DATA Pill #195 – Platform Engineering, Container Scaling, Analytics Skills, Agent Frameworks, File Formats & Hybrid LLMs

ARTICLES

Santander’s “Catalyst” platform reduces infrastructure provisioning from weeks to hours. Built on Amazon EKS and Crossplane, the platform defines resources through Git repositories; ArgoCD syncs changes into multi‑cloud clusters and Open Policy Agent enforces compliance centrally. The bank has rolled out generative‑AI agent stacks using Bedrock, S3 and KMS with custom IAM policies; proof‑of‑concept preparation dropped from 90 days to 1 hour and monthly data‑platform tickets fell by 3,000. The result is a self‑service culture where developers can launch secure environments rapidly and platform teams maintain governance.

Netflix rethought its container strategy to harness modern CPUs. Containers provide flexible deployment and shorten time to production, but new CPUs with hyper‑threading and larger caches required revisiting resource allocation. Engineering teams identified bottlenecks in per‑container mount operations and implemented advanced monitoring to track resource usage and AI‑driven orchestration that anticipates demand spikes and reallocates workloads

Xebia offers a step‑by‑step guide to build a modern data transformation pipeline using free tools. It recommends a serverless warehouse such as Google BigQuery and dbt for SQL‑based transformations; Git/GitHub handle version control and GitHub Actions automate CI/CD. The roadmap covers provisioning BigQuery, connecting dbt Cloud to a GitHub repository, organising models into staging/intermediate/marts layers, adding tests and exposures, and integrating BI tools

RockBot is an open‑source framework for building multi‑agent systems where agents and user proxies communicate through a message bus. The author created it after finding existing frameworks insecure and monolithic; many run LLM‑generated code in the host process, making it hard to swap providers and isolating runaway code. RockBot solves this by running each agent as an isolated process that subscribes to topics on a RabbitMQ‑backed bus, calls tools/LLMs as needed and emits responses

Iceberg has finalized a File Format API that makes file formats pluggable and engine‑agnostic. The new layer replaces fragmented logic across Spark, Flink and Java readers, eliminating large switch statements and uneven feature support. Core concepts include a FormatModel describing a format’s name, reader/writer construction and capabilities, and a FormatModelRegistry that engines query to obtain read/write builders.

The Allen Institute for AI released Olmo Hybrid 7B, a language model that mixes transformer attention with Gated DeltaNet linear recurrence. By replacing 75 % of attention layers with DeltaNet heads, the model achieves the same accuracy as Olmo 3 7B while using 49 % fewer tokens

FREE COURSES

Anthropic’s course portal offers a range of free learning paths. Courses include: Claude 101, AI Fluency, Building with the Claude API, Introduction to Agent Skills and more

DATATube

A short explainer on the hidden technical debt created when AI generates code that humans no longer fully understand. It cautions teams to pair LLM‑generated code with documentation and code reviews to avoid long‑term maintainability issues.

A practical tutorial comparing Python’s uv package manager to pip. The presenter demonstrates faster installs, offline caching and dependency resolution, showing why uv improves the developer experience.

TOOLS

A terminal‑based tool that inspects your hardware and suggests large language models that fit your CPU, GPU and memory. The interactive TUI lists models with estimated token/s throughput, context length and quantization; CLI mode supports searching, planning required hardware and exposing a REST API

Modal’s research platform hosts an API endpoint for Z.ai’s GLM‑5 model (745 billion parameters). The endpoint allows one concurrent request and is free until 30 April 2026; a curl example shows how to call the /chat/completions API with the zai-org/GLM-5-FP8 model

Fluss is a streaming storage layer for real‑time analytics and lakehouse architectures. The Rust client enables table management and log streaming operations; it details how to start a local Fluss cluster (requiring Java 17+), create tables and append logs.

The official Anthropic marketplace (claude-plugins-official) comes preconfigured; categories include code intelligence(connecting language‑server plugins for C++, Python, Rust, etc.), external integrations (GitHub, Jira, Slack), development workflows (commit commands, PR review agents) and output styles for educational or learning‑focused responses

CONFS, EVENTS, WEBINARS & MEETUPS

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