November 2025 Summaries

Chalk has expanded its platform across flexibility, visibility, and efficiency, enhancing its capabilities as a data platform for AI and ML applications. The updates introduce model registration directly in Python, allowing models to be versioned, type-checked, and integrated into workflows alongside features. Tracing capabilities have been improved to diagnose query performance, while vector aggregations now support advanced embedding workflows, beneficial for recommendation systems and fraud detection. Dynamic Expressions enable runtime logic adaptation without redeployment, and the expanded expression library now includes over 50 new Velox functions and sklearn classifiers. Improved visibility features include enhanced diff viewing for code changes and offline input exploration for dataset building, which aid in introspection and validation. New CLI workflows support the creation of structured prompt templates for LLM-assisted development, standardizing workflows and accelerating onboarding. Notable adopters like Whatnot, iwoca, and Mission Lane leverage Chalk for real-time recommendations and decision-making, with resources available for deep-dives into Chalk's integration into ML stacks.

Mission Lane's recent live demo showcased how they efficiently manage thousands of features for over two million customers in production, with Mike Kuhlen detailing their use of Chalk to streamline complex feature dependencies across various processes like training, live decisioning, and batch evaluation. By defining features once and reusing them across multiple applications, such as real-time apps and monthly batch jobs, they significantly reduce coordination overhead that previously required collaboration across data science, data engineering, and ML engineering teams. Chalk enables their models to request features by name, automatically fetching the necessary data without altering decisioning system code, which simplifies handling dependencies at scale and ensures up-to-date data for batch scoring. This approach has transformed their process, allowing data scientists to self-serve with Python code that is universally applicable, thus eliminating the need for extensive coordination across teams and systems.