaling App Store search relevance by augmenting the ranker with fine-tuned LLM-generated textual judgments to complement behavioral signals, improving offline NDCG and online conversion as shown in A/B tests.

TabPFN AI accelerates business transformation on Databricks by delivering production-grade, seconds-fast predictions on structured data with a pre-trained, ready-to-use model, reducing preprocessing and retraining within a governed Lakehouse and scalable ML workflows.

Databricks at MWC 2026 showcases how telecoms turn fragmented data into real-time, AI-powered decisions across marketing, network investment, and customer care via a unified data platform and a forward-looking data strategy.

Leverages a top-down, retrieval-augmented generation pipeline with LLMs to deliver cohesive, personalized Shopping Hub recommendations at scale, spanning design, evaluation, and ranking.

A practical blueprint for COBOL modernization on the mainframe that combines AI-assisted reverse and forward engineering with deterministic, platform-aware analysis to deliver traceable and production-ready results at scale across large enterprise portfolios.