Welcome to The Weekly Five - your curated list of 5
exceptional open source projects I discovered this week.
The Weekly Five: Intelligent Automation Across the Stack
From autoscaling your Kubernetes workloads based on real-time events to building AI agent harnesses that orchestrate LLM interactions, this week's picks show how intelligent automation is reshaping every layer of the modern stack. Whether you're scaling infrastructure, training ML models, or streamlining content workflows, these tools share a common thread: they handle the complex orchestration so you can focus on building.
Top takeaways
Event-driven autoscaling (KEDA) and distributed ML training (XGBoost) bring production-grade automation to infrastructure and data pipelines without custom glue code
AI-assisted development is moving beyond code completion into full CMS platforms (Webiny) and agent harness patterns (basically-ai-harness) that wrap LLMs with tools, context, and guardrails
Native tooling for everyday tasks (Markdown Preview) rounds out a productive stack by eliminating context switches for documentation review
Who this issue is for
Platform engineers, ML practitioners, and full-stack developers building systems where automation, scalability, and AI-assisted workflows intersect.
KEDA
Why this made the cut: The de facto standard for event-driven autoscaling in Kubernetes, with 60+ scalers covering queues, databases, and custom metrics.
Why it matters
Traditional Horizontal Pod Autoscaler relies on CPU and memory, which misses the mark for queue-driven or schedule-based workloads. KEDA extends Kubernetes with scalers that react to external event sources (Kafka, RabbitMQ, Azure Queue, cron, and more), scaling deployments down to zero when idle and up instantly when demand spikes.
Key features
60+ built-in scalers: Connect to message queues, databases, HTTP metrics, and cloud services out of the box
Scale to zero: Eliminate idle compute costs by scaling replicas to zero and waking them on demand
ScaledObject and ScaledJob resources: Declarative CRDs that integrate cleanly with GitOps workflows
CNCF graduated project: Production-hardened with broad community support
How to use
Install KEDA via Helm: helm install keda kedacore/keda --namespace keda --create-namespace
Define a ScaledObject referencing your deployment and desired scaler (for example, Redis list length or cron schedule)
Apply the manifest and watch pods scale in response to events
To test the waters, start with a cron-based ScaledObject that schedules scale-up windows at predictable times, then graduate to queue-based scaling (such as Redis list length triggers) once you have confidence in the behavior
Monitor scaling events with kubectl get scaledobjects and check replica counts to verify KEDA is responding correctly
🔗 View on GitHub | GitHub stars: 10,542
Learning resources
Event-driven autoscaling in Kubernetes with KEDA Tutorial - Civo - Step-by-step tutorial covering KEDA basics, ScaledObject configuration, and practical examples with cron-based and Redis queue-based scaling.
XGBoost
Why this made the cut: Nearly 29K stars and battle-tested across Kaggle competitions and enterprise pipelines, XGBoost remains the go-to gradient boosting library for tabular data.
Why it matters
Gradient boosting dominates structured data problems, and XGBoost delivers a portable, high-performance implementation that scales from laptops to distributed clusters. The same codebase runs on Spark, Dask, Flink, and Kubernetes, so you can prototype locally and deploy at scale without rewriting training logic.
Key features
Multi-language bindings: Python, R, Java, Scala, C++, and more
Distributed training: Native support for Hadoop, Spark, Dask, Flink, and Kubernetes
GPU acceleration: CUDA-enabled training for faster iteration on large datasets
Built-in regularization: L1 and L2 penalties reduce overfitting out of the box
How to use
Install via pip: pip install xgboost
Load your dataset with pandas or use DMatrix for memory efficiency on large files: dtrain = xgb.DMatrix(X_train, label=y_train)
Train with the native API (xgb.train(params, dtrain)) or the scikit-learn interface (XGBClassifier, XGBRegressor) for familiar fit/predict patterns
For distributed training, configure your cluster backend (Spark, Dask) and pass the appropriate parameters to the training call
Serialize the model with model.save_model('model.json') for deployment, then load it anywhere with xgb.Booster().load_model('model.json')
🔗 View on GitHub | GitHub stars: 28,784
Learning resources
Why XGBoost Beats Deep Learning on Tables - DataMListic - Clear breakdown of why gradient boosting still dominates tabular data, with practical comparisons against deep learning approaches.
Webiny JS
Why this made the cut: A serverless, self-hosted CMS that pairs enterprise features (multi-tenancy, GraphQL) with AI-assisted development via an MCP server.
Why it matters
Headless CMS platforms often force a choice between flexibility and operational overhead. Webiny runs entirely on AWS serverless primitives (Lambda, DynamoDB, S3), eliminating server management while offering lifecycle hooks and a GraphQL API for deep customization. The integrated MCP server brings AI-assisted development into the CMS layer itself.
Key features
AWS serverless architecture: Lambda, DynamoDB, and S3 with infrastructure-as-code deployment
Multi-tenancy built in: Serve multiple brands or clients from a single deployment
GraphQL API: Flexible content queries and mutations for any frontend
AI-assisted development (MCP server): Use LLM tooling to scaffold content models and extensions
How to use
Bootstrap a new project: npx create-webiny-project my-project
Configure AWS credentials and deploy with yarn webiny deploy
Access the Admin app to define content models via the UI or programmatically through lifecycle hooks
Query content from your Next.js or React frontend using the auto-generated GraphQL endpoint
Enable the MCP server integration to let AI tools assist with schema and plugin development
🔗 View on GitHub | GitHub stars: 8,042
Learning resources
Learn Webiny Course - Webiny - Six-chapter course covering foundations, deployment, headless CMS, website builder, and AI-assisted development with the MCP server.
basically-ai-harness
Why this made the cut: A minimal, educational repo that demystifies AI agent harnesses with step-by-step branches and a companion deep-dive article.
Why it matters
As LLM-powered agents proliferate, understanding the infrastructure that wraps around models (tools, context injection, guardrails, verification loops) becomes essential. This repo, built for an AI Engineer Europe 2026 talk, breaks harness engineering into digestible steps so you can build your own orchestration layer rather than depend on opaque frameworks.
Key features
Branch-per-step structure: Each Git branch introduces one concept (tools, context, guardrails, verification)
TypeScript implementation: Approachable codebase for web-focused teams
Companion walkthrough: The tej.as blog post explains every decision in depth
How to use
Clone the repo and check out the first branch to see a bare-bones harness that wraps a single LLM call
Step through subsequent branches to add tool registration, context management, and guardrails incrementally
Implement a verification loop that validates LLM output before acting: this pattern catches hallucinations and enforces business rules before side effects occur
For login handling, the harness demonstrates how to inject user context and manage session state across multi-turn conversations
Adapt the patterns to your own agent project, swapping in your preferred LLM provider while keeping the orchestration logic intact
🔗 View on GitHub | GitHub stars: 182
Learning resources
Harnesses in AI: A Deep Dive - Tejas Kumar at AI Engineer Europe 2026 - The conference talk this repo was built for, walking through building the harness live on stage.
What Is an Agent Harness? Harness Engineering Explained - tej.as - Step-by-step walkthrough building the harness repo, covering tools, context, guardrails, verification loops, and login handling.
Markdown Preview
Why this made the cut: A lightweight, native macOS app that renders Markdown beautifully, including Quick Look integration for zero-friction documentation review.
Why it matters
Switching to a browser or heavy editor to preview Markdown breaks flow. Markdown Preview uses a WKWebView pipeline backed by swift-markdown to deliver fast, native rendering with heading anchors and automatic URL linking, all inside Finder's Quick Look.
Key features
Native macOS rendering: Swift 6 and WKWebView for responsive, high-fidelity previews
Quick Look extension: Press Space in Finder to preview any .md file instantly
Clickable links: Bare http:// and https:// URLs become links without extra markup
Heading anchors: Jump to sections directly in long documents
How to use
Download the latest release from GitHub or build from source with Xcode
Move the app to Applications and open it once to register the Quick Look extension
In Finder, select any .md file and press Space for an instant preview
Double-click to open in the full app for scrolling and link navigation
🔗 View on GitHub | GitHub stars: 2,367
Learning resources
Obscure (and Excellent) Mac Apps - Ep 854 - Mac Power Users - Podcast episode featuring Markdown Preview alongside other under-the-radar macOS utilities.
If you only try one
KEDA is the single best starting point this week. Event-driven autoscaling touches nearly every production Kubernetes deployment, and KEDA's 60+ scalers mean you can connect it to whatever queue, database, or custom metric already powers your workload. Once KEDA is in place, you unlock cost savings (scale to zero) and responsiveness (instant scale-up) without writing custom controllers. Start here, and the benefits ripple across your entire infrastructure.
If you are doing Open Source I have a good news for you, I work at CodeRabbit which is an AI review tool and its free for Open Source, please reach out to me on X or LinkedIn or just send an email on [email protected] if you need help on adopting CodeRabbit.
You can visit our portal below to create a new account and connect your repository and start reviewing your code.
