ARTICLES
Building an MCP Ecosystem at Pinterest | Tan Wang | Pinterest| 6 min | AI Agents
Pinterest describes how it is building an internal ecosystem based on the Model Context Protocol (MCP). The system standardizes how agents interact with tools, data sources, and services, enabling reusable integrations across teams. This approach simplifies agent development and supports scaling internal AI use cases.
The RL Algorithm Behind DeepSeek’s Breakthroughs | Miguel Otero Perido, Antoni Zarauz Moreno |The Neural Maze | 8 min | LLM Trainin
This deep dive explores the reinforcement learning techniques powering DeepSeek models. It focuses on scalable RL strategies, reward modeling, and optimization methods that improve reasoning and alignment. The piece highlights how RL pipelines are evolving beyond traditional RLHF toward more efficient and scalable training paradigms.
FEAST MEETS ORACLE: UNLOCKING FEATURE STORE FOR ORACLE DATABASE USERS | Aniket Paluskar, Srihari Venkataramaiah | Feast | 5 min | ML Infrastructure
Feast introduces an Oracle-based offline store, enabling enterprises to integrate feature stores with existing data warehouse ecosystems. The setup allows teams to use Oracle for historical feature retrieval while maintaining consistency with online serving. This approach supports governance, scalability, and enterprise-grade security requirements when deploying feature platforms.
Agentic FinOps for AI: Autonomous Cost Optimization Across Cloud Platforms | Preeti Shrimal, Niladri Ray| Flexera | 7 min | FinOps / AI
Agentic FinOps introduces AI-driven systems that automatically optimize cloud and AI infrastructure costs. These systems monitor usage, adjust resources, and enforce policies across platforms like Snowflake and Databricks. The approach moves FinOps from reactive reporting toward autonomous cost management.
From Parquet to Iceberg: The Evolution of Lakehouse Storage | Marek Wiewiorka | Xebia | 6 min | Data Architecture
This article explains how modern lakehouse architectures evolved from Parquet-based storage to table formats like Apache Iceberg. These formats add features such as schema evolution, time travel, and transactional guarantees. The result is a more reliable and flexible storage layer for analytics and AI workloads.
NEWS
Announcing General Availability of Real-Time Mode for Apache Spark Structured Streaming on Databricks| Navneeth Nair, Giselle Goicochea | Databricks | 4 min | Data Infrastructure
Databricks announces general availability of real-time mode for Spark Structured Streaming. The feature reduces latency by processing data continuously instead of micro-batching, enabling near real-time pipelines. This improvement makes Spark more suitable for streaming use cases such as fraud detection, monitoring, and real-time analytics.
Polars-bio Available as a Skill on KDense AI | 2 min | AI Tools
Polars-bio is now available as a skill, enabling bioinformatics and genomic workflows directly through LLM interfaces. This integration brings high-performance data processing to scientific workloads, making it easier to analyze large-scale biological datasets with AI.
Open SWE: An Open-Source Framework for Internal Coding Agents | LangChain | 4 min | AI Agents
LangChain introduces Open SWE, a framework for building internal coding agents that automate software engineering tasks. It supports workflows such as code generation, debugging, and repository interactions, helping teams integrate AI deeper into development processes.
Introducing NVIDIA NeMoClaw for Agentic Systems | NVIDIA| 3 min | AI Infrastructure
NVIDIA introduces NeMoClaw, a framework designed for building and orchestrating agentic AI systems. It focuses on scalable deployment, tool integration, and enabling complex multi-agent workflows across enterprise environments.
DATATube
GitHub Actions are DEAD. (Use Agentic Workflows instead) | Better Stack | 7 min
GitHub is moving beyond traditional automation with Agentic Workflows, enabling teams to manage CI/CD pipelines using natural language instead of rigid scripts. The video shows how to build an AI DevOps assistant that applies judgment through “productive ambiguity,” while still respecting security guardrails.
A practical breakdown of a real-world AI tool stack used to run a software agency. Instead of listing tools, the video shows how they are connected into a system using MCP integrations, what each tool replaces, and common mistakes that break workflows.
TOOLS
GitNexus is a developer tool that helps navigate and understand large codebases. It provides contextual insights, repository exploration, and AI-assisted code understanding, making it easier to work with complex projects.
CONFS, EVENTS, WEBINARS & MEETUPS
Operationalizing AI Agents: From Experimentation to Production | MLOps Community | March 25, 2026
A livestream panel exploring the challenges of deploying AI agents at scale. Speakers from Databricks, GrottoAI and Sentick discuss reliability, cost, governance and security barriers; they compare LLMOps/AIOps/AgentOps with traditional MLOps and outline success criteria for generative‑AI frameworks, including evaluation and observability
Data & AI That Matter From systems to people: building data and AI with real-world impact| Webinar Online | March 26, 2026
Not all data and AI initiatives deliver value. This webinar series focuses on impact over hype — how data platforms, AI systems, and teams can be designed to genuinely support people, decisions, and outcomes
PINNACLE PICKS
Data Lakehouse Explained: Architecture Powering Modern Data & AI Platforms |Marek Wiewiórka | Xebia | Data Architecture | 7 min
The lakehouse architecture combines the flexibility of data lakes with the reliability of warehouses. By using open table formats on object storage and separating compute from storage, organizations can support BI, streaming, and machine learning on the same platform. Modern lakehouses rely on open catalogs, multi-engine compute, and governance layers to avoid vendor lock-in and enable scalable data platforms.
Feast + MLflow + Kubeflow: A Unified AI/ML Lifecycle | Francisco Javier Arceo, Nikhil Kathole | Feast | ML Infrastructure | 12 min
Building production ML systems requires coordinating feature management, experimentation, and model deployment. This architecture combines Feast for feature storage and retrieval, MLflow for experiment tracking and model registry, and Kubeflow for orchestration. Together they create a modular open-source stack that helps teams move from feature engineering to production pipelines with consistent workflows.
The lakehouse architecture combines the flexibility of data lakes with the reliability of warehouses. By using open table formats on object storage and separating compute from storage, organizations can support BI, streaming, and machine learning on the same platform. Modern lakehouses rely on open catalogs, multi-engine compute, and governance layers to avoid vendor lock-in and enable scalable data platforms.
Feast + MLflow + Kubeflow: A Unified AI/ML Lifecycle | Francisco Javier Arceo, Nikhil Kathole | Feast | ML Infrastructure | 12 min
Building production ML systems requires coordinating feature management, experimentation, and model deployment. This architecture combines Feast for feature storage and retrieval, MLflow for experiment tracking and model registry, and Kubeflow for orchestration. Together they create a modular open-source stack that helps teams move from feature engineering to production pipelines with consistent workflows.
_____________________
Have any interesting content to share in the DATA Pill newsletter? Reach Out!
