engblogs

Technical case study of an AI-powered scam network that used multilingual translation and cross-platform outreach (SMS to WhatsApp to Telegram) to recruit workers, outlining the ping–zing–sting workflow and detection indicators for disruption.

A technical case study of an AI-assisted task scam traced to Cambodia, detailing translation-driven workflows, cross-platform cold outreach, recruitment-style messages, and a three-phase ping, zing, sting pattern culminating in deposits and cryptocurrency payments.

A technical analysis of a China-origin covert influence operation that used AI-generated social media comments and internal documents to orchestrate engagement across TikTok, X, Reddit, and more, revealing patterns, targets, and policy implications.

A technical analysis of OpenAI's detection and disruption of a China-origin influence operation that used AI-generated multilingual social media comments and internal documents to simulate authentic engagement and influence discourse.

Pinterest leverages user surveys to enable healthier personalization that balances engagement and wellbeing, creating more relevant and inspiring content while avoiding pitfalls of depersonalized or purely engagement-based feeds.

Product Key Memory (PKM) sparse coders offer a faster training alternative to traditional TopK skip transcoders with comparable reconstruction error and improved interpretability through natural latent grouping.

Snorkel AI unveils two new products—Snorkel Evaluate and Expert Data-as-a-Service—offering enterprises a unified platform to develop, evaluate, and deploy specialized, expert-level agentic AI systems with scalable, trustable data solutions.

Leveraging Snorkel's AI Data Development Platform, this post details the data-centric creation, evaluation, and tuning of a specialized enterprise AI underwriting agent using benchmark datasets, custom evaluators, and reinforcement learning.

A comprehensive overview of enhancing the Pinterest home feed pre-ranking stage through advanced model architectures, efficient system design, and unbiased training data pipelines to significantly improve user engagement and recommendation quality.

Qwen3 32B, a 32-billion parameter dense LLM with hybrid reasoning, multilingual support, and agentic capabilities, is now accessible via Lambda's cost-efficient Inference API for advanced coding, reasoning, and multi-task applications.

A comprehensive guide to deploying real-world applications using Kamal 2.0, covering essential production practices, Docker integration, security, and practical tips for effective Rails 8 deployment.

Airbnb revolutionizes voice support by integrating machine learning models for speech recognition, intent detection, help article retrieval, and paraphrasing, enhancing real-time, self-service customer interactions and reducing reliance on human agents.