engblogs

Modular, community-driven update on research milestones and a growing Modverse ecosystem, highlighting easy MAX installs for local GenAI deployment and access to 500+ optimized models.

A unified embedding framework that combines context and user intent to enable scalable Text-to-SQL conversion.

A lightweight two-branch fusion of handcrafted cues—Wavelet-Denoised Feature with phase-only SPSL or Local Binary Patterns—built on an Xception baseline, boosts robustness in video face forgery detection with higher AUC and a smaller model footprint.

A semidiscrete optimal transport (SD-OT) approach to flow matching that replaces costly batch OT with a learned dual potential via SGD, enabling efficient SD-FM for training time-dependent velocity fields.

LogSentinel is a Databricks-powered, LLM-driven data classification and governance workflow that tracks schema evolution and labeling drift to automatically label PII, trigger remediation tickets, and feed improvements back into the Databricks Data Classification product via MLflow and a Mixture-of-Experts approach.

Explores how Descript enables scalable multilingual video dubbing, detailing the architecture, tooling, and end-to-end workflows required to deliver synchronized multilingual voiceovers.

A concise analysis of Codex Security as it enters a research preview, outlining its security posture, potential use cases, and considerations for developers and researchers.

Build a near real-time operational app by directly ingesting events with Zerobus Ingest into Delta tables and surfacing live analytics via Lakebase and Databricks Apps on the Databricks Data Intelligence Platform.

Guide to building and deploying multimodal AI agents for the Gemini Live Agent Challenge, using Gemini models, the ADK, and Google Cloud services to deliver real-time, immersive cross-modal experiences across Live Agent, Creative Storyteller, and UI Navigator.

Proactive, defense-in-depth playbook for hardening IT/OT, cloud, and virtualization against destructive cyberattacks in 2026, covering external-facing risk, segmentation, strong authentication, privileged access workstations, and immutable backups.

A formal framework for the controllability of generative models, treating human–model interaction as a control process and providing PAC-style guarantees on controllable-set estimation, with Plan-then-Generate approaches for data-to-text.

Technical exploration of how Balyasny Asset Management built an AI-driven research engine to accelerate data-driven investment insights and quantitative decision-making.