Insights
Practical, opinionated writing on building AI and data systems that hold up in production.
Showing 15 of 50
WOSS
Your API gateway can't see a single MCP tool call. agentgateway is the Rust proxy that puts policy, security and observability in front of agent traffic - and it just became AAIF's fourth project.
WOSS
Governance is member-gated. Code isn't. Here's exactly where an individual engineer can land a first PR across all four Agentic AI Foundation projects - and the house rules each one will close your PR for breaking.
AI Models
Anthropic's most capable model is the first from its Mythos tier, a rung above Opus. The positioning, the specs that matter, the benchmarks, and the practitioner's rule for when its 2x price is actually worth paying.
AI Tooling
Everyone publishes the concept. Almost nobody publishes the fields, the ingest contract, or the dispatcher. This is a working LLM wiki taken apart piece by piece - the permission model, the three frontmatter fields that carry the system, the five-step ingest, the invariant that makes automation safe, and the two things that actually break.
WOSS
CNAI is the practice of running AI workloads the cloud-native way - on Kubernetes, declaratively, at scale. Here's the definition, the CNCF working group behind it, and where to go deeper.
WOSS
Not everything in open source ladders up to the Linux Foundation. The ASF is a separate root, 300+ projects strong, and it governs the data layer that modern AI runs on.
GenAI & Agents
A 95%-per-step agent completes a 10-step workflow 60% of the time - that math, not model quality, is why agents fail. Where agents genuinely work in 2026, and the verification-first architecture that gets them there.
WOSS
PyTorch stopped being just a framework. As a full umbrella foundation it now governs the open stack from training to inference to agents - PyTorch, vLLM, DeepSpeed, and Ray.
WOSS
Kubernetes, Prometheus, Envoy, OpenTelemetry - and 230 more projects. A decade in, the CNCF is the largest open-source foundation on earth and the center of gravity for cloud native.
AI Strategy
Model metrics are proxies, proxies degrade under optimization (Goodhart), and the fix is organizational: metric trees, guardrail pairs, attribution via experiment, and a named owner for the business number.
WOSS
Spark made distributed data processing feel like writing ordinary code. Spark 4.0 pried the client loose from the cluster and made strict SQL the default.
WOSS
Airflow turned data pipelines into version-controlled Python. Airflow 3 rebuilt the engine underneath - service-oriented, versioned, and no longer chaining tasks to the metadata database.
WOSS
OpenSearch began as a community fork of Elasticsearch. In 2024 it moved under the Linux Foundation, trading single-vendor control for neutral governance - with AWS, IBM, SAP, and Uber backing it.
WOSS
Kafka is the durable, high-throughput event log at the ingestion edge of nearly every modern data platform. In 4.0 it finally cut its last external dependency: ZooKeeper is gone.
WOSS
Modern software is mostly other people's open source. The OpenSSF builds the tools - Sigstore, Scorecard, SLSA - that let you trust the supply chain underneath it.
The people building the systems. Posts come from the delivery side of the firm, which is why they tend to cover failure modes and trade-offs rather than announcements.
Data science, predictive analytics, AI engineering including RAG and agents, MLOps, and the product engineering that ships them. If it touches getting a model into production and keeping it there, it is in scope.
Open-source and community writing runs through WOSS, the Welzin Open Source Society, which takes article submissions and residency applications.
Posts are the general case; your system is not. Send the specifics and get a reply from someone who builds these for a living.
Talk to a senior engineer